Bumps [@cursor/sdk](https://github.com/cursor/cursor) from 1.0.18 to
1.0.19.
<details>
<summary>Commits</summary>
<ul>
<li>See full diff in <a
href="https://github.com/cursor/cursor/commits">compare view</a></li>
</ul>
</details>
<br />
[](https://docs.github.com/en/github/managing-security-vulnerabilities/about-dependabot-security-updates#about-compatibility-scores)
Dependabot will resolve any conflicts with this PR as long as you don't
alter it yourself. You can also trigger a rebase manually by commenting
`@dependabot rebase`.
[//]: # (dependabot-automerge-start)
[//]: # (dependabot-automerge-end)
---
<details>
<summary>Dependabot commands and options</summary>
<br />
You can trigger Dependabot actions by commenting on this PR:
- `@dependabot rebase` will rebase this PR
- `@dependabot recreate` will recreate this PR, overwriting any edits
that have been made to it
- `@dependabot show <dependency name> ignore conditions` will show all
of the ignore conditions of the specified dependency
- `@dependabot ignore this major version` will close this PR and stop
Dependabot creating any more for this major version (unless you reopen
the PR or upgrade to it yourself)
- `@dependabot ignore this minor version` will close this PR and stop
Dependabot creating any more for this minor version (unless you reopen
the PR or upgrade to it yourself)
- `@dependabot ignore this dependency` will close this PR and stop
Dependabot creating any more for this dependency (unless you reopen the
PR or upgrade to it yourself)
</details>
Signed-off-by: dependabot[bot] <support@github.com>
Co-authored-by: dependabot[bot] <49699333+dependabot[bot]@users.noreply.github.com>
## Thinking Path
> - Paperclip is the open source app people use to manage AI agents for
work
> - Local adapters (Claude, Codex, Cursor, Gemini, Grok, OpenCode, Pi,
ACPX) ship bundled "skills" — opinionated Markdown prompt bundles
materialized into the agent's runtime
> - One of those bundled skills, `paperclip-dev`, existed to let agents
develop Paperclip itself; it has now moved to its own external repo and
no longer belongs in the core tree
> - The adapter skill model also carried a `required` / `requiredReason`
attribute plus a `paperclip_required` `AdapterSkillOrigin` variant, all
of which only existed to mark bundled skills as non-optional in the UI
and adapter sync logic
> - With `paperclip-dev` gone, no bundled skill is "required" anymore,
and the type / runtime surface for `required` is dead weight — but it is
computed at request time and never persisted, so a clean removal is safe
(no compatibility shim needed)
> - This pull request deletes `skills/paperclip-dev/` and removes every
trace of the `required` / `requiredReason` field and the
`paperclip_required` origin across shared types, validators,
adapter-utils, all eight local adapters, server routes, the
company-skills service, the UI, the storybook fixtures, and the test
suite
> - The benefit is a smaller, simpler adapter-skill surface: one origin
(`company_managed`) for managed bundled skills,
`resolvePaperclipDesiredSkillNames` collapses to "just the configured
desired set", and the AgentDetail skills tab no longer renders a
"Required by Paperclip" section that no longer applies
## Linked Issues or Issue Description
<!-- No existing public GitHub issue; describing the underlying work
inline (feature_request template fields). -->
**Summary**
Remove the bundled `paperclip-dev` skill (now maintained in its own
external repo) and retire the `required` / `requiredReason` skill
attribute and the `paperclip_required` skill origin, which only existed
to support it.
**Problem or motivation**
`paperclip-dev` is the only bundled skill that was ever marked
"required". Now that it lives in a separate repository, shipping it
inside the core tree is wrong, and the entire `required` surface (a type
field, a validator field, a synthesized `paperclip_required` origin, UI
"Required by Paperclip" section, and required-skill merging in the
desired-skills calculation) becomes dead weight. The `required` value is
computed at request time and never persisted, so it can be removed
cleanly without a migration or compatibility shim.
**Proposed solution**
Delete `skills/paperclip-dev/`, drop the `required` / `requiredReason`
fields and `paperclip_required` origin everywhere they are produced or
consumed, collapse managed-skill origin to a single `company_managed`
value, and simplify `resolvePaperclipDesiredSkillNames` to return only
the configured desired set.
**Alternatives considered**
Keeping the `required` attribute as a no-op for forward compatibility —
rejected because it is request-time only (nothing persists it), so
leaving it in place is pure dead surface area with no callers.
**Roadmap alignment**
Internal cleanup / dead-code removal that simplifies the adapter-skill
surface; it does not introduce or duplicate any planned core feature in
ROADMAP.md.
## What Changed
- Deleted bundled `skills/paperclip-dev/` (moved to a separate repo).
- Dropped `required`, `requiredReason`, and the `paperclip_required`
origin from `packages/shared/src/types/adapter-skills.ts`,
`packages/shared/src/validators/adapter-skills.ts`, and
`packages/adapter-utils/src/types.ts`.
- In `packages/adapter-utils/src/server-utils.ts`: removed
`readSkillRequired()`; dropped `required`/`requiredReason` from
`listPaperclipSkillEntries()`,
`normalizeConfiguredPaperclipRuntimeSkills()`,
`buildPersistentSkillSnapshot()`, and `PaperclipSkillEntry`; collapsed
`buildManagedSkillOrigin()` to always return `company_managed`;
simplified `resolvePaperclipDesiredSkillNames()` to return only the
configured desired set (signature preserved so adapter call sites are
untouched).
- Walked all eight local adapters (`acpx-local`, `claude-local`,
`codex-local`, `cursor-local`, `gemini-local`, `grok-local`,
`opencode-local`, `pi-local`) and removed every remaining
`requiredReason` / `paperclip_required` reference.
- `server/src/services/company-skills.ts`: dropped the `required =
sourceKind === "paperclip_bundled"` synthesis when listing runtime skill
entries.
- `server/src/routes/agents.ts`: removed required-skill merging from the
desired-skills calculation in the persist-config path and the
unsupported-snapshot path (keeping the current version-aware
`desiredSkillEntries` structure).
- `ui/src/pages/AgentDetail.tsx`: dropped required-based filters, the
required tooltip, and the entire "Required by Paperclip" section from
the agent skills tab; storybook fixtures in
`ui/storybook/stories/acpx-local.stories.tsx` cleaned up to match.
- Tests: deleted the `required: false` case in
`paperclip-skill-utils.test.ts` and the "keeps required bundled skills
installed" case in every `*-local-skill-sync.test.ts`;
`acpx-local-execute.test.ts`, `cursor-local-execute.test.ts`,
`cursor-local-skill-sync.test.ts`, `agent-skills-routes.test.ts`, and
`packages/adapter-utils/src/server-utils.test.ts` were updated to drop
removed fields and map `origin: "paperclip_required"` →
`"company_managed"`.
- `server/src/adapters/registry.ts`: two `as unknown as
ServerAdapterModule["..."]` casts on `hermesListSkills` /
`hermesSyncSkills` (matching the existing `executeHermesLocal` pattern).
`hermes-paperclip-adapter@0.2.0` still depends on the published
`@paperclipai/adapter-utils` which keeps the retired
`paperclip_required` variant; the cast bridges the
workspace-vs-published type mismatch at the registry seam and can drop
once hermes upgrades.
## Verification
Run from the workspace root:
```sh
grep -rn "skills/paperclip-dev" .
grep -rn "paperclip_required" --include="*.ts" --include="*.tsx" .
grep -rn "requiredReason" --include="*.ts" --include="*.tsx" .
pnpm -w typecheck
pnpm --filter @paperclipai/server exec vitest run paperclip-skill-utils
pnpm --filter @paperclipai/server exec vitest run skill-sync
```
The first three greps return only the explanatory comment in
`server/src/adapters/registry.ts` (no live `paperclip_required` /
`requiredReason` usage) and zero `skills/paperclip-dev` source hits.
Locally:
- `pnpm -w typecheck` → all packages this PR touches pass
(adapter-utils, shared, server, ui, cli, and the
cursor/gemini/opencode/pi adapters).
- Affected vitest suites pass: `paperclip-skill-utils`, `server-utils`,
all eight `*-local-skill-sync`, `agent-skills-routes`, and the
`acpx`/`cursor`/`pi` execute suites.
## Risks
- Behavioral shift in the agent skills UI: the "Required by Paperclip"
section disappears. No bundled skill is required anymore, so this only
affects environments that previously surfaced `paperclip-dev` as a
forced-on row; those installs will see the skill move into the regular
"company-managed" list (and be uninstalled on next sync unless
explicitly listed as desired).
- Existing agents may still have the string `"paperclip-dev"` in their
persisted `desiredSkills`. That entry is inert (no source for it to
install from); a one-time DB cleanup is out of scope. Low risk.
- Hermes adapter type bridge: two casts in `registry.ts` paper over a
type-only divergence between the workspace `@paperclipai/adapter-utils`
and the published version still pinned by
`hermes-paperclip-adapter@0.2.0`. Runtime behavior is unaffected because
the retired `paperclip_required` value is no longer produced by anything
in this tree. The casts can be removed once hermes upgrades its
dependency.
## Model Used
- Provider: Anthropic
- Model: Claude Opus 4.7 (`claude-opus-4-7`)
- Capability: agent tool use via Paperclip's `claude_local` adapter
## Checklist
- [x] I have included a thinking path that traces from project context
to this change
- [x] I have specified the model used (with version and capability
details)
- [x] I have checked ROADMAP.md and confirmed this PR does not duplicate
planned core work
- [x] I have searched GitHub for duplicate or related PRs and linked
them above
- [x] I have either (a) linked existing issues with `Fixes: #` / `Closes
#` / `Refs #` OR (b) described the issue in-PR following the relevant
issue template
- [x] I have not referenced internal/instance-local Paperclip issues or
links (only public GitHub `#NNN` / `github.com/paperclipai/paperclip`
URLs)
- [ ] My branch name describes the change (e.g. `docs/...`, `fix/...`)
and contains no internal Paperclip ticket id or instance-derived details
- [x] I have run tests locally and they pass
- [x] I have added or updated tests where applicable
- [x] I have updated relevant documentation to reflect my changes
- [x] I have considered and documented any risks above
- [ ] All Paperclip CI gates are green
- [ ] Greptile is 5/5 with no open P2s, recommendations, or follow-ups
- [x] I will address all Greptile and reviewer comments before
requesting merge
Co-authored-by: Paperclip <noreply@paperclip.ing>
## Thinking Path
> - Paperclip orchestrates AI agents for zero-human companies
> - The `codex_local` adapter runs `codex exec` as a child process and
streams JSONL events from its stdout
> - NEE-79 caught a real-world `codex_local` orphan: codex sat in
`read()` on stdin for 1h+, no JSON events emitted past startup, no LLM
call in flight. The adapter had no inactivity timer; the only safety net
was the platform-level 1h silent-run detector
> - This is precisely the failure shape NEE-80 scoped: an
adapter-detected fault that should be killed *by the adapter* and
surfaced as `failed`, not waited out for an hour
> - This pull request adds an output-inactivity watchdog inside the
codex-local adapter that resets on every parsed JSONL event from stdout,
kills the child via SIGTERM → 5s grace → SIGKILL when it fires, and
resolves the run with a structured watchdog failure
> - The benefit is that NEE-79-class hangs shorter than 1h stop reaching
the platform-level safety net — they fail fast and visibly at the
adapter layer, with diagnostic logs that don't require host shell access
## Linked Issues or Issue Description
No existing GitHub issue covers this; describing the underlying bug
in-PR per the bug report template
(`.github/ISSUE_TEMPLATE/bug_report.yml`):
- **What happened?** A `codex_local` run sat in `read()` on stdin for
over an hour: codex emitted its startup JSONL events, then nothing — no
further events, no LLM call in flight, no exit. The adapter kept the
child alive indefinitely; the run was only reaped by the platform-level
1h silent-run safety net, an hour after it had effectively died.
- **Expected behavior:** The adapter should detect that its child has
stopped producing output long before the platform-level safety net, kill
it, and surface the run as `failed` with a diagnostic that explains what
happened.
- **Steps to reproduce:** Run any `codex_local` issue where `codex exec`
hangs after startup (e.g. codex blocks reading stdin and never emits
another JSONL event). Observe the run stays alive until the 1h platform
safety net fires.
- **Paperclip version or commit:** reproduced on master prior to this
branch.
- **Agent adapter(s) involved:** codex_local.
Related PRs found while searching for duplicates (none implement an
event-aware inactivity watchdog inside `codex_local`):
- #4004 — generic idle + wall watchdogs for `runChildProcess` in
adapter-utils; complementary, operates below the JSONL parse layer and
is not codex-event-aware
- #4742 — fail-fast on codex-local *startup* hang; this PR covers the
post-startup hang class
- #6861 — zombie-run termination for codex-local; reaping after exit,
not inactivity detection
- #7811 — the analogous output-idle timeout for grok-local
## What Changed
- `packages/adapters/codex-local/src/server/watchdog.ts` *(new)* — pure
watchdog primitive with injectable timers/clock:
`resolveCodexInactivityTimeout`, `createCodexInactivityWatchdog`,
`formatWatchdogErrorMessage`. Default `7 * 60_000` ms; honors `null` as
the disabled escape hatch
- `packages/adapters/codex-local/src/server/execute.ts` — `runAttempt`
now wraps `onSpawn` to capture `pid`/`processGroupId`, feeds stdout
chunks through `noteStdoutChunk`, and on watchdog fire sends SIGTERM to
the process group, schedules SIGKILL after 5s, and returns an
`AdapterExecutionResult` with `exitCode: null`, `signal:
SIGTERM|SIGKILL`, `errorMessage: "watchdog: no codex output for {N}m
{S}s"`, `errorCode: "codex_output_inactivity_watchdog"`. With
`errorMessage` set and `timedOut: false`, `heartbeat.ts:5860` maps the
run to `outcome === "failed"` (not `cancelled`)
- `packages/adapters/codex-local/src/index.ts` — `agentConfigurationDoc`
documents `outputInactivityTimeoutMs` (number ms; `null` disables;
non-positive falls back to default with a warning log at spawn)
- `packages/adapters/codex-local/src/server/watchdog.test.ts` *(new)* —
13 tests covering acceptance criteria 2 and 3 plus supporting cases:
fires after silence, no-fire across 12× (threshold − 1s) cycles,
multi-event chunks, non-JSON ignoring, single-fire idempotency,
formatter shape, full resolution table, `null` → disabled
-
`packages/adapters/codex-local/src/server/watchdog.integration.test.ts`
*(new)* — real Node subprocess that prints one JSONL event then sleeps;
`runChildProcess` reaps it within `threshold + 6s`; signal is SIGTERM or
SIGKILL; `parsedEventCount === 1` (acceptance criteria 1 and 4)
## Verification
```
$ pnpm --filter @paperclipai/adapter-codex-local typecheck
> tsc --noEmit
# clean
$ pnpm --filter @paperclipai/adapter-codex-local exec vitest run
✓ src/server/quota-spawn-error.test.ts (1 test)
✓ src/server/codex-home.test.ts (3 tests)
✓ src/server/codex-args.test.ts (3 tests)
✓ src/server/watchdog.test.ts (13 tests)
✓ src/server/parse.test.ts (9 tests)
✓ src/server/execute.remote.test.ts (4 tests)
✓ src/ui/parse-stdout.test.ts (3 tests)
✓ src/ui/build-config.test.ts (1 test)
✓ src/server/watchdog.integration.test.ts (1 test)
Test Files 9 passed (9)
Tests 38 passed (38)
```
Acceptance criteria check:
1. ✅ Simulated child emits one event then sleeps → killed at threshold;
result `errorMessage` matches `watchdog: no codex output for {N}m {S}s`
(`watchdog.integration.test.ts`)
2. ✅ Child emits events every (threshold − 1s) → not killed
(`watchdog.test.ts: "does not fire when events arrive every (threshold -
1s)"`)
3. ✅ `outputInactivityTimeoutMs: null` disables the watchdog
(`resolveCodexInactivityTimeout` returns `disabled`; `execute.ts` skips
watchdog construction and logs a startup warning)
4. ✅ Real subprocess reaped well within `threshold + 6s` — 250 ms
threshold, 290 ms wall clock in CI
5. ✅ Adapter-level fault → `outcome === "failed"` per
`heartbeat.ts:5860`. NEE-79-class hangs <1h get caught at the adapter,
not the platform-level safety net
Post-rebase verification (head `2a8703fab`, rebased onto master
`69a368ed5`):
```
$ pnpm --filter @paperclipai/adapter-codex-local typecheck # clean
$ pnpm --filter @paperclipai/adapter-codex-local exec vitest run
Test Files 11 passed (11)
Tests 64 passed (64)
```
## Risks
- Low risk. Behind a default-on watchdog that only fires after 7m of
zero parsed JSON events. Operators can disable it with
`outputInactivityTimeoutMs: null` for known-slow tasks
- The kill path reuses the same `process.kill(-pgid, signal)` pattern
that `runChildProcess` already uses for its terminal-result cleanup, so
signal semantics match the existing code path
- `timedOut: false` is preserved on watchdog fire — the platform-level
timeout outcome is unchanged, only the `failed`-vs-success
classification flips. No behavioral shift for already-failing runs
- Sandbox/SSH execution targets: the watchdog fires and emits the
structured log, but the kill is best-effort because remote pids aren't
owned by this process. The platform-level 1h safety net still applies.
Out of scope for NEE-81 by design
## Model Used
- Provider: Anthropic Claude
- Model: claude-opus-4-7
- Mode: Claude Code (extended thinking, tool use)
## Checklist
- [x] I have included a thinking path that traces from project context
to this change
- [x] I have specified the model used (with version and capability
details)
- [x] I have checked ROADMAP.md and confirmed this PR does not duplicate
planned core work
- [x] I have searched GitHub for duplicate or related PRs and linked
them above
- [x] I have either (a) linked existing issues with `Fixes: #` / `Closes
#` / `Refs #` OR (b) described the issue in-PR following the relevant
issue template
- [x] I have run tests locally and they pass
- [x] I have added or updated tests where applicable
- [ ] If this change affects the UI, I have included before/after
screenshots (N/A — no UI changes)
- [x] I have updated relevant documentation to reflect my changes
(`agentConfigurationDoc` for `outputInactivityTimeoutMs`)
- [x] I have considered and documented any risks above
- [ ] All Paperclip CI gates are green (re-running on the rebased head)
- [x] Greptile is 5/5 with no open P2s, recommendations, or follow-ups
(P2 addressed; review threads resolved)
- [x] I will address all Greptile and reviewer comments before
requesting merge
---------
Co-authored-by: Neeraj Kumar Singh <b.nirajkumarsingh@hotmail.com>
Co-authored-by: Devin Foley <devin@paperclip.ing>
Co-authored-by: Paperclip <noreply@paperclip.ing>
## Thinking Path
> - Paperclip is the open source app people use to manage AI agents for
work
> - The `codex_local` adapter isolates local Codex runs by assigning
managed `CODEX_HOME` state per company and per agent
> - PR #8272 tightened that isolation, but it left a gap: once
`CODEX_HOME` became explicit, the adapter treated it like a user-managed
override and skipped auth seeding
> - That meant newly isolated agents could launch with no usable
`auth.json`, hit OpenAI unauthenticated, and fail with `401 Missing
bearer`
> - Users who had already persisted one of those broken managed homes
could remain stranded even after config changes unless Paperclip
repaired the home itself
> - This pull request teaches Paperclip to seed managed homes correctly,
backfill already-stranded managed homes on startup, and reject
credential-less managed homes before they reach the provider
> - The benefit is that affected managed `codex_local` agents recover
automatically after upgrade and restart, without manual `CODEX_HOME`
surgery
## Linked Issues or Issue Description
- Fixes#497
- Refs #5028
- Related PR: #8272
- Related PR: #8399
## What Changed
- Distinguished Paperclip-managed `CODEX_HOME` paths from genuine
external overrides and always seeded auth into managed homes, even when
`CODEX_HOME` is explicit in config.
- Wrote API-key-backed `auth.json` files for managed homes when
`OPENAI_API_KEY` is configured, otherwise symlinked the shared Codex
auth for subscription/OAuth flows.
- Added a startup reconciliation pass that backfills already-isolated
managed homes created by the broken release so upgrade plus restart
repairs stranded agents automatically.
- Preserved previously resolved API-key auth when the stored
`OPENAI_API_KEY` binding is secret-backed and startup cannot resolve the
secret value directly.
- Hardened the managed-home preflight to require a credential-bearing
`auth.json`, not just file presence, and documented the recovery
behavior.
- Added regression tests covering managed-home seeding, fail-fast
behavior, and server-side startup reconciliation.
## Verification
```bash
pnpm exec vitest run packages/adapters/codex-local/src/server/codex-home.test.ts packages/adapters/codex-local/src/server/execute.auth.test.ts server/src/__tests__/codex-auth-reconciliation.test.ts server/src/__tests__/codex-local-execute.test.ts server/src/__tests__/server-startup-feedback-export.test.ts
pnpm --filter @paperclipai/adapter-codex-local typecheck
pnpm --filter @paperclipai/server typecheck
pnpm check:tokens
pnpm exec vitest run server/src/__tests__/server-startup-feedback-export.test.ts
```
GitHub Actions `PR` workflow is green on latest head `e5b1e08d4`,
including policy, typecheck, test shards, build, e2e, serialized server
suites, and canary dry run. Greptile Review is green on latest head with
0 comments added. No UI changes.
## Risks
- Startup reconciliation now mutates persisted managed Codex homes at
boot. Risk is low because it only touches Paperclip-managed
company/agent home paths and no-ops when a home already has usable auth.
- Genuine external `CODEX_HOME` overrides remain intentionally
self-managed, so those users still own repair steps inside their custom
home.
- Hosts with neither shared Codex auth nor an explicit per-agent API key
now fail earlier with a clearer adapter error instead of surfacing a
downstream `401`, which changes timing but not capability.
## Model Used
- OpenAI Codex, GPT-5-based coding agent in a local Codex session; exact
served model ID/context window were not exposed to the session. Tool use
and code execution were enabled.
## Checklist
- [x] I have included a thinking path that traces from project context
to this change
- [x] I have specified the model used (with version and capability
details)
- [x] I have checked ROADMAP.md and confirmed this PR does not duplicate
planned core work
- [x] I have searched GitHub for duplicate or related PRs and linked
them above
- [x] I have either (a) linked existing issues with `Fixes: #` / `Closes
#` / `Refs #` OR (b) described the issue in-PR following the relevant
issue template
- [x] I have not referenced internal/instance-local Paperclip issues or
links (only public GitHub `#NNN` / `github.com/paperclipai/paperclip`
URLs)
- [x] My branch name describes the change and contains no internal
Paperclip ticket id or instance-derived details
- [x] I have run tests locally and they pass
- [x] I have added or updated tests where applicable
- [x] If this change affects the UI, I have included before/after
screenshots
- [x] I have updated relevant documentation to reflect my changes
- [x] I have considered and documented any risks above
- [x] All Paperclip CI gates are green
- [x] Greptile is 5/5 with no open P2s, recommendations, or follow-ups
- [x] I will address all Greptile and reviewer comments before
requesting merge
---------
Co-authored-by: Paperclip <noreply@paperclip.ing>
## Thinking Path
> - Paperclip is the open source app people use to manage AI agents for
work
> - The claude-local adapter launches the Claude CLI inside execution
environments (local, SSH, and ephemeral sandboxes such as Daytona)
> - Newer adapter code passes `--effort` to the CLI, but the Claude
binary baked into some sandbox images is older and rejects it with
`error: unknown option '--effort'`, so every run in those environments
fails
> - This needs addressing because the failure is environment-dependent
and silent from the operator's perspective — the run just dies with a
CLI usage error
> - This pull request probes the in-sandbox CLI for `--effort` support
once, caches the result per environment, and strips the flag (with a
warning) when the CLI does not support it
> - The benefit is that sandboxes with older Claude CLIs keep working
instead of failing, with negligible probe overhead because the
capability check is cached and reused across ephemeral leases
## Linked Issues or Issue Description
No public GitHub issue exists. Describing the bug inline following the
bug report template:
### What happened?
Runs using the `claude-local` adapter inside certain sandbox/execution
environments fail with `error: unknown option '--effort'`. The adapter
unconditionally appends `--effort` to the Claude CLI invocation, but the
Claude CLI version present in some sandbox base images predates that
flag, so the process exits with a usage error and the run dies.
### Expected behavior
The adapter should detect that the target environment's Claude CLI does
not support `--effort` and degrade gracefully — drop the flag and emit a
warning — rather than failing the run.
### Steps to reproduce
1. Configure an execution environment (e.g. a sandbox image) whose
bundled Claude CLI is old enough to predate the `--effort` option.
2. Run any claude-local task that resolves to an effort level (so
`--effort` is appended).
3. Observe the run fail immediately with `error: unknown option
'--effort'`.
### Paperclip version or commit
`master` at the time of this PR (branch forked from current `master`).
### Deployment mode
Self-hosted / local instance using execution environments (reproducible
with ephemeral sandbox providers such as Daytona where `reuseLease:
false`).
### Agent adapter(s) involved
Claude Code (`claude-local`).
## What Changed
- Add a CLI capability probe (`cli-capabilities.ts`) that runs the
target Claude binary's `--help` inside the execution environment to
detect `--effort` support.
- Strip `--effort` from the CLI args (emitting a warning) when the probe
reports the flag is unsupported; keep it otherwise.
- Cache probe results keyed by
`sandbox:providerKey:environmentId:command` (no lease id) so the probe
is reused across ephemeral leases — important for `reuseLease: false`
sandbox configs like Daytona, which would otherwise re-probe on every
run.
- Conservative fallback: if the probe itself can't run/parse, assume the
flag is supported (preserves prior behavior).
## Verification
- `node_modules/.bin/vitest run
src/__tests__/claude-local-execute.test.ts
src/__tests__/claude-local-adapter-environment.test.ts` → **2 files, 30
tests passed**.
- Regression test issues two `execute()` calls with distinct lease ids
and asserts the in-sandbox `--help` probe runs exactly once (cache reuse
across leases).
- Added tests covering: flag stripped when unsupported, flag retained
when supported, warning emitted, and conservative fallback when the
probe fails.
## Risks
Low risk. The change is additive and gated behind a probe with a
conservative default (assume supported on probe failure), so existing
environments that support `--effort` are unaffected. Worst case for an
environment where the probe is unreliable is the prior behavior (flag
passed through).
## Model Used
Claude Opus 4.8 (claude-opus-4-8), extended thinking, with tool use /
code execution via Claude Code.
## Checklist
- [x] I have included a thinking path that traces from project context
to this change
- [x] I have specified the model used (with version and capability
details)
- [x] I have checked ROADMAP.md and confirmed this PR does not duplicate
planned core work
- [x] I have searched GitHub for duplicate or related PRs and linked
them above
- [x] I have either (a) linked existing issues with `Fixes: #` / `Closes
#` / `Refs #` OR (b) described the issue in-PR following the relevant
issue template
- [x] I have not referenced internal/instance-local Paperclip issues or
links (only public GitHub `#NNN` / `github.com/paperclipai/paperclip`
URLs)
- [x] My branch name describes the change (e.g. `docs/...`, `fix/...`)
and contains no internal Paperclip ticket id or instance-derived details
- [x] I have run tests locally and they pass
- [x] I have added or updated tests where applicable
- [ ] If this change affects the UI, I have included before/after
screenshots (N/A — no UI change)
- [ ] I have updated relevant documentation to reflect my changes (N/A —
no doc-facing behavior change)
- [x] I have considered and documented any risks above
- [ ] All Paperclip CI gates are green (pending CI)
- [ ] Greptile is 5/5 with no open P2s, recommendations, or follow-ups
(pending review)
- [x] I will address all Greptile and reviewer comments before
requesting merge
## Thinking Path
> - Paperclip is the control plane for running AI-agent companies, so
long-running remote work needs to stay observable to human operators.
> - Cloud / sandbox agents are an active roadmap area, and their
workspace sync path is part of the runtime substrate every remote coding
run depends on.
> - In the sandbox and SSH execution-target flows, Paperclip logged that
sync had started, then often went silent for the full transfer window.
> - That made large remote syncs feel stalled and also hid a real
performance problem in the command-managed sandbox upload path.
> - The first part of this pull request threads a throttled
progress-reporting surface through the adapter execution-target stack so
sync and restore work can emit meaningful updates.
> - The second part fixes the command-managed sandbox transport itself:
it removes the old serial 32KB append bottleneck, but also falls back
away from the single-stream path when a provider-backed sandbox runner
cannot surface mid-flight stdin progress.
> - The result is that sandbox and SSH transfers are both faster and
more observable, including the live Daytona-style sandbox case that
previously only emitted `0%` and `100%`.
## Linked Issues or Issue Description
No public GitHub issue exists for this bug, so it is described inline
below following the bug report template.
### What happened
- Remote sandbox and SSH workspace syncs could spend a long time
transferring data while only logging a start line (`Syncing workspace
and runtime assets to sandbox environment`) and, at best, a terminal
line.
- In the command-managed sandbox path, the original upload
implementation also paid a large performance cost by appending base64
data in many small sequential remote writes (thousands of serial 32KB
round-trips on a large workspace).
- After the initial transport rewrite, live provider-backed sandbox runs
still only emitted `0%` and `100%` because the single-stream stdin RPC
buffered progress until completion.
### Expected behavior
- Long-running sandbox and SSH syncs should periodically report how much
of the transfer is complete (a percentage and/or MB transferred) so an
operator can tell the run is healthy and making progress rather than
stuck.
- The main sandbox upload path should not be artificially slow.
- A transfer that fails partway should leave an explicit failure marker
in the log rather than a dangling intermediate percentage.
### Steps to reproduce
1. Run an agent against a sandbox (command-managed) or SSH
(remote-managed) execution target with a non-trivial workspace.
2. Watch the run log during the workspace/runtime asset sync phase.
3. Observe that the log shows the sync start line and then stays silent
for the full transfer (live provider-backed sandbox runs only show `0%`
then `100%`).
### Paperclip version or commit
- Branch `PAPA-825-provide-status-updates-when-syncing-sandboxes` off
`master`.
### Deployment mode
- Self-hosted / local instance using sandbox (command-managed) and SSH
(remote-managed) execution targets, including provider-backed sandbox
runners.
## What Changed
- Added shared throttled runtime progress reporting and threaded
`onProgress` through the adapter execution-target surface and adapter
`execute.ts` entrypoints.
- Added sync and restore progress reporting for the command-managed
sandbox path and the SSH/remote-managed path, including git
import/export progress where totals are known.
- Reworked command-managed sandbox transfer behavior so uploads use the
faster single-stream path when appropriate, but fall back to chunked
progress-emitting writes when the runner cannot expose mid-stream stdin
progress.
- Marked provider-backed environment sandbox runners as not supporting
single-stream stdin progress so live sandbox runs emit meaningful
intermediate updates instead of only `0%` and `100%`.
- Emit an explicit terminal failure marker (`failed at NN% (x/y MB)`)
when an SSH/tar transfer rejects, so a failed sync no longer leaves a
dangling intermediate percentage in the log.
- Run the SSH sync/restore size estimate (local directory walk / remote
`du` probe) concurrently with the transfer instead of awaiting it before
opening the pipe, so progress instrumentation no longer adds startup
latency proportional to workspace file count.
- Added and extended focused regression coverage for runtime progress
throttling and the new failure marker, command-managed sandbox
transfers, sandbox orchestration, SSH transfer progress, and environment
execution-target wiring.
## Verification
- `pnpm exec vitest run
packages/adapter-utils/src/runtime-progress.test.ts
packages/adapter-utils/src/ssh-fixture.test.ts
packages/adapter-utils/src/command-managed-runtime.test.ts
packages/adapter-utils/src/sandbox-managed-runtime.test.ts`
- `pnpm exec vitest run
packages/adapter-utils/src/command-managed-runtime.test.ts
server/src/__tests__/environment-execution-target.test.ts`
- `npx tsc --noEmit` for `packages/adapter-utils`
## Risks
- The provider-backed sandbox fallback now prefers chunked
command-managed writes when progress hooks are active, so
small-to-medium uploads may trade some raw throughput for observable
intermediate progress on runtimes that cannot surface true mid-stream
stdin progress.
- Progress percentages on tar-based transfers still depend on estimates
in some cases, so operators may briefly see MB-only lines before the
estimate resolves, then near-final clamping before the terminal `100%`
line.
- This PR changes shared execution-target behavior used by multiple
adapters, so regressions would most likely appear in remote runtime
setup/teardown flows rather than in a single adapter.
## Model Used
- Initial implementation: OpenAI GPT-5.4 via Codex local agent
(`codex_local`), high reasoning mode.
- Observability follow-ups (failure marker, concurrent size estimate,
added tests): Claude Opus 4.8 via Claude Code (`claude_local`).
## Checklist
- [x] I have included a thinking path that traces from project context
to this change
- [x] I have specified the model used (with version and capability
details)
- [x] I have checked ROADMAP.md and confirmed this PR does not duplicate
planned core work
- [x] I have searched GitHub for duplicate or related PRs and linked
them above
- [x] I have either (a) linked existing issues with `Fixes: #` / `Closes
#` / `Refs #` OR (b) described the issue in-PR following the relevant
issue template
- [x] I have not referenced internal/instance-local Paperclip issues or
links (only public GitHub `#NNN` / `github.com/paperclipai/paperclip`
URLs)
- [x] I have run tests locally and they pass
- [x] I have added or updated tests where applicable
- [ ] If this change affects the UI, I have included before/after
screenshots
- [ ] I have updated relevant documentation to reflect my changes
- [x] I have considered and documented any risks above
- [ ] All Paperclip CI gates are green
- [ ] Greptile is 5/5 with no open P2s, recommendations, or follow-ups
- [ ] I will address all Greptile and reviewer comments before
requesting merge
---------
Co-authored-by: Paperclip <noreply@paperclip.ing>
## Thinking Path
> - Paperclip is the open source app people use to manage AI agents for
work
> - Agents run through adapters; `@paperclipai/adapter-claude-local`
shells out to the Claude CLI and maps its result JSON into Paperclip's
run outcome (`errorCode` / `errorFamily`)
> - A Fable 5 policy refusal exits the CLI cleanly (`exitCode=0`,
`is_error=false`) with `stop_reason: "refusal"`, which the adapter
mapped to `errorCode: null`
> - Paperclip therefore recorded the run as a silent success, so the
agent's heartbeat stalled with no signal for operators and no hook to
retry or alert
> - This pull request adds a refusal detector to the adapter so a
refusal resolves to a distinct `errorCode: "claude_refusal"`
(`errorFamily: "model_refusal"`)
> - The benefit is that policy refusals become observable: the server
persists the code, operators can see it on the run, and downstream
retry/alert policy can key off it
## Linked Issues or Issue Description
No public GitHub issue exists. Describing the underlying bug in-PR,
following `.github/ISSUE_TEMPLATE/bug_report.yml`:
**What happened?**
When the Claude CLI returns `stop_reason: "refusal"` (a model policy
refusal), it still exits cleanly (`exitCode=0`, `is_error=false`).
`@paperclipai/adapter-claude-local` keyed refusal detection off the
failure flag during error-code resolution, so it returned `errorCode:
null` — a refused run was indistinguishable from a successful one,
recorded by Paperclip as a silent success with no signal to retry or
alert.
**Expected behavior**
A refusal should resolve to a distinct, non-null error code (`errorCode:
"claude_refusal"`, `errorFamily: "model_refusal"`) so the server can
surface it to operators and downstream policy can react.
**Steps to reproduce**
1. Run any agent on the `claude-local` adapter with a prompt the model
refuses on policy grounds.
2. The Claude CLI exits `0` with `is_error=false` and `stop_reason:
"refusal"`.
3. Observe the run resolves `errorCode: null` (pre-fix) instead of a
refusal-specific code.
**Paperclip version or commit**
`adapter-claude-local` v2026.609.0 (`dist/server/execute.js` error-code
resolution block); reproduced on `master`.
**Deployment mode**
Local / self-hosted.
**Agent adapter(s) involved**
`@paperclipai/adapter-claude-local` (Claude Code, local).
## What Changed
- **`packages/adapters/claude-local/src/server/parse.ts`** — new
`isClaudeRefusalResult()` helper, parallel to
`isClaudeMaxTurnsResult()`. Detects `stop_reason` / `stopReason` /
`error_code` / `errorCode` == `refusal` and `subtype` ==
`model_refusal`, case/whitespace tolerant.
- **`packages/adapters/claude-local/src/server/execute.ts`** — compute
the refusal flag independent of the `failed` flag (a refusal exits
cleanly, so keying off `failed` would miss it); resolve `errorCode:
"claude_refusal"` and `errorFamily: "model_refusal"`; surface
`stopReason: "refusal"` in `resultJson`.
- **`packages/adapter-utils/src/types.ts`** — widen
`AdapterExecutionErrorFamily` with `"model_refusal"`.
- **`packages/adapters/claude-local/src/server/index.ts`** — export the
new helper.
- **`packages/adapters/claude-local/src/server/parse.test.ts`** — 7 new
unit tests.
## Verification
```
npx vitest run packages/adapters/claude-local # 35 passed (7 new)
npx tsc --noEmit -p packages/adapter-utils # clean
npx tsc --noEmit -p packages/adapters/claude-local # clean
```
Manual: a result JSON with `stop_reason: "refusal"` on a clean exit now
resolves `errorCode: "claude_refusal"` and `errorFamily:
"model_refusal"`; non-refusal results are unaffected.
## Risks
Low risk. Additive only — it introduces a new error code on a path that
previously returned `null`; no existing error code or success path
changes. `claude_refusal` is deliberately **not** added to the
transient-upstream retry set: a refusal is deterministic, so retrying
the same prompt yields the same refusal. It surfaces as a distinct
non-transient code for operators; downstream retry/alert policy can key
off `claude_refusal` / `errorFamily: model_refusal` later.
## Model Used
Claude Opus 4.8 (`claude-opus-4-8`), extended-thinking mode, run via
Claude Code with tool use.
## Checklist
- [x] I have included a thinking path that traces from project context
to this change
- [x] I have specified the model used (with version and capability
details)
- [x] I have checked ROADMAP.md and confirmed this PR does not duplicate
planned core work
- [x] I have searched GitHub for duplicate or related PRs and linked
them above (no related PRs found)
- [x] I have either (a) linked existing issues with `Fixes: #` / `Closes
#` / `Refs #` OR (b) described the issue in-PR following the relevant
issue template
- [x] I have not referenced internal/instance-local Paperclip issues or
links (only public GitHub `#NNN` / `github.com/paperclipai/paperclip`
URLs)
- [x] I have run tests locally and they pass
- [x] I have added or updated tests where applicable
- [x] If this change affects the UI, I have included before/after
screenshots (N/A — no UI changes)
- [x] I have updated relevant documentation to reflect my changes (N/A —
internal adapter error-code addition; no user-facing docs)
- [x] I have considered and documented any risks above
- [ ] All Paperclip CI gates are green (this push re-runs the gates;
will confirm once green)
- [ ] Greptile is 5/5 with no open P2s, recommendations, or follow-ups
(prior run was 4/5 on a PR-description note now addressed; will confirm
on re-run)
- [x] I will address all Greptile and reviewer comments before
requesting merge
Co-authored-by: Full-Stack Engineer <noreply@paperclip.ing>
## Thinking Path
> - Paperclip is the open source app people use to manage AI agents for
work.
> - Local adapters turn Paperclip runs into unattended CLI invocations.
> - The `gemini_local` adapter depends on Gemini CLI behaving
non-interactively in headless worker sessions.
> - When Gemini CLI falls back to browser-based auth, the process can
stall at startup instead of producing stream-json output.
> - The adapter should make headless intent explicit and turn missing
auth into a classified, fast failure.
> - This pull request hardens the Gemini child-process environment and
parser classification around that failure mode.
> - The benefit is that operators get actionable `gemini_auth_required`
failures instead of silent hung runs.
## Linked Issues or Issue Description
No exact public issue exists for this runtime stall.
Related: #2344 covers a Gemini CLI adapter environment/auth probe
failure. This PR addresses a different runtime path: unattended
`gemini_local` executions that can stall when Gemini CLI attempts
interactive browser auth in a headless session.
Bug summary:
- Adapter: `gemini_local`
- Symptom: child Gemini CLI process starts but produces no stream-json
output when auth requires interactive browser flow
- Expected: unattended runs either produce stream-json output or fail
quickly with a classified auth error
- Actual: the run can hang at invocation until an external watchdog
kills it
- Scope: Gemini adapter process env, auth-required parsing, and adapter
docs/tests
Duplicate search performed:
- `gh pr list --state all --search "gemini headless invocation stall"`
found only this PR.
- `gh pr list --state all --search "gemini NO_BROWSER NO_COLOR"` found
only this PR.
- `gh issue list --state all --search "gemini headless authentication"`
found related issue #2344 but no exact duplicate.
## What Changed
- Set a headless-safe Gemini invocation env at the final child-process
boundary: `TERM=xterm-256color`, `COLORTERM=truecolor`, and
`NO_BROWSER=1`.
- Delete inherited `NO_COLOR` from the child env so Gemini CLI keeps
color-capable terminal behavior when deciding whether it can run
non-interactively.
- Classify Gemini `FatalAuthenticationError: Manual authorization is
required...` failures as `gemini_auth_required`.
- Update Gemini adapter docs to describe the non-interactive `--prompt`
path and headless env behavior.
- Add/extend focused parser and remote execution tests for the
auth-required and env invariants.
## Verification
- `pnpm exec vitest run
packages/adapters/gemini-local/src/server/parse.test.ts
packages/adapters/gemini-local/src/server/execute.remote.test.ts` — 23
tests passed.
- `pnpm --filter @paperclipai/adapter-gemini-local typecheck` — passed.
- Local Gemini CLI probe confirmed `NO_BROWSER=1` turns the browser-auth
stall into a fast auth error instead of an interactive wait.
## Risks
Low risk. The change is limited to `gemini_local` invocation env,
parsing, docs, and tests. It does not touch schemas, API routes,
persisted data, or other adapters.
Operational risk: environments that intentionally rely on `NO_COLOR`
being inherited by Gemini child processes will no longer pass that
variable through. That is intentional here because Gemini CLI
auth/headless behavior should not depend on inherited color suppression.
> For core feature work, check [`ROADMAP.md`](ROADMAP.md) first and
discuss it in `#dev` before opening the PR. Feature PRs that overlap
with planned core work may need to be redirected — check the roadmap
first. See `CONTRIBUTING.md`.
## Model Used
OpenAI GPT-5 Codex via Codex CLI, with shell/file-editing tool use for
repository inspection, code edits, tests, and GitHub PR maintenance.
## Checklist
- [x] I have included a thinking path that traces from project context
to this change
- [x] I have specified the model used (with version and capability
details)
- [x] I have checked ROADMAP.md and confirmed this PR does not duplicate
planned core work
- [x] I have searched GitHub for duplicate or related PRs and linked
them above
- [x] I have either (a) linked existing issues with `Fixes: #` / `Closes
#` / `Refs #` OR (b) described the issue in-PR following the relevant
issue template
- [x] I have not referenced internal/instance-local Paperclip issues or
links (only public GitHub `#NNN` / `github.com/paperclipai/paperclip`
URLs)
- [x] My branch name describes the change (e.g. `docs/...`, `fix/...`)
and contains no internal Paperclip ticket id or instance-derived details
- [x] I have run tests locally and they pass
- [x] I have added or updated tests where applicable
- [ ] If this change affects the UI, I have included before/after
screenshots
- [x] I have updated relevant documentation to reflect my changes
- [x] I have considered and documented any risks above
- [ ] All Paperclip CI gates are green
- [ ] Greptile is 5/5 with no open P2s, recommendations, or follow-ups
- [x] I will address all Greptile and reviewer comments before
requesting merge
/cc @codex — please review.
## Thinking Path
> - Paperclip orchestrates AI agents for zero-human companies.
> - The `codex_local` adapter runs local Codex CLI processes and builds
their environment from persisted agent config plus host process env.
> - A host-level `OPENAI_API_KEY` or shared Codex auth home can silently
make new agents spend through shared credentials.
> - Existing agents can be repaired manually, but new and updated agents
need a persistent guard at the agent configuration boundary.
> - This pull request isolates new and updated `codex_local` agents with
per-agent `CODEX_HOME` and an empty `OPENAI_API_KEY` override.
> - The benefit is that future agent creation or adapter updates cannot
silently fall back to shared OpenAI credentials.
## Linked Issues or Issue Description
Paperclip work item: [ZOL-5477](/ZOL/issues/ZOL-5477).
No matching GitHub issue exists, so the bug is described inline
following `.github/ISSUE_TEMPLATE/bug_report.yml`.
**Pre-submission checklist**
- [x] I have searched existing open and closed issues and this is not a
duplicate.
- [x] I am on the latest `master` commit for this PR branch.
- [x] I have confirmed the error originates in Paperclip's `codex_local`
adapter configuration boundary, not in a provider outage.
**What happened?**
New or updated `codex_local` agents could inherit a host-level
`OPENAI_API_KEY` or use a shared Codex home when their adapter config
did not explicitly isolate those values. That made it possible for
future agents or manual adapter edits to silently fall back to shared
OpenAI credentials.
**Expected behavior**
Creating, hiring, or updating a `codex_local` agent should either
persist isolated per-agent configuration or reject unsafe shared Codex
home configuration with a clear 422 response. The guard must not print
secret values.
**Steps to reproduce**
1. Create or update a `codex_local` agent without an explicit
`adapterConfig.env.OPENAI_API_KEY` override.
2. Run it on a host where the Paperclip server process has
`OPENAI_API_KEY` set.
3. Observe that the adapter process can inherit the host key unless
Paperclip persists a blocking empty override.
4. Set `adapterConfig.env.CODEX_HOME` to a shared path such as
`~/.codex` or the company-level `codex-home`.
5. Observe that the old code allowed the shared auth home instead of
returning a validation error.
**Paperclip version or commit**
- Reproduced by inspection against `master` before this PR.
**Deployment mode**
- Local dev / self-hosted server with `codex_local` agents.
**Installation method**
- Built from source.
**Agent adapter(s) involved**
- Codex.
**Database mode**
- Not database-related.
**Access context**
- Board and agent configuration paths.
**Relevant logs or output**
- No secret-bearing logs included.
**Relevant config**
- Unsafe shape: missing `adapterConfig.env.OPENAI_API_KEY`, or shared
`adapterConfig.env.CODEX_HOME`.
- Fixed shape: per-agent `CODEX_HOME` plus empty `OPENAI_API_KEY`
override.
**Additional context**
Related PR search for `codex_local OPENAI_API_KEY CODEX_HOME` found:
- #3681 `fix: preserve managed Codex auth and repo-root env loading`
- #5621 `fix: copy worktree codex auth locally`
Those are adjacent auth-handling changes, but they do not add the agent
create/update guard implemented here.
**Privacy checklist**
- [x] I have reviewed all pasted output for PII, usernames, file paths,
API keys, tokens, company names, and redacted where necessary.
## What Changed
- Added a `codex_local` config guard in agent create, hire, and update
routes.
- The guard assigns `adapterConfig.env.CODEX_HOME` to
`companies/<companyId>/agents/<agentId>/codex-home` when missing.
- The guard persists `adapterConfig.env.OPENAI_API_KEY = ""` when
missing, preventing host env inheritance.
- Shared `CODEX_HOME` values for the company codex-home, host
`$CODEX_HOME`, or `~/.codex` now fail with a 422 error.
- Added route tests for create, hire, update, and rejected shared host
Codex home.
- Updated `codex_local` and development docs to describe the per-agent
home contract.
## Verification
- `pnpm exec vitest run
server/src/__tests__/agent-adapter-validation-routes.test.ts`
- `pnpm exec vitest run
server/src/__tests__/agent-skills-routes.test.ts`
- `pnpm typecheck`
- `git diff --check upstream/master...HEAD`
- `gh pr list --repo paperclipai/paperclip --state all --search
"codex_local OPENAI_API_KEY CODEX_HOME" --limit 20 --json
number,title,state,url`
- `rg -n "codex|OPENAI_API_KEY|CODEX_HOME|adapter" ROADMAP.md` returned
no roadmap overlap.
## Risks
- Existing legacy `codex_local` agents with shared `CODEX_HOME` will get
a clear 422 when their adapter config is updated until the shared path
is replaced. This is intentional because silent fallback is the bug
being guarded.
- Low migration risk: no database migration and no secret values are
printed or persisted beyond the empty override.
## Model Used
- OpenAI GPT-5.5 Codex, Codex coding-agent session with repository tool
use.
## Checklist
- [x] I have included a thinking path that traces from project context
to this change
- [x] I have specified the model used (with version and capability
details)
- [x] I have checked ROADMAP.md and confirmed this PR does not duplicate
planned core work
- [x] I have searched GitHub for duplicate or related PRs and linked
them above
- [x] I have either (a) linked existing issues with `Fixes: #` / `Closes
#` / `Refs #` OR (b) described the issue in-PR following the relevant
issue template
- [x] I have run tests locally and they pass
- [x] I have added or updated tests where applicable
- [x] If this change affects the UI, I have included before/after
screenshots
- [x] I have updated relevant documentation to reflect my changes
- [x] I have considered and documented any risks above
- [ ] All Paperclip CI gates are green
- [ ] Greptile is 5/5 with no open P2s, recommendations, or follow-ups
- [x] I will address all Greptile and reviewer comments before
requesting merge
## Paperclip
- Issue: [ZOL-5477](/ZOL/issues/ZOL-5477)
- Owner: Разработчик (`6625498c-66c9-429f-b578-4463ddc3ba16`)
- Status: waiting reviewer
- Next action: merge after approval and green CI
---------
Co-authored-by: Paperclip <noreply@paperclip.ing>
## Thinking Path
> - Paperclip is the open source app people use to manage AI agents for
work
> - The `openclaw_gateway` adapter is how operators wire Paperclip
agents to an OpenClaw gateway over WebSocket
> - The adapter UI previously only exposed a handful of config fields in
edit mode; many timeout / auth / session-routing knobs were unreachable
through the form
> - The serializer also forgot to inject the configured `authToken` into
the `x-openclaw-token` header, and the server-side execute path lacked
retries on transient gateway errors and an `OPENCLAW_TOKEN` env fallback
> - This pull request exposes the full set of config fields in both
create and edit modes, fixes the serializer, hardens the server-side
execute path, and pins the existing default request timeouts (120s /
120000ms) — see the dedicated commit and the new unit tests
> - The benefit is operators can configure and reconfigure an
`openclaw_gateway` agent end-to-end through the UI, with no silent
change to the defaults documented in the adapter README and
`doc/ONBOARDING_AND_TEST_PLAN.md`
## Linked Issues or Issue Description
Closes#414Closes#1901Closes#2309
## What Changed
- **UI**: Removed the `!isCreate` guard so all `openclaw_gateway` config
fields are visible in both create and edit modes (`authToken`,
`agentId`, `sessionKeyStrategy`, `sessionKey`, `timeoutSec`,
`waitTimeoutMs`, `disableDeviceAuth`, `autoPairOnFirstConnect`, `role`,
`scopes`, `paperclipApiUrl`, `headersJson`, `payloadTemplate`,
`runtimeServices`).
- **Serialization**
(`packages/adapters/openclaw-gateway/src/ui/build-config.ts`): inject
`authToken` into headers as `x-openclaw-token`; apply safe defaults on
create (`timeoutSec=120`, `waitTimeoutMs=120000`,
`sessionKeyStrategy="issue"`, `role="operator"`,
`scopes=["operator.admin"]`).
- **Backend**
(`packages/adapters/openclaw-gateway/src/server/execute.ts`): add
`OPENCLAW_TOKEN` env-var fallback for `authToken`, retry logic (max 2
retries with backoff for transient gateway errors), session-key prefix
`agent:{agentId}:{sessionId}` when `agentId` is configured.
- **Defaults restoration** (dedicated commit): an earlier revision of
this PR lowered the default request timeouts to `60` / `30000`. The
current branch restores the historical `timeoutSec=120` /
`waitTimeoutMs=120000` defaults that match the values documented in
`packages/adapters/openclaw-gateway/src/index.ts`,
`src/server/execute.ts` on master, and the worked example in
`doc/ONBOARDING_AND_TEST_PLAN.md`.
- **Tests** (new):
`packages/adapters/openclaw-gateway/src/ui/build-config.test.ts` pins
the documented timeout and identity defaults so the silent-halve
regression cannot recur.
## Verification
- `pnpm --filter @paperclipai/adapter-openclaw-gateway typecheck`
- `pnpm typecheck` (root)
- Manual: create a new `openclaw_gateway` agent — all fields visible,
defaults populate as documented.
- Manual: edit an existing `openclaw_gateway` agent — every field
round-trips correctly and saves.
- Manual: unset `authToken` in the form and set `OPENCLAW_TOKEN` env var
— adapter picks up the env-var fallback.
- Manual: simulate a transient gateway error — execute retries up to 2
times with backoff before failing.
## Risks
- Low risk. Surface area is one adapter, behind explicit operator
configuration. The defaults change in this PR is a restoration of values
that already exist on master and in the adapter docs, so no production
agent sees a behavioral shift relative to the prior release. Field
exposure in edit mode is purely additive — existing values are preserved
on save.
## Model Used
- Provider/model: Claude (Anthropic) — `claude-opus-4-7`
- Mode: standard tool use, no extended thinking
- Capability notes: code execution + repository file edits via Claude
Code
## Cross-references and status (maintainer)
Closes#414Closes#1901Closes#2309
## Checklist
- [x] I have included a thinking path that traces from project context
to this change
- [x] I have specified the model used (with version and capability
details)
- [x] I have checked ROADMAP.md and confirmed this PR does not duplicate
planned core work
- [x] I have searched GitHub for duplicate or related PRs and linked
them above
- [x] I have either (a) linked existing issues with `Fixes: #` / `Closes
#` / `Refs #` OR (b) described the issue in-PR following the relevant
issue template
- [x] I have run tests locally and they pass
- [x] I have added or updated tests where applicable
- [ ] If this change affects the UI, I have included before/after
screenshots
- [x] I have updated relevant documentation to reflect my changes
- [x] I have considered and documented any risks above
- [ ] All Paperclip CI gates are green
- [ ] Greptile is 5/5 with no open P2s, recommendations, or follow-ups
- [x] I will address all Greptile and reviewer comments before
requesting merge
---------
Co-authored-by: Paperclip Bot <bot@paperclip.dev>
Co-authored-by: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
Co-authored-by: Devin Foley <devin@paperclip.ing>
Co-authored-by: Paperclip <noreply@paperclip.ing>
## Thinking Path
> - Paperclip is the open source app people use to manage AI agents for
work
> - The gemini-local adapter runs gemini-cli headlessly, including on
remote/sandboxed execution targets where the adapter manages a dedicated
HOME under the runtime root
> - gemini-cli hard-refuses headless runs with "Invalid auth method
selected." unless `$HOME/.gemini/settings.json` persists an auth
selection; setting `GEMINI_DEFAULT_AUTH_TYPE` alone does NOT satisfy it
(proven in an isolated pod)
> - With a managed HOME the runtime root replaces the image home, so any
settings.json baked into the agent image (or the user's real home) is
invisible to the CLI, and every sandboxed gemini run dies before doing
any work
> - This affects any sandbox provider that runs gemini with API-key auth
through the managed-HOME path (SSH, E2B, Daytona, Kubernetes, or any
other remote execution target); it is a headless-execution bug fix, not
gateway- or deployment-specific behavior
> - This pull request makes the adapter pre-select the `gemini-api-key`
auth type in the managed `$HOME/.gemini/settings.json` whenever a
Gemini/Google API key is present, writing both settings schema
generations and never touching an existing settings.json
> - The benefit is that gemini agents actually run headlessly on remote
and sandboxed execution targets without any manual settings provisioning
## Linked Issues or Issue Description
No existing issue; describing the bug in-PR (bug template fields):
- **What happened:** Headless gemini-local runs on remote/sandboxed
execution targets fail immediately with `Invalid auth method selected.`
even though `GEMINI_API_KEY` is provided.
- **Expected:** Providing the API key should be enough for a headless
run to authenticate and proceed.
- **Root cause:** gemini-cli requires an auth selection persisted in
`$HOME/.gemini/settings.json` for non-interactive runs; the
`GEMINI_DEFAULT_AUTH_TYPE` env var does not substitute for it (verified
in an isolated pod with only the env var set). The adapter's
managed-HOME execution path points HOME at the runtime root, so any
pre-existing settings.json (image-baked or user home) is hidden and the
CLI finds no auth selection.
- **Reproduction:** Run the gemini-local adapter against a
remote/sandboxed execution target with `GEMINI_API_KEY` set and no
settings.json under the managed HOME; the run aborts with the error
above.
- Duplicate/related search: no existing PR or issue addresses this;
closest related is #7693 (bundles gemini-cli in the Docker image), which
makes the CLI available but does not fix headless auth selection.
## What Changed
- `packages/adapters/gemini-local/src/server/execute.ts`: after
provisioning the managed HOME, when a Gemini/Google API key is present,
write `$HOME/.gemini/settings.json` pre-selecting `gemini-api-key` auth.
Both settings schema generations are written (legacy top-level
`selectedAuthType` and current `security.auth.selectedType`) so old and
new gemini-cli versions are covered.
- The write is strictly scoped to the managed HOME (the per-run runtime
root on sandbox transports). On non-managed remote targets (SSH), where
the remote home is the user's real home and existing settings remain
visible to the CLI, the adapter creates nothing (review feedback, P1).
- The write is guarded by `[ -f ... ] ||` so a user-shipped
settings.json (e.g. via workspace) is never overwritten.
- The key-presence gate checks the run env AND the host process env
(`GEMINI_API_KEY` / `GOOGLE_API_KEY`): in sandboxed paths the key never
enters the adapter's run env; it reaches the agent pod via the sandbox
provider's per-run secret (env passthrough from the host env), so the
host env is the correct signal there.
- `packages/adapters/gemini-local/src/server/execute.remote.test.ts`: a
new sandbox-transport test asserts the settings.json write lands under
the per-run runtime root (path + `gemini-api-key` content), and the SSH
test asserts no settings.json is created on a non-managed home.
## Verification
- `npx vitest run packages/adapters/gemini-local`: 3 files, 17 tests,
all pass.
- `pnpm --filter @paperclipai/adapter-gemini-local typecheck` and
`build`: clean (test file is covered by the package tsconfig `include`).
- Negative control: in an isolated pod, gemini-cli with `GEMINI_API_KEY`
+ `GEMINI_DEFAULT_AUTH_TYPE` set but no settings.json still fails with
`Invalid auth method selected.`; with the settings.json written by this
change, the run proceeds.
- Verified end-to-end: a gemini agent in a hardened Kubernetes (gVisor)
sandbox completed a real task (with `GOOGLE_GEMINI_BASE_URL` pointing at
a GenAI-compatible endpoint), producing a billed usage row. That
deployment supplies the verification evidence; the fix applies to any
sandbox provider running gemini with API-key auth.
## Risks
- Low risk. The new write only fires on the managed-HOME path (per-run
runtime root) when an API key is present, and only when no settings.json
exists yet, so existing setups, real user homes on SSH targets, and
user-provided settings are unaffected.
- If a future gemini-cli changes the settings schema again, the file may
need a third generation key; both current generations are written today.
## Model Used
- Claude (Anthropic), Claude Opus 4.8, 1M context, extended thinking,
with tool use (code execution / shell) via Claude Code.
## Checklist
- [x] I have included a thinking path that traces from project context
to this change
- [x] I have specified the model used (with version and capability
details)
- [x] I have checked ROADMAP.md and confirmed this PR does not duplicate
planned core work
- [x] I have searched GitHub for duplicate or related PRs and linked
them above
- [x] I have either (a) linked existing issues with `Fixes: #` / `Closes
#` / `Refs #` OR (b) described the issue in-PR following the relevant
issue template
- [x] I have run tests locally and they pass
- [x] I have added or updated tests where applicable
- [ ] If this change affects the UI, I have included before/after
screenshots (no UI change)
- [x] I have updated relevant documentation to reflect my changes (code
comments document the behavior; no doc pages cover managed-home auth)
- [x] I have considered and documented any risks above
- [x] All Paperclip CI gates are green
- [ ] Greptile is 5/5 with no open P2s, recommendations, or follow-ups
(review requested)
- [x] I will address all Greptile and reviewer comments before
requesting merge
🤖 Generated with [Claude Code](https://claude.com/claude-code)
---------
Co-authored-by: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
## Thinking Path
> - Paperclip is the open source app people use to manage AI agents for
work
> - The `codex-local` adapter runs the OpenAI Codex CLI; Paperclip
already maintains a managed `CODEX_HOME` per company and ships it to
remote/sandboxed execution targets
> - Deployments increasingly put an OpenAI-compatible LLM gateway
between the harness and the model for cost, governance, or
data-residency reasons: LiteLLM, OpenRouter, Portkey, Kong, a corporate
proxy, self-hosted models (vLLM/Ollama), or region-pinned/sovereign
endpoints. But Codex has no CLI flag or env var for a custom endpoint:
its only mechanism is `[model_providers.<id>]` tables (with `base_url`,
`env_key`, `wire_api`) in `$CODEX_HOME/config.toml`, selected by a
root-level `model_provider` key
> - Today there is no supported way to get such provider config into the
managed `CODEX_HOME`, so gateway routing requires hand-editing files the
adapter owns and regenerates
> - This pull request adds the codex analogue of #7837's opencode
mechanism: a `PAPERCLIP_CODEX_PROVIDERS` JSON env var whose shape maps
1:1 onto codex's TOML schema, merged into the managed `config.toml` so
the existing asset-shipping + `env.CODEX_HOME` mechanics deliver it to
local and sandboxed runs alike; nothing here is specific to one hosting
setup
> - The benefit is Codex works behind any OpenAI-compatible gateway with
config only; with no env set, behavior is unchanged
## Linked Issues or Issue Description
No existing issue; describing in-PR (feature / adapter enhancement).
- **Gap:** there is no supported way to register a custom/gateway
`[model_providers.*]` endpoint for `codex-local`. Codex's only
custom-endpoint mechanism is `config.toml` (`base_url` + `env_key` +
`wire_api`, selected via the root `model_provider` key), and the adapter
owns/regenerates the managed `CODEX_HOME`, so operators cannot durably
hand-edit it.
- Related: #7837 (the opencode-local analogue of this change, same
env-driven gateway-routing pattern). Searched for duplicate/related PRs:
no existing codex-local gateway/provider-routing PR found.
> Note on ROADMAP: this is adapter-level, opt-in config (defaults
unchanged) that *enables* gateway routing for one harness; it is not the
core "Cloud / Sandbox agents" platform work itself.
## What Changed
- New `prepareCodexRuntimeConfig()`
(`packages/adapters/codex-local/src/server/runtime-config.ts`): reads
`PAPERCLIP_CODEX_PROVIDERS` (run env first, then `process.env`), shaped
as `{"providers": {"<id>": {base_url, env_key, wire_api, ...}},
"model_provider": "<id>"}`, and merges it into the managed
`CODEX_HOME`'s `config.toml`. No-op when unset or empty.
- A malformed value (invalid JSON, not a JSON object, no `providers`
object, no usable provider entries, or individual entries with empty
names or non-object values, which are skipped by name) is never silently
dropped: each case surfaces a distinct, user-visible note (via the
prepare notes, which flow into command notes + `onLog`) and unusable
input leaves `config.toml` untouched.
- Merge is marker-delimited and TOML-correct: existing `config.toml`
content is preserved between two managed blocks. Root keys (e.g.
`model_provider`) are prepended **before the first table header** (TOML
root-region rule), `[model_providers.*]` tables are appended.
Pre-existing same-name provider sections and root `model_provider` keys
are excised so the managed definitions win without duplicate-table parse
errors.
- `{env:VAR}` placeholders are expanded server-side for
literal-credential fields; `env_key` indirection remains the preferred
path.
- Crash-safe restore: prepare writes a pre-run backup
(`config.toml.paperclip-backup`) before the merged file; `cleanup()`
restores the original in the execute `finally` and removes the backup.
If a run never reaches `cleanup()` (a throw during the setup between
prepare and execution, or SIGKILL), the next prepare restores the
original from the backup with full fidelity, including user
`[model_providers.*]` sections the merge excised (review feedback, P2);
plain block-stripping remains the fallback for pre-backup state.
- An explicit adapter-config `env.CODEX_HOME` override is treated as
user-managed: no merge, surfaced as a command note.
- Dependency-free hand-emitted TOML (strings/numbers/booleans, arrays of
scalars, plain objects as inline tables); basic strings escape
U+0000-U+001F and U+007F per TOML 1.0 (review feedback, P2). Merged
output was additionally validated locally with python tomllib during
development; the committed tests assert the structural invariants.
- `execute.ts` wiring: `prepareCodexRuntimeConfig` runs after
`prepareManagedCodexHome` (before the home ships to the remote target),
notes surface via `onLog` + command notes, and the `finally` calls
`cleanup()`.
**Note for reviewers:** current codex removed `wire_api = "chat"`
(openai/codex#10157, Feb 2026), so gateway provider configs must use
`wire_api = "responses"`, i.e. the gateway must speak `/v1/responses`.
The adapter passes the value through verbatim; this is a codex-side
constraint worth knowing when configuring it.
## Verification
- `pnpm --filter @paperclipai/adapter-codex-local build` and
`typecheck`: tsc clean against current `master`
- `pnpm exec vitest run packages/adapters/codex-local`: 45 passing
(incl. 17 `runtime-config` tests: fresh-merge + cleanup restore,
root-region placement, same-name provider override, inline
tables/arrays, DEL escaping, `{env:}` expansion from run env +
`process.env`, per-case malformed-input notes with `config.toml`
untouched, skipped-entry notes alongside a successful merge, silent
no-op when unset/empty, explicit-`CODEX_HOME` skip note, backup restore
of excised user sections after an interrupted run, backup removal on
cleanup, stale-block self-heal, re-run replacement)
- Verified end-to-end: a codex agent in a hardened Kubernetes (gVisor)
sandbox completed a real task routed through an OpenAI-compatible
gateway's `/v1/responses`, with a billed usage row recorded on the
gateway. That deployment supplies the verification evidence; the
mechanism is gateway-agnostic.
## Risks
Low. Entirely env-driven and opt-in; with `PAPERCLIP_CODEX_PROVIDERS`
unset the adapter never touches `config.toml` and behavior is
byte-identical to before. The merge preserves user content, restores the
original file on cleanup, and survives interrupted runs via the pre-run
backup; malformed input surfaces a visible note and is ignored without
touching `config.toml`. No migration/UI impact.
## Model Used
Claude Opus 4.8 (`claude-opus-4-8`, 1M context), extended thinking +
tool use, via Claude Code.
## Checklist
- [x] I have included a thinking path that traces from project context
to this change
- [x] I have specified the model used (with version and capability
details)
- [x] I have checked ROADMAP.md (adapter-level opt-in config enabling
gateway routing; not the core sandbox-platform work, noted above)
- [x] I have searched GitHub for duplicate or related PRs and linked
them above (#7837 is the opencode analogue; no codex-local duplicate
found)
- [x] I have either (a) linked existing issues OR (b) described the
issue in-PR following the relevant issue template
- [x] I have run tests locally and they pass
- [x] I have added or updated tests where applicable
- [ ] If this change affects the UI, I have included before/after
screenshots (n/a, no UI)
- [ ] I have updated relevant documentation to reflect my changes (env
var documented inline; no central doc references the adapter env yet)
- [x] I have considered and documented any risks above
- [x] All Paperclip CI gates are green (green on the previous head;
re-running on the final note-copy polish commit)
- [x] Greptile is 5/5 with no open P2s, recommendations, or follow-ups
(both review P2s are fixed at head: the interrupted-run restore via the
pre-run backup and the U+007F escaping; a re-review is requested for the
note-copy polish)
- [x] I will address all Greptile and reviewer comments before
requesting merge
🤖 Generated with [Claude Code](https://claude.com/claude-code)
---------
Co-authored-by: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
## Thinking Path
> - Paperclip is the open source app people use to manage AI agents for
work
> - The `pi-local` adapter runs the Pi coding agent, including inside
remote/sandboxed execution targets; Pi resolves `--provider P --model M`
by an exact (provider, id) match against its model registry, and it has
no base-url CLI flag or env var: a `models.json` in its agent config dir
(`$PI_CODING_AGENT_DIR`, falling back to `$HOME/.pi/agent`) is its only
mechanism for custom or OpenAI/Anthropic-compatible endpoints
> - Deployments increasingly put an LLM gateway between the harness and
the model for cost, governance, or data-residency reasons: LiteLLM,
OpenRouter, Portkey, Kong, a corporate proxy, self-hosted models
(vLLM/Ollama), or region-pinned/sovereign endpoints. Today there is no
supported way to get such provider config into Pi's registry for
orchestrated runs
> - The opencode adapter gained the equivalent capability in #7837 and
codex in #7919; this pull request is the Pi analogue, so the harness
layer stays gateway-agnostic regardless of which CLI an agent uses;
nothing here is specific to one hosting setup
> - This pull request reads `PAPERCLIP_PI_PROVIDERS` (Pi's `models.json`
`providers` shape), materialises a managed `models.json` in a temp
agent-config dir, points `PI_CODING_AGENT_DIR` at it, and ships it to
remote execution targets with the run
> - The benefit is Pi works behind any compatible gateway with config
only; with no env set, behavior is unchanged
## Linked Issues or Issue Description
No existing issue; describing in-PR (feature / adapter enhancement).
- **Gap:** there is no supported way to register custom/gateway
providers + models for `pi-local`. Pi's only custom-endpoint mechanism
is a `models.json` in its agent config dir, and orchestrated (especially
sandboxed) runs have no way to provision one declaratively.
- Related: #7837 (the opencode-local analogue, same env-driven
gateway-routing pattern) and #7919 (the codex-local analogue). Searched
for duplicate or related PRs: no existing pi-local
gateway/provider-routing PR found.
> Note on ROADMAP: this is adapter-level, opt-in config (defaults
unchanged) that *enables* gateway routing for one harness; it is not the
core "Cloud / Sandbox agents" platform work itself.
## What Changed
- New `packages/adapters/pi-local/src/server/runtime-config.ts`:
`preparePiRuntimeConfig()` reads `PAPERCLIP_PI_PROVIDERS` (a JSON object
in pi's `models.json` `providers` shape) from the run env, then
`process.env`. When set, it expands `{env:VAR}` placeholders (run env
first, then process env; unresolvable placeholders left intact), writes
`{"providers": ...}` to a managed temp dir as `models.json`, and returns
env with `PI_CODING_AGENT_DIR` pointing at it plus a cleanup handle.
- `execute.ts`: the prepared dir ships to remote execution targets as
the managed-runtime asset `agentConfig` (same mechanism as opencode's
`xdgConfig`), and `PI_CODING_AGENT_DIR` is repointed to the in-target
path; cleanup runs in `finally`.
- Misconfiguration is visible, not silent: a set-but-unusable
`PAPERCLIP_PI_PROVIDERS` (invalid JSON, not an object, no provider
objects) surfaces an explanatory note instead of proceeding unconfigured
into an opaque model-not-found failure later, and provider entries with
non-object values are skipped with a note naming them. Unset/empty stays
a silent no-op (feature off).
- Defaults unchanged: with `PAPERCLIP_PI_PROVIDERS` unset, the adapter
behaves byte-for-byte as before, for local runs and for every existing
sandbox provider.
## Verification
- All pi-local tests green against this base (new: providers written
verbatim, `{env:VAR}` expansion from run env/process env/unresolvable,
no-op when unset, `PI_CODING_AGENT_DIR` set and shipped, the
misconfiguration notes incl. skipped non-object entries, remote asset
sync + env repoint). Typecheck and build clean.
- Production end-to-end evidence (our deployment, used as verification,
not as the scope of the change): a pi agent in a Kubernetes gVisor
sandbox resolved a custom provider from the shipped `models.json`,
completed an assigned issue through an Anthropic-compatible gateway, and
landed a billed usage row.
## Risks
Low. The entire feature is opt-in behind one env var; the only behavior
change when it is set is the intended one. The managed dir replaces the
host agent dir for the run by design (credentials travel inside the
provider config or via env-key indirection), which is the correct
posture for orchestrated runs. No migration/UI impact.
## Model Used
Claude Opus 4.8 (`claude-opus-4-8`, 1M context), extended thinking +
tool use, via Claude Code.
## Checklist
- [x] I have included a thinking path that traces from project context
to this change
- [x] I have specified the model used (with version and capability
details)
- [x] I have checked ROADMAP.md (adapter-level opt-in config enabling
gateway routing; not the core sandbox-platform work, noted above)
- [x] I have searched GitHub for duplicate or related PRs and linked
them above (#7837 and #7919 are the opencode/codex analogues; no
pi-local duplicate found)
- [x] I have either (a) linked existing issues OR (b) described the
issue in-PR following the relevant issue template
- [x] I have run tests locally and they pass
- [x] I have added or updated tests where applicable
- [ ] If this change affects the UI, I have included before/after
screenshots (n/a, no UI)
- [ ] I have updated relevant documentation to reflect my changes (env
var documented inline; no central doc references the adapter env yet)
- [x] I have considered and documented any risks above
- [x] All Paperclip CI gates are green (green on the previous head;
re-running on the final note-copy polish commit)
- [x] Greptile is 5/5 with no open P2s, recommendations, or follow-ups
(both prior review findings are fixed at head: the indirect notes-based
guard is now an explicit `agentConfigDir` handle, and a failed
`models.json` write no longer leaks the temp dir; a re-review is
requested for the note-copy polish)
- [x] I will address all Greptile and reviewer comments before
requesting merge
🤖 Generated with [Claude Code](https://claude.com/claude-code)
---------
Co-authored-by: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
## Thinking Path
> - Paperclip is the open source app people use to manage AI agents for
work
> - The `opencode-local` adapter runs the OpenCode harness; its
model/provider routing assumes built-in providers (anthropic/openai/...)
and their default models
> - Deployments increasingly put an OpenAI/Anthropic-compatible LLM
gateway between the harness and the model for cost, governance, or
data-residency reasons: LiteLLM, OpenRouter, Portkey, Kong, a corporate
proxy, self-hosted models (vLLM/Ollama), or region-pinned/sovereign
endpoints. But OpenCode only resolves `--model provider/model` when the
model is registered in a provider's `models` map, and
`OPENCODE_ALLOW_ALL_MODELS` does NOT bypass its internal `getModel()`
> - Several lanes also fall back to built-in default models the gateway
may not serve: the auxiliary/title model (e.g. `claude-haiku-*`) and the
budget/recovery "cheap" lane (`openai/gpt-5.1-codex-mini`); these abort
runs with "no keys found that support model"
> - This pull request makes the adapter's provider/model wiring
declarative via env, so any such deployment can register gateway models
+ pin the auxiliary/budget lanes without code changes; nothing here is
specific to one hosting setup
> - The benefit is OpenCode works behind any compatible gateway with
config only; with no env set, behavior is unchanged
## Linked Issues or Issue Description
No existing issue; describing in-PR (feature / adapter enhancement).
- **Gap:** there is no supported way to register custom/gateway
providers + models for `opencode-local`, nor to pin the auxiliary
(title-gen) and budget (recovery) model lanes, so routing OpenCode
through a gateway fails at `getModel()` or on the default helper models.
- Related: #5737 (exe.dev sandbox installs for gemini/opencode local),
#5823 (unblock claude_local on remote sandbox providers).
> Note on ROADMAP: this is adapter-level, opt-in config (defaults
unchanged) that *enables* gateway routing for one harness; it is not the
core "Cloud / Sandbox agents" platform work itself. Happy to
redirect/discuss in #dev if preferred.
## What Changed
- `PAPERCLIP_OPENCODE_PROVIDERS`: merge custom/extended providers
(OpenCode `provider` shape) into the runtime `opencode.json`, so gateway
models are registered and `--model provider/model` resolves. `{env:VAR}`
placeholders are expanded server-side (so a key need not depend on the
sandbox run env).
- A malformed `PAPERCLIP_OPENCODE_PROVIDERS` is no longer silently
ignored: invalid JSON, a non-object value, and individual provider
entries with non-object values (which are skipped by name) each append a
visible note to the run notes so the misconfiguration is diagnosable
(addresses both review P1s).
- `PAPERCLIP_OPENCODE_SMALL_MODEL` / `PAPERCLIP_OPENCODE_CHEAP_MODEL`:
pin the auxiliary (title-generation) and budget (recovery-retry) lanes
to gateway-served models; defaults unchanged.
- Honour `OPENCODE_ALLOW_ALL_MODELS` on the **remote** execution path
too (was local-only, a parity gap).
- `PAPERCLIP_OPENCODE_PRINT_LOGS`: optional toggle adding `--print-logs`
so OpenCode logs surface on stderr for diagnosing remote/sandbox runs.
- `buildOpenCodeModelProfiles()` guards its `process.env` default with
`typeof process` so the shared client/server module stays browser-safe
(a bare `process.env` at module load threw ReferenceError in the browser
under Vite dev middleware and broke UI rendering in the e2e lane).
## Verification
- `pnpm --filter @paperclipai/adapter-opencode-local build` and
`typecheck` (tsc clean)
- `pnpm exec vitest run packages/adapters/opencode-local/src` shows 33
passing (incl. new tests for the provider merge, `{env:}` expansion, the
malformed/non-object/skipped-entry provider notes, small/cheap-model
resolution, and the remote allow-all bypass)
- Manually verified end-to-end against a real
OpenAI-/Anthropic-compatible gateway: with the providers + small/cheap
model set, both the title-gen and main task route to the configured
gateway model and the agent completes (a real completion is returned and
billed). That deployment supplies the verification evidence; the
mechanism is gateway-agnostic.
## Risks
Low. Everything is env-driven and opt-in; with no env set the generated
config output is unchanged, and the cheap model profile keeps its model
(the only difference is its updated human-readable description).
Defaults preserved: built-in providers, Codex-mini cheap lane with
`variant: low`, no `--print-logs`. No migration/UI impact.
## Model Used
Claude Opus 4.8 (`claude-opus-4-8`, 1M context), extended thinking +
tool use, via Claude Code.
## Checklist
- [x] I have included a thinking path that traces from project context
to this change
- [x] I have specified the model used (with version and capability
details)
- [x] I have checked ROADMAP.md (adapter-level opt-in config enabling
gateway routing; not the core sandbox-platform work, noted above)
- [x] I have searched GitHub for duplicate or related PRs and linked
them above (#5737, #5823)
- [x] I have either (a) linked existing issues OR (b) described the
issue in-PR following the relevant issue template
- [x] I have run tests locally and they pass
- [x] I have added or updated tests where applicable
- [ ] If this change affects the UI, I have included before/after
screenshots (n/a, no UI)
- [ ] I have updated relevant documentation to reflect my changes (env
vars documented inline via comments; no central doc references the
adapter env yet)
- [x] I have considered and documented any risks above
- [x] All Paperclip CI gates are green
- [x] Greptile is 5/5 with no open P2s, recommendations, or follow-ups
(the P1 about silently dropped malformed providers JSON is addressed in
6eeb803, the follow-up P1 about silently skipped non-object entries in
2f4045a; the latest review has no further findings, and a re-review is
requested for the final note-copy/test-fixture polish at head)
- [x] I will address all Greptile and reviewer comments before
requesting merge
🤖 Generated with [Claude Code](https://claude.com/claude-code)
---------
Co-authored-by: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
## Thinking Path
> - Paperclip orchestrates AI agents through pluggable local adapters;
codex_local wraps OpenAI's `codex` CLI.
> - The codex_local adapter declares a hard-coded
`DEFAULT_CODEX_LOCAL_MODEL = "gpt-5.3-codex"` and multiple Paperclip
consumers (UI build-config, server route, OnboardingWizard, NewAgent
form, AgentConfigForm) fall back to it when the operator doesn't pick a
model.
> - That model — and every `*-codex` model plus the older
`gpt-5/5.1/5.2` lines — is API-key-only. Codex CLI rejects them on
ChatGPT subscription auth with "The 'gpt-5.3-codex' model is not
supported when using Codex with a ChatGPT account."
> - Every codex_local agent created through the default onboarding path
inherits this pin and breaks on its first heartbeat for any user authed
via `codex login` (ChatGPT).
> - claude_local already takes the right shape: its build-config only
sets `adapterConfig.model` when the operator actually picked one, and
falls through to whatever default `claude` CLI uses.
> - Codex CLI's own default is auth-mode-aware. ChatGPT-subscription
accounts get `gpt-5.5`; API-key accounts get the codex-tuned default. A
Paperclip-side pin masks this and downgrades whichever group it wasn't
built for.
> - This PR makes codex_local match claude_local's shape: omit
`adapterConfig.model` when the user picks "default," and let the CLI
choose. Subscription users stop breaking; API-key users stop getting
downgraded.
> - The benefit is auth-mode-correct defaults with no Paperclip-side
hard pin, plus future-proofing: when OpenAI bumps the CLI default we
inherit it for free.
## What Changed
- `packages/adapters/codex-local/src/ui/build-config.ts` — only set
`adapterConfig.model` when the operator picked one (parity with
`packages/adapters/claude-local/src/ui/build-config.ts`).
- `server/src/routes/agents.ts` — drop the codex_local-specific
`next.model = DEFAULT_CODEX_LOCAL_MODEL` fallback in
`applyCreateDefaultsByAdapterType`. Bypass-sandbox default is left in
place (security posture, not a model choice).
- `ui/src/pages/NewAgent.tsx`, `ui/src/components/AgentConfigForm.tsx`,
`ui/src/components/OnboardingWizard.tsx` — stop pre-populating the model
field with `DEFAULT_CODEX_LOCAL_MODEL` when the user selects the Codex
adapter. Other adapters' defaults (gemini_local, cursor, opencode_local)
are unchanged.
- `DEFAULT_CODEX_LOCAL_MODEL` is preserved as an exported constant for
downstream consumers / plugin authors who want to opt in to a pin; we
just stop forcing it on operators who didn't ask for one.
- Test: assert `buildCodexLocalConfig` omits `model` when input is
blank.
## Verification
- `pnpm exec vitest run
packages/adapters/codex-local/src/ui/build-config.test.ts
packages/adapters/codex-local/src/server/codex-args.test.ts
server/src/__tests__/adapter-registry.test.ts
server/src/__tests__/heartbeat-model-profile.test.ts
server/src/__tests__/agent-permissions-routes.test.ts` → 74/74 passing
- `pnpm exec vitest run ui/src/lib/duplicate-agent-payload.test.ts
ui/src/lib/acpx-model-filter.test.ts` → passing
- `pnpm tsc --noEmit -p .` → clean
- Live: I separately verified live during initial investigation that on
ChatGPT-subscription auth, `gpt-5.3-codex` is rejected and `gpt-5.5` is
what Codex CLI picks by default. Omitting model lets the CLI handle
that.
## Risks
- Telemetry: any sink that reads `adapterConfig.model` for cost
attribution will now see the empty/omitted case more often. The CLI
emits the actually-used model in its event stream; downstream telemetry
should already read from there for accuracy, but worth a check.
- Operator UX: "default" now means "whatever the CLI picks" instead of a
Paperclip-known model. The selectable catalog still includes `gpt-5.5`,
`gpt-5.4`, `gpt-5.3-codex`, etc. for operators who want to pin
explicitly.
- Existing agents are unaffected — their `adapterConfig.model` is
already set; this only changes the *new-agent* default flow.
## Related work
- Depends on: an open catalog-add PR adding `gpt-5.5` to the selectable
model list and to `CODEX_LOCAL_FAST_MODE_SUPPORTED_MODELS`. Operators
who want to switch to `gpt-5.5` explicitly need that PR merged first;
this PR is the structural change that makes "default" mean "let the CLI
choose."
- Closes#5371 — codex_local default model selection persists
`gpt-5.3-codex` instead of adapter default (this PR is the exact fix
#5371 proposes).
- Related: #5132 (opencode-local: hire-time default model fails on
ChatGPT-OAuth accounts) — same problem shape on a sibling adapter; not
fixed here but worth tracking for a parallel.
- Related: #5939 (codex_local adapter hardcodes `gpt-5.3-codex-spark`
validation, fails on ChatGPT OAuth accounts regardless of configured
model) — separate validation-path bug; not fixed here.
## Model Used
Claude (Sonnet-class), running inside Paperclip as a claude_local
executor.
## Checklist
- [x] I have included a thinking path that traces from project context
to this change
- [x] I have specified the model used (with version and capability
details)
- [x] I have checked ROADMAP.md and confirmed this PR does not duplicate
planned core work
- [x] I have searched GitHub for duplicate or related PRs and linked
them above
- [x] I have either (a) linked existing issues with `Fixes: #` / `Closes
#` / `Refs #` OR (b) described the issue in-PR following the relevant
issue template
- [x] I have run tests locally and they pass
- [x] I have added or updated tests where applicable
- [x] If this change affects the UI, I have included before/after
screenshots
- [x] I have updated relevant documentation to reflect my changes
- [x] I have considered and documented any risks above
- [x] All Paperclip CI gates are green
- [x] Greptile is 5/5 with no open P2s, recommendations, or follow-ups
- [x] I will address all Greptile and reviewer comments before
requesting merge
Co-authored-by: Paperclip <noreply@paperclip.ing>
## Related work
This PR is the cleanest "add `gpt-5.5` to the codex-local catalog"
change open against master. Several other PRs propose the same
catalog/fast-mode update; they should close as duplicates once this
lands:
- #4646 — Add Codex gpt-5.5 model option
- #6044 — feat(codex-local): add gpt-5.5 to model catalog, default
reasoning to medium, cheap profile xhigh
- #6045 — feat(codex-local): add gpt-5.5 to model catalog, default
medium reasoning, xhigh cheap profile
- #6595 — feat(adapters): add new Codex models (gpt-5.5, gpt-5.4-mini,
gpt-5.3-codex, gpt-5.2)
Related issues this enables (catalog-level surface area):
- #5371 — codex_local default model selection persists `gpt-5.3-codex`
instead of adapter default. This PR makes `gpt-5.5` selectable in the
dropdown; a separate follow-up changes the *default* behavior so users
who don't pick a model are subscription-compatible.
- #5132 — opencode-local: hire-time default model fails on ChatGPT-OAuth
accounts. Sibling adapter, same problem shape; not fixed here but worth
tracking as a parallel for the opencode side.
---
## Thinking Path
> - Paperclip orchestrates AI agents through adapter-backed local and
remote runtimes.
> - The `codex_local` adapter declares built-in model options that feed
the server model list and, in turn, the agent configuration UI dropdown.
> - GPT-5.5 is available in newer Codex environments but was missing
from Paperclip's fallback `codex_local` model list.
> - Operators could still type a manual model ID, but the default
dropdown made the supported path look unavailable.
> - Codex fast mode support is declared separately, so adding GPT-5.5 to
the visible list should also include it in the supported fast-mode set.
> - This pull request adds GPT-5.5 to the built-in Codex local model
options and updates focused tests around argument generation and adapter
model listing.
> - The benefit is a clearer default setup path for agents using GPT-5.5
without changing existing defaults or migrations.
## What Changed
- Added `gpt-5.5` to the `codex_local` fallback model list.
- Added `gpt-5.5` to `CODEX_LOCAL_FAST_MODE_SUPPORTED_MODELS`.
- Updated Codex argument tests to cover GPT-5.5 fast mode and preserve
manual-model fast mode behavior.
- Updated adapter model listing tests to assert the Codex fallback list
includes GPT-5.5.
## Verification
- `pnpm exec vitest run
packages/adapters/codex-local/src/server/codex-args.test.ts
server/src/__tests__/adapter-models.test.ts`
- `git diff --check`
- UI note: this is a dropdown data-source change rather than a
layout/component change; the adapter model listing test covers the list
consumed by the UI.
## Risks
- Low risk. This only extends a static fallback model list and fast-mode
allowlist.
- Existing defaults remain unchanged (`gpt-5.3-codex`).
- If a local Codex CLI does not support `gpt-5.5`, selecting it will
still fail at execution time the same way any unavailable manual model
would.
> For core feature work, check [`ROADMAP.md`](ROADMAP.md) first and
discuss it in `#dev` before opening the PR. Feature PRs that overlap
with planned core work may need to be redirected — check the roadmap
first. See `CONTRIBUTING.md`.
## Model Used
- OpenAI Codex desktop coding agent, GPT-5-family model. The exact
backing model ID was not exposed by the local runtime; the session used
shell, Git, test execution, and GitHub CLI tool access.
## Checklist
- [x] I have included a thinking path that traces from project context
to this change
- [x] I have specified the model used (with version and capability
details)
- [x] I have checked ROADMAP.md and confirmed this PR does not duplicate
planned core work
- [x] I have searched GitHub for duplicate or related PRs and linked
them above
- [x] I have either (a) linked existing issues with `Fixes: #` / `Closes
#` / `Refs #` OR (b) described the issue in-PR following the relevant
issue template
- [x] I have run tests locally and they pass
- [x] I have added or updated tests where applicable
- [x] If this change affects the UI, I have included before/after
screenshots
- [x] I have updated relevant documentation to reflect my changes
- [x] I have considered and documented any risks above
- [x] All Paperclip CI gates are green
- [x] Greptile is 5/5 with no open P2s, recommendations, or follow-ups
- [x] I will address all Greptile and reviewer comments before
requesting merge
---------
Co-authored-by: apple <apple@appledeMacBook-Pro.local>
Co-authored-by: Paperclip <noreply@paperclip.ing>
Co-authored-by: Devin Foley <devin@paperclip.ing>
## Thinking Path
The same 2026-04-30 audit that produced PR #4118 (`Invalid session`
regex extension) and the ENOTFOUND classifier (#4931) identified a third
stuck-session pattern: **13 failures in 7 days, all on a single agent
(Ernest)**, with stderr matching:
```
_ApiError: {"error":{"code":400,"message":"The input token count exceeds
the maximum number of tokens allowed 1048576","status":"INVALID_ARGUMENT"}}
at ChatCompressionService.compress
```
The root cause is that gemini-cli's `ChatCompressionService` blew the 1M
token context limit **during its compression step itself**. Resuming the
same session ID will hit the same wall on the next attempt — the session
is effectively dead the same way it is when "Invalid session identifier"
fires (PR #4118).
## What Changed
Extends the `isGeminiUnknownSessionError` regex in `parse.ts` with two
phrases:
- `exceeds\s+the\s+maximum\s+number\s+of\s+tokens`
- `input\s+token\s+count\s+exceeds`
Both trigger the **existing** fresh-session retry path in
`execute.ts:596` — no new code path. Same extension pattern as PR #4118.
## Verification
- `npx vitest run --project @paperclipai/adapter-gemini-local` → 14/14
pass (11 in `parse.test.ts` + 3 existing in `execute.remote.test.ts`)
- 2 new tests cover the token-overflow patterns
- `pnpm --filter @paperclipai/adapter-gemini-local typecheck` → clean
- Audit query against `heartbeat_runs.stderr_excerpt` confirms regex
matches all 13 occurrences
## Stacking
This PR is stacked on top of #4931 (the ENOTFOUND classifier) which adds
the `parse.test.ts` file. If #4931 merges first, this PR's diff is just
the regex + 2 tests. If this PR is reviewed first, please merge #4931
first to avoid touching the same test scaffolding twice.
## Risks
- **Low.** Single-line regex extension. No new code paths.
- The session-reset path is well-trodden (PR #4118 in flight).
- If a non-Gemini caller produces a stderr containing "exceeds the
maximum number of tokens" by coincidence, they would trigger one
unnecessary fresh-session retry. Not plausible in the gemini-cli output
context where this stderr is sourced.
## Model Used
Claude Opus 4.7 (1M context), Anthropic SDK via Claude Code CLI.
## Checklist
- [x] Thinking path traces from audit data to single-line regex change
- [x] Model specified
- [x] No duplicate of planned core work
- [x] Tests pass locally
- [x] Tests added (2 new)
- [x] N/A — server-side regex
- [x] Internal pattern; no docs change
- [x] Risks documented
- [x] Will address Greptile + reviewer comments before merge
- [x] I searched the GitHub PR list for similar PRs and confirmed this
is not a duplicate (related: #4118 covers the "Invalid session
identifier" regex; this PR extends the same regex with token-overflow
phrases)
🤖 Generated with [Claude Code](https://claude.com/claude-code)
---------
Co-authored-by: Devin Foley <devin@paperclip.ing>
Co-authored-by: Paperclip <noreply@paperclip.ing>
## Thinking Path
> - The `codex-local` adapter sets up a per-company Codex home with an
auth symlink. Between `lstat` and `symlink` there is a race where two
concurrent setups can both try to create the same symlink, surfacing
`EEXIST`.
> - Master already handles this at runtime via `createExpectedSymlink`,
which accepts `EEXIST` only when the raced-in entry resolves to the
expected source, and ships a regression test for the tolerated-race path
(symlink already points at the right place).
> - The symmetric path — `EEXIST` raised by a symlink pointing somewhere
else — must stay strictly rejected so a future refactor cannot silently
weaken the guard.
> - This PR locks that in with a single additive test. No production
code change.
## What Changed
- Added one regression test in
`packages/adapters/codex-local/src/server/codex-home.test.ts` that
injects an `EEXIST` whose raced-in symlink target points at a different
file, and asserts:
- `prepareManagedCodexHome` rejects with `code: "EEXIST"`.
- The mismatched symlink is left on disk (we do not blindly overwrite
the raced-in entry).
Complements the existing "treats a concurrently-created expected auth
symlink as success" test already on master.
Refs #5240 (Stack B — codex-home adapter session/auth handling).
## Verification
- `pnpm --filter @paperclipai/adapter-codex-local exec vitest run
src/server/codex-home.test.ts` — passes.
- `pnpm --filter @paperclipai/adapter-codex-local typecheck` — clean.
## Risks
- Test-only change. No production code is modified.
## Model Used
- Provider: Anthropic
- Model: Claude (Opus 4.7)
- Mode/capabilities: tool-using coding agent with shell execution and
test verification
## Checklist
- [x] I have included a thinking path that traces from project context
to this change
- [x] I have specified the model used (with version and capability
details)
- [x] I have checked ROADMAP.md and confirmed this PR does not duplicate
planned core work
- [x] I have run tests locally and they pass
- [x] I have added or updated tests where applicable
- [x] If this change affects the UI, I have included before/after
screenshots
- [x] I have updated relevant documentation to reflect my changes
- [x] I have considered and documented any risks above
- [x] I will address all Greptile and reviewer comments before
requesting merge
- [x] I searched the GitHub PR list for similar PRs and confirmed this
is not a duplicate
Co-authored-by: Devin Foley <devin@paperclip.ing>
Co-authored-by: Paperclip <noreply@paperclip.ing>
## Thinking Path
> - Paperclip orchestrates AI agents for zero-human companies
> - codex_local runs Codex CLI under a per-company "managed home" so
multiple companies don't trample on each other's session state
> - For `auth.json` specifically, the managed home keeps a SYMLINK to
the user's real `~/.codex/auth.json` rather than a copy — Codex refresh
tokens rotate and are single-use, so any copy goes stale the moment the
source rotates and every subsequent run dies with `401
refresh_token_reused`
> - Older Paperclip versions copied `auth.json` instead. After
upgrading, `ensureSymlink()` saw a regular file at the target, hit `if
(!existing.isSymbolicLink()) return;`, and silently kept the stale copy
> - This pull request makes the upgrade path self-healing inside
`ensureSymlink()` itself: when the target is a regular file, unlink it
and create the symlink, since the target lives under the
Paperclip-managed home and is safe to delete. Directories are skipped to
avoid `EISDIR` on Unix (and inconsistent behavior on Windows)
> - The benefit is operators who upgraded from a copy-based version stop
getting refresh-token-reused failures without having to manually purge
`companies/<id>/codex-home/auth.json`, and the healing is
defense-in-depth even outside the `prepareManagedCodexHome` cleanup path
## What Changed
- `packages/adapters/codex-local/src/server/codex-home.ts` —
`ensureSymlink()` previously bailed out of the
`!existing.isSymbolicLink()` branch, leaving any pre-existing regular
file untouched. Now unlinks and recreates the symlink in that branch via
the existing `createExpectedSymlink()` helper (preserves the EEXIST
race-tolerance behavior added in #5119). A guard skips directories so
the call never throws `EISDIR` and aborts `prepareManagedCodexHome`.
Inline comment explains the safety: target is always under the
company-scoped managed home
(`<paperclipHome>/instances/<id>/companies/<companyId>/codex-home/`),
never the user's real `~/.codex`.
- `packages/adapters/codex-local/src/server/codex-home.test.ts` — adds a
regression test for #5028: pre-seed a stale copy at the target, run
`prepareManagedCodexHome`, assert the target is now a symlink and reads
through to the fresh source. The existing concurrent-symlink test is
preserved.
## Verification
```
pnpm --filter @paperclipai/adapter-codex-local exec vitest run
# Test Files 8 passed (8)
# Tests 26 passed (26)
pnpm --filter @paperclipai/adapter-codex-local exec tsc --noEmit
# clean
```
Manual repro flow that the regression test mirrors:
1. Create a stale copy: `echo '{"token":"old"}' >
<managedHome>/auth.json`.
2. Rotate source: `echo '{"token":"new"}' > ~/.codex/auth.json`.
3. Trigger any codex_local run — `prepareManagedCodexHome` is called
from the execute path, the managed file is now a symlink to the source,
and the CLI sees the fresh token.
## Risks
- **Low risk.** The new branch only fires when the target file is a
regular file (the upgrade path) — a pure copy that Codex couldn't have
written, since Codex never writes into the managed home. Operators in
steady-state on the symlink-based version are unaffected.
- The `fs.unlink` only runs against the per-company managed-home path,
never the user's real `~/.codex`. Inline comment makes this guarantee
explicit.
- A directory at the auth.json path is left in place (no silent `EISDIR`
crash) — this requires operator inspection rather than autonomous
deletion.
- The healing uses `createExpectedSymlink()` so it remains tolerant of
EEXIST races with concurrent prepare calls (the concurrent-symlink test
still passes).
- No DB / migration / schema impact.
## Model Used
- Anthropic Claude Opus 4.7 (claude-opus-4-7), via Claude Code CLI with
extended tool use (Read / Edit / Bash / Grep). No extended-thinking
budget consumed beyond default.
## Checklist
- [x] I have included a thinking path that traces from project context
to this change
- [x] I have specified the model used (with version and capability
details)
- [x] I have checked ROADMAP.md and confirmed this PR does not duplicate
planned core work
- [x] I have run tests locally and they pass
- [x] I have added or updated tests where applicable
- [ ] If this change affects the UI, I have included before/after
screenshots — N/A, adapter-only
- [x] I have updated relevant documentation to reflect my changes —
inline comment explains the why and the safety of the unlink
- [x] I have considered and documented any risks above
- [x] I will address all Greptile and reviewer comments before
requesting merge
- [x] I searched the GitHub PR list for similar PRs and confirmed this
is not a duplicate
Fixes#5028.
---------
Co-authored-by: Devin Foley <devin@paperclip.ing>
Co-authored-by: Paperclip <noreply@paperclip.ing>
## Thinking Path
> - Paperclip orchestrates AI agents for zero-human companies.
> - The Claude-local adapter resumes prior sessions via `claude --resume
<session-id>` so work continues across heartbeats.
> - When a resumed session contains an image whose content is no longer
accessible, Claude returns a 400 "Could not process image" — but the
session itself is poisoned and will keep returning the same error on
every resume.
> - The existing retry path only catches the "unknown session" 400 case;
image-processing 400s on resume fall through and the run fails for the
user.
> - This PR adds an `isClaudeImageProcessingError` detector mirroring
`isClaudeUnknownSessionError` and wires it into the same fresh-session
retry branch in `execute.ts`.
> - The benefit is that a poisoned-image resume self-recovers by
retrying once with a fresh session, exactly like the existing
unknown-session path.
## Linked Issues or Issue Description
Fixes#3275
Refs #3123
## What Changed
- Added `isClaudeImageProcessingError()` in
`packages/adapters/claude-local/src/server/parse.ts` that matches `Could
not process image` in 400 error messages.
- Wired the new detector into the existing session-resume retry branch
in `packages/adapters/claude-local/src/server/execute.ts` alongside
`isClaudeUnknownSessionError`.
- Retry only fires when `sessionId` is present (i.e. we were resuming),
so fresh-session runs that hit the same error are not retried (no
infinite loop).
## Verification
- `pnpm --filter @paperclipai/adapter-claude-local test` covers
`parse.ts` patterns and the resume-retry decision branch.
- `pnpm --filter @paperclipai/adapter-claude-local typecheck`
## Risks
Low. Behavior change is narrowly additive: a previously-fatal 400 on
resume now triggers a single fresh-session retry. No effect on
fresh-session runs, unknown-session retries, or non-image 400s.
## Model Used
Claude (Opus 4.6) — used to mirror the existing unknown-session pattern
and verify the guard against infinite loops.
## Checklist
- [x] I have included a thinking path that traces from project context
to this change
- [x] I have specified the model used (with version and capability
details)
- [x] I have checked ROADMAP.md and confirmed this PR does not duplicate
planned core work
- [x] I have searched GitHub for duplicate or related PRs and linked
them above
- [x] I have either (a) linked existing issues with `Fixes: #` / `Closes
#` / `Refs #` OR (b) described the issue in-PR following the relevant
issue template
- [x] I have run tests locally and they pass
- [x] I have added or updated tests where applicable
- [x] If this change affects the UI, I have included before/after
screenshots (N/A — no UI changes)
- [x] I have updated relevant documentation to reflect my changes
- [x] I have considered and documented any risks above
- [ ] All Paperclip CI gates are green (in progress)
- [ ] Greptile is 5/5 with no open P2s, recommendations, or follow-ups
- [x] I will address all Greptile and reviewer comments before
requesting merge
Co-authored-by: Devin Foley <devin@paperclip.ing>
Co-authored-by: Paperclip <noreply@paperclip.ing>
## Thinking Path
> - Paperclip orchestrates AI agents on pluggable adapters
(`claude_local`, `opencode_local`, `codex_local`, …); each adapter wraps
an external CLI.
> - The heartbeat service stores a session ID per agent and replays it
back to the adapter via `--resume` so within-task continuity is
preserved.
> - Session IDs are adapter-specific in format: claude expects a UUID,
opencode emits `ses_…`, etc. They cannot be cross-replayed.
> - When the cross-adapter session ID does slip through (operator
changes `adapterType`, edge cases in the resume path, foreign-format ID
in stored task sessions), the claude CLI hard-fails with a validation
error and every subsequent heartbeat loops on the same error until the
stored ID is manually cleared.
> - Master now ships a canonical-session-ID guard at `heartbeat.ts:8450`
(via #5972) that prevents most of this at the source, and
`isClaudePoisonedPreviousMessageIdError` recovers from the 400-class API
error.
> - This PR adds defense-in-depth at the adapter layer: the `--resume
requires a valid session ID … not a UUID …` validation error from the
claude CLI is now classified as an unknown-session signal, so the
existing fresh-session retry recovers instead of hard-failing.
## Linked Issues or Issue Description
Refs #5972 — sibling fix on the same cluster (recovers from poisoned
`previous_message_id` 400). This PR complements it by handling the
CLI-layer `--resume` validation error class.
## What Changed
- `packages/adapters/claude-local/src/server/parse.ts` — broaden
`isClaudeUnknownSessionError` regex to also match `--resume requires a
valid session`, `is not a UUID`, and `does not match any session title`.
The existing fresh-session retry at `execute.ts:612-625` now fires for
this error class.
- `packages/adapters/claude-local/src/server/parse.test.ts` — adds 4 new
test cases for `isClaudeUnknownSessionError` covering the legacy and new
patterns plus a negative case.
**Dropped from the original PR on rebase** (already on master, would
conflict):
- `server/src/services/heartbeat.ts` runtimeSessionFallback gate —
superseded by the stricter `isCanonicalSessionIdForAdapter` check on
master (#5972 lineage).
- `packages/adapters/claude-local/vitest.config.ts` and
`vitest.config.ts` projects entry — both already in master.
## Verification
```sh
pnpm --filter @paperclipai/adapter-claude-local vitest run
# 19/19 passed (3 files, includes 4 new isClaudeUnknownSessionError cases)
```
Pre-existing failure on
`server/src/__tests__/heartbeat-process-recovery.test.ts > queues
exactly one retry when the recorded local pid is dead` reproduces on
`origin/master` — unrelated to this PR.
## Risks
- **Low-to-medium.** The added regex fragments are narrow. `--resume
requires a valid session` and `does not match any session title` are
unambiguously session-related. `is not a UUID` is more generic; worst
case is one extra retry on an unrelated CLI validation error that would
also fail on the same root issue. Happy to drop `is not a UUID` if
reviewers prefer.
- **No DB migration; no schema change; no behavior change when adapter
types match (the common path).**
## Model Used
- Provider: Anthropic (Claude)
- Model: `claude-opus-4-7` (Opus 4.7), 1M context window
- Tool: Claude Code CLI with extended thinking + tool use; human review
on the rebase and the regex narrowing tradeoffs
## Checklist
- [x] I searched the GitHub PR list for similar PRs and confirmed this
is not a duplicate (related: #5972 already merged, complementary scope)
- [x] I have included a thinking path that traces from project context
to this change
- [x] I have specified the model used (with version and capability
details)
- [x] I have checked ROADMAP.md and confirmed this PR does not duplicate
planned core work
- [x] I have run tests locally and they pass (19/19 claude-local)
- [x] I have added or updated tests where applicable
- [x] I have considered and documented any risks above
- [x] I will address all Greptile and reviewer comments before
requesting merge
Co-authored-by: Devin Foley <devin@paperclip.ing>
Co-authored-by: Paperclip <noreply@paperclip.ing>
## Thinking Path
> - Paperclip orchestrates AI agents for zero-human companies.
> - The Claude-local adapter uses `claude --resume <session-id>` to
continue prior sessions; the `--resume` value MUST be a UUID per
Claude's CLI contract.
> - Paperclip internally uses session IDs prefixed with `ses_` (not
UUIDs); these get passed straight through to `--resume` and crash the
run.
> - On top of the crash, when the underlying error path triggers a
secret-decryption failure or heartbeat setup failure, the diagnostics
are too thin to tell key-mismatch from other failures, and the heartbeat
error code is mis-classified as `adapter_failed` instead of
`setup_failed`.
> - This PR validates `runtimeSessionId` against a UUID regex before
letting `canResumeSession` become true, adds `not a valid UUID` to
Claude's own retry-error regex, improves AES-256-GCM decryption
diagnostics in the local encrypted provider, and re-classifies
pre-adapter setup failures.
> - The benefit is that Paperclip session IDs are detected and skipped
gracefully (logged, no crash), legitimate Claude UUID-rejection errors
are treated as retriable, and operators can diagnose decryption/setup
failures from the run log.
## Linked Issues or Issue Description
**What happened?**
The `claude-local` adapter passes Paperclip's internal session
identifiers (e.g. `ses_…`) straight to `claude --resume <session-id>`.
Because Claude's CLI requires the `--resume` argument to be a UUID, the
run crashes with a `not a valid UUID` error. When the surrounding code
path also hits a secret-decryption failure, the heartbeat reports it as
`adapter_failed`, hiding the real `setup_failed` cause and making
diagnosis hard.
**Expected behavior**
Non-UUID session IDs should be detected before `--resume` is called, the
run should fall back to a fresh session with a clear log line, and any
decryption / setup failure should be reported with enough detail (and
the correct error code) for an operator to tell what failed.
**Steps to reproduce**
1. Have a persisted task session whose ID is not a UUID
(Paperclip-issued `ses_…` form).
2. Trigger a heartbeat that resumes that session via the `claude-local`
adapter.
3. Observe: the adapter crashes with a UUID-validation error; if the
path also involves a decryption failure, the heartbeat surfaces
`adapter_failed` instead of `setup_failed`.
## What Changed
- `packages/adapters/claude-local/src/server/execute.ts`: Validates
`runtimeSessionId` against a UUID regex before setting
`canResumeSession`; non-UUID IDs are logged and skipped gracefully.
Guards the cwd-mismatch log block on `isValidUuid` so it does not fire
for non-UUID session IDs.
- `packages/adapters/claude-local/src/server/parse.ts`: Adds `not a
valid UUID` to the session-error retry regex so Claude's own UUID
rejection is treated as a retriable error.
- `server/src/services/secrets/local-encrypted-provider.ts`: Wraps
AES-256-GCM decryption in try/catch and re-throws with a key fingerprint
hint to aid key-mismatch diagnosis.
- `server/src/services/heartbeat.ts`: Corrects the outer-catch
`errorCode` from `adapter_failed` to `setup_failed` for pre-adapter
setup failures.
- `AGENTS.md`: Adds task/PR/CI governance sections (10–13) and expands
the Definition of Done.
## Verification
- `pnpm --filter @paperclipai/adapter-claude-local test` covers UUID
validation and the parse retry regex.
- `pnpm --filter @paperclipai/server test src/services/secrets` covers
decryption diagnostics.
- `pnpm --filter @paperclipai/server typecheck`
## Risks
Low. UUID validation is strictly additive (non-UUIDs that previously
crashed now log and skip). Decryption diagnostics only fire on failure
paths. The `setup_failed` error code change is a clearer classification,
not a behavior change.
## Model Used
Claude (Opus 4.6) — used to identify the UUID-validation root cause,
mirror existing parse patterns, and re-classify the heartbeat setup
error code.
## Checklist
- [x] I have included a thinking path that traces from project context
to this change
- [x] I have specified the model used (with version and capability
details)
- [x] I have checked ROADMAP.md and confirmed this PR does not duplicate
planned core work
- [x] I have searched GitHub for duplicate or related PRs and linked
them above
- [x] I have either (a) linked existing issues with `Fixes: #` / `Closes
#` / `Refs #` OR (b) described the issue in-PR following the relevant
issue template
- [x] I have run tests locally and they pass
- [x] I have added or updated tests where applicable
- [x] If this change affects the UI, I have included before/after
screenshots (N/A — no UI changes)
- [x] I have updated relevant documentation to reflect my changes
- [x] I have considered and documented any risks above
- [ ] All Paperclip CI gates are green (in progress)
- [ ] Greptile is 5/5 with no open P2s, recommendations, or follow-ups
- [x] I will address all Greptile and reviewer comments before
requesting merge
---------
Co-authored-by: CTO Agent <cto@paperclip.ing>
Co-authored-by: Paperclip <noreply@paperclip.ing>
Co-authored-by: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
Co-authored-by: Devin Foley <devin@paperclip.ing>
## Thinking Path
> - Paperclip's `claude_local` adapter persists Claude Code session
jsonls under `~/.claude/projects/…/{sessionId}.jsonl` and resumes them
on the next heartbeat
> - When Claude Code injects `<synthetic>` placeholder assistant
messages (after rate-limit, max-turn exhaustion, or transient-upstream
failures) those placeholders get UUID-format `message.id`s rather than
`msg_…`-format ids
> - On the next `--resume`, Claude Code passes that UUID as
`previous_message_id` and Anthropic's API rejects it with a 400:
``diagnostics.previous_message_id: must be the `id` from a prior
/v1/messages response (starts with `msg_`)``
> - The adapter had a session-rotation fallback only for "unknown
session" errors, so the poisoned session was `--resume`-d indefinitely
and the agent flipped between `idle` and `error` every heartbeat
> - Even worse, the *result* event of the failing run still carried a
`session_id`, and the adapter was persisting that id into the
issue-scoped session store (`agentTaskSessions`). So even after we
detected the 400, every subsequent continuation re-loaded the same
poisoned id and hit the same 400 again — the issue was permanently
stranded
> - We observed this on multiple agents in our deployment; the only
manual fix was to rename the `.jsonl`, which is not a viable long-term
workaround
> - This PR detects the 400, runs the same session-rotation fallback the
unknown-session path uses **and** stops persisting the poisoned id, so
the next attempt starts genuinely fresh
## Linked Issues or Issue Description
No external GitHub issue is linked. Describing the problem inline
following the bug-report template:
**What happened:** `claude_local` agents flipped between `idle` and
`error` on every heartbeat because the persisted session jsonl carried a
synthetic UUID `previous_message_id` (from `<synthetic>` assistant
placeholders injected after rate-limit/max-turn/upstream errors).
Anthropic's API rejected every `--resume` with a 400:
``diagnostics.previous_message_id: must be the `id` from a prior
/v1/messages response (starts with `msg_`)``.
**Expected behavior:** When the persisted session is poisoned and
unrecoverable, the adapter should rotate to a fresh session — the same
fallback path already used for unknown-session errors — and stop
re-persisting the poisoned `session_id`.
**Actual behavior:** The session-rotation fallback only matched the
"unknown session" pattern, so the poisoned session was `--resume`-d
forever. The result event of the failing run still carried `session_id`,
which was being persisted into `agentTaskSessions`, so every subsequent
continuation reloaded the same poisoned id and hit the same 400.
**Reproduction:** Inject any flow that causes Claude Code to emit a
`<synthetic>` placeholder (rate-limit, max-turn exhaustion, transient
upstream failure). The next `--resume` will fail with the 400 and the
agent will not self-recover.
**Scope of fix:** Add a `previous_message_id` 400 detector; route it
through the existing unknown-session fallback; drop the poisoned
`sessionId` and emit `clearSession: true` so the heartbeat service wipes
the persisted row; best-effort delete the local poisoned `.jsonl`.
## What Changed
Two commits:
1. **`adapter-claude-local: auto-rotate session on previous_message_id
400 (synthetic-msg poisoning)`** — detector + execute-time rotation
2. **`adapter-claude-local: guard against persisting poisoned
sessionId`** — validate-before-persist + `clearSession`
Combined diff:
- `parse.ts`: new `isClaudePoisonedPreviousMessageIdError(parsed)`
matching ``/diagnostics\.previous_message_id.*starts with `msg_`/i``
against `parsed.result` and `extractClaudeErrorMessages(parsed)`
- `parse.ts`: `isClaudeTransientUpstreamError()` excludes the new error
from transient classification so it isn't masked as retryable upstream
noise
- `execute.ts`: expand the resume-fallback branch so it triggers on both
`isClaudeUnknownSessionError` and the new
`isClaudePoisonedPreviousMessageIdError`, with a distinct log line
(`"returned a poisoned message-id"` vs `"is unavailable"`)
- `execute.ts`: for local (non-remote) execution targets, best-effort
delete the poisoned `~/.claude/projects/.../{sessionId}.jsonl` before
retrying so the file can't be accidentally resumed by an out-of-band
caller. The `fs.unlink` and follow-up log call are in separate try/catch
blocks so a closed log stream cannot mask a successful unlink (and vice
versa)
- `execute.ts` / `toAdapterResult`: when a result carries the poisoned
400, **drop** `sessionId`/`sessionParams`/`sessionDisplayId` (return
`null`) and emit `clearSession: true` so the heartbeat service's
`resolveNextSessionState` wipes the persisted row. The result also
surfaces `errorCode: "claude_poisoned_previous_message_id"` for
observability
- `docs/adapters/claude-local.md`: runbook entry — symptom,
auto-recovery flow, on-call checklist
- Tests:
- 4 new `parse.test.ts` cases covering positive detection in `result`
and `errors[]`, negative cases, and non-transient classification
- 3 new `claude-local-execute.test.ts` cases: (a) fresh run reports the
poisoned error → sessionId dropped + `clearSession: true`; (b) recovery
retry also reports the poisoned error → same guards apply; (c)
session-rotation success on retry
## Verification
```bash
pnpm --filter @paperclipai/adapter-claude-local exec vitest run src/server/parse.test.ts
pnpm --filter @paperclipai/server exec vitest run src/__tests__/claude-local-execute.test.ts
```
Both suites green locally. This patch is also currently running as a
hot-patch over the published `2026.513.0` adapter on the reporting
deployment — sessions that previously looped indefinitely now
self-recover on the first heartbeat after the 400 surfaces.
## Risks
- Low risk. The detector is conservative (regex over `result` +
`errors[]` only) and the rotation reuses the existing unknown-session
fallback path
- The local-only `fs.unlink` of the poisoned `.jsonl` is wrapped in
`try/catch` and ignored on failure — strictly an optimization; the
server-side session clear is the authoritative reset
- Remote execution targets (`executionTargetIsRemote`) skip the disk
cleanup because the file lives on a remote host that we can't safely
reach from the adapter
- The `clearSession: true` + nulled session fields path is a no-op on
healthy runs; it only fires when the new detector matches, so existing
successful continuations are unaffected
- No DB schema changes, no public API changes, no new dependencies
## Model Used
- Provider: Anthropic Claude
- Model: `claude-opus-4-7` (Opus 4.7)
- Context window: 1M
- Capabilities: extended reasoning, tool use, code execution
- Role: implemented the detector, expanded the fallback branch, added
the persist-guard + `clearSession`, wrote the unit + integration tests,
validated locally, and applied the equivalent hot-patch to the deployed
`2026.513.0` install while this PR is in review
## Checklist
- [x] I have included a thinking path that traces from project context
to this change
- [x] I have specified the model used (with version and capability
details)
- [x] I have checked ROADMAP.md and confirmed this PR does not duplicate
planned core work
- [x] I have searched GitHub for similar or duplicate PRs and linked
them — closed#2295, #2361, #3572, #5438 as duplicates of this canonical
fix; complementary fixes#4838 (heartbeat_timer reset) and #4932 (gemini
context-overflow rotation) target different code paths
- [x] I have run tests locally and they pass
- [x] I have added or updated tests where applicable
- [x] If this change affects the UI, I have included before/after
screenshots — N/A, adapter-only change
- [x] I have updated relevant documentation
(`docs/adapters/claude-local.md` runbook entry)
- [x] I have considered and documented any risks above
- [x] I will address all Greptile and reviewer comments before
requesting merge
---------
Co-authored-by: Danial Jawaid <danial.jawaid@gmail.com>
Co-authored-by: Paperclip <noreply@paperclip.ing>
Co-authored-by: Devin Foley <devin@paperclip.ing>
## Summary
Adds the newly released Claude models from the [models
overview](https://platform.claude.com/docs/en/about-claude/models/overview)
to the `claude_local` adapter's model selector:
- **Claude Fable 5** (`claude-fable-5`) — generally available as of
2026-06-09, Anthropic's most capable widely-released model.
- **Claude Mythos 5** (`claude-mythos-5`) — limited availability
(Project Glasswing).
**Opus 4.8 stays first in the list so it remains the default selection**
— per the request, the new flagship models are *offered* but not
defaulted (not Fable, not Mythos).
## Changes
- `packages/adapters/claude-local/src/index.ts` — add `claude-fable-5`
and `claude-mythos-5` to the adapter model list, right after
`claude-opus-4-8`.
- `packages/adapters/claude-local/src/server/models.ts` — add the Fable
5 Bedrock identifier (`us.anthropic.claude-fable-5-v1`) to the Bedrock
fallback list. Mythos 5 is limited-availability on Bedrock, so it's
intentionally left out of that fallback.
- `server/src/__tests__/adapter-models.test.ts` — assert the new models
are present and that `claude-opus-4-8` remains first (the default).
These flow through the single `claudeModels` source, so they also appear
in the ACPX combined list (`registry.ts` prefixes them with `Claude:`)
and are recognized by the ACPX Claude model filter. The UI selector
reads models dynamically from the adapter, so no UI changes are needed.
## Testing
- `npx vitest run src/__tests__/adapter-models.test.ts` — 13 passed.
🤖 Generated with [Claude Code](https://claude.com/claude-code)
Co-authored-by: Claude Opus 4.8 <noreply@anthropic.com>
Bumps [@cursor/sdk](https://github.com/cursor/cursor) from 1.0.12 to
1.0.18.
<details>
<summary>Commits</summary>
<ul>
<li>See full diff in <a
href="https://github.com/cursor/cursor/commits">compare view</a></li>
</ul>
</details>
<br />
[](https://docs.github.com/en/github/managing-security-vulnerabilities/about-dependabot-security-updates#about-compatibility-scores)
Dependabot will resolve any conflicts with this PR as long as you don't
alter it yourself. You can also trigger a rebase manually by commenting
`@dependabot rebase`.
[//]: # (dependabot-automerge-start)
[//]: # (dependabot-automerge-end)
---
<details>
<summary>Dependabot commands and options</summary>
<br />
You can trigger Dependabot actions by commenting on this PR:
- `@dependabot rebase` will rebase this PR
- `@dependabot recreate` will recreate this PR, overwriting any edits
that have been made to it
- `@dependabot show <dependency name> ignore conditions` will show all
of the ignore conditions of the specified dependency
- `@dependabot ignore this major version` will close this PR and stop
Dependabot creating any more for this major version (unless you reopen
the PR or upgrade to it yourself)
- `@dependabot ignore this minor version` will close this PR and stop
Dependabot creating any more for this minor version (unless you reopen
the PR or upgrade to it yourself)
- `@dependabot ignore this dependency` will close this PR and stop
Dependabot creating any more for this dependency (unless you reopen the
PR or upgrade to it yourself)
</details>
Signed-off-by: dependabot[bot] <support@github.com>
Co-authored-by: dependabot[bot] <49699333+dependabot[bot]@users.noreply.github.com>
## Thinking Path
> - Paperclip orchestrates AI-agent companies through adapter-backed
local and external runtimes.
> - The agent configuration UI lets operators choose adapter models and
refresh model lists when adapters support live discovery.
> - Codex already had a live refresh path, but Claude Local only exposed
static fallback models and the UI hid the refresh action for Claude.
> - A newly available Claude Opus model should not require a code
release every time the model catalog changes.
> - This pull request adds Anthropic model discovery for Claude Local,
keeps the static fallback current with Claude Opus 4.8, and exposes the
existing refresh button in the Claude Local dropdown.
> - The benefit is that operators can refresh Claude models from the
same model selector flow they already use for Codex.
## What Changed
- Added `claude-opus-4-8` to the Claude Local fallback model list.
- Added Claude model discovery through Anthropic-compatible `GET
/v1/models` when `ANTHROPIC_API_KEY` is available.
- Added normal cache reuse, forced refresh support, a SHA-256-based
API-key fingerprint for cache keys, and warning logging for discovery
errors before fallback.
- Wired `claude_local.refreshModels` into the server adapter registry.
- Enabled the existing `Refresh models` dropdown action for
`claude_local` in `AgentConfigForm`.
- Added tests for Claude fallback, live discovery, API-failure fallback,
forced refresh, and the UI refresh-button gate.
## Verification
- `pnpm exec vitest run server/src/__tests__/adapter-models.test.ts`
- `pnpm exec vitest run ui/src/components/AgentConfigForm.test.ts`
- `pnpm --filter @paperclipai/adapter-claude-local typecheck`
- `pnpm --filter @paperclipai/server typecheck`
- `pnpm --filter @paperclipai/ui typecheck`
- Greptile review reached Confidence Score: 5/5 on commit `b796cf4f1`
with addressed threads resolved.
UI note: the visible change is a conditional action row inside the
existing model dropdown; the regression test covers that `claude_local`
now receives the refresh action.
## Risks
- Low risk. Without `ANTHROPIC_API_KEY`, Claude Local still uses the
static fallback list.
- If Anthropic model discovery fails or times out, Paperclip falls back
to the existing cached or static list.
- Bedrock environments remain on Bedrock-native model IDs.
## Model Used
OpenAI GPT-5 via Codex local coding agent, with repository file access,
shell command execution, git operations, and targeted test/typecheck
verification. Exact context window is not exposed by the runtime.
## Checklist
- [x] I have included a thinking path that traces from project context
to this change
- [x] I have specified the model used (with version and capability
details)
- [x] I have checked ROADMAP.md and confirmed this PR does not duplicate
planned core work
- [x] I have run tests locally and they pass
- [x] I have added or updated tests where applicable
- [x] If this change affects the UI, I have included before/after
screenshots
- [x] I have updated relevant documentation to reflect my changes
- [x] I have considered and documented any risks above
- [x] I will address all Greptile and reviewer comments before
requesting merge
## Thinking Path
> - Paperclip orchestrates AI agents across isolated execution
workspaces; the local cwd is the only persistence boundary between runs.
> - Workspace lifecycle (worktree_prepare → execute →
workspace_finalize) and the wake/accept flow are what guarantee that
dependent issues see a consistent worktree.
> - PAPA-380 / PAPA-431 / PAPA-432 / PAPA-440 surfaced three holes in
that contract: silent env reuse across assignees, dependent wakes firing
before finalize, and `issue.interaction.accept` advancing before
finalize landed.
> - PAPA-441 / PAPA-442 then needed to document the "no remote git"
contract and prevent future adapter/runtime code from quietly
reintroducing `git push` as a backdoor sync.
> - This pull request lands those server fixes, the static
`check-no-git-push` enforcement, the AUTHORING.md cross-link, and the
Cody-review follow-ups on the PAPA-430 thread.
> - The benefit is that finalize is a real barrier — board accepts,
dependent wakes, and operator-set env all respect it — and adapter code
can't bypass it via raw `git push`.
## What Changed
- **server (PAPA-380, PAPA-431):** `execution-workspace-policy` refuses
silent env reuse when the assignee's resolved env disagrees with the
workspace it would inherit. The inheritance protection is now scoped to
the actual inheritance signal — explicit issue-level `environmentId` is
honored even when the agent's default env is `null`.
- **server (PAPA-432):** `heartbeat.ts` gates dependent wakes on
`listUnfinalizedExecutionWorkspaceIds`, and writes a
`workspace_finalize` row on the succeeded path. Write failures now
surface instead of being swallowed so dependents aren't silently
stranded behind a missing row.
- **server (PAPA-440):** `issue-thread-interactions.acceptInteraction`
adds a workspace_finalize precondition for `request_confirmation` (not
`suggest_tasks`). Accept returns 409 if finalize hasn't succeeded for
the latest workspace operation.
- **ci (PAPA-442):** new `scripts/check-no-git-push.mjs` static check
scans `packages/adapters/`, `packages/adapter-utils/`, `server/src/`,
and `cli/src/` for any `git push` invocation (string or args-array).
Wired into the `policy` PR job and `test:release-registry`. Operators
can opt in per-call with `// paperclip:allow-git-push: <reason>`.
Release scripts are out of scope by design.
- **docs (PAPA-441):** `AUTHORING.md` documents the no-remote-git
contract and cross-links the static check so adapter authors learn the
rule and the enforcement together.
- **review follow-up (PAPA-430, Cody):** three fixes — env resolver bug,
accept-gate scope (request_confirmation only), and finalize record write
on the succeeded path.
## Verification
- `pnpm exec vitest run
server/src/__tests__/execution-workspace-policy.test.ts
server/src/__tests__/issue-thread-interactions-service.test.ts` → 33/33
pass
- `node scripts/check-no-git-push.test.mjs` → check covers string form,
args-array form, comment exclusions, and per-line allow-comment.
- Manual: server compiles; the policy job runs the check in <1s before
heavier jobs.
## Risks
- **Behavioral shift in accept:** boards accepting
`request_confirmation` while finalize is in-flight now get 409s. This is
intentional — they can retry — but it changes timing on a hot path.
`suggest_tasks` is unaffected.
- **Workspace policy:** the env-reuse refusal is a new error path.
Issues that previously silently reused an env from a different-assignee
workspace will now fail-loud; the resolver still honors explicit
issue-level `executionWorkspaceSettings.environmentId`.
- **CI rule:** any future legitimate `git push` in scoped dirs must be
marked with the allow-comment, which is the intended ergonomic.
## Model Used
- Claude Opus 4.7 (`claude-opus-4-7`, extended thinking), via Claude
Code in the Paperclip executor adapter.
## Checklist
- [x] I have included a thinking path that traces from project context
to this change
- [x] I have specified the model used (with version and capability
details)
- [x] I have checked ROADMAP.md and confirmed this PR does not duplicate
planned core work
- [x] I have run tests locally and they pass
- [x] I have added or updated tests where applicable
- [ ] If this change affects the UI, I have included before/after
screenshots (N/A — server/CI/docs only)
- [x] I have updated relevant documentation to reflect my changes
- [x] I have considered and documented any risks above
- [x] I will address all Greptile and reviewer comments before
requesting merge
Closes related issues: PAPA-430, PAPA-380, PAPA-431, PAPA-432, PAPA-440,
PAPA-441, PAPA-442
---------
Co-authored-by: Paperclip <noreply@paperclip.ing>
## Thinking Path
> - Paperclip orchestrates AI agents for zero-human companies through
company-scoped control-plane workflows.
> - Agents need reusable, inspectable skills that can be installed,
reset, audited, exported, and assigned without bespoke local setup.
> - The existing skill truth model needed cleanup so bundled skills,
optional catalog skills, runtime skills, and adapter-provided skills
have clear provenance.
> - Operators also need a practical CLI and board UI for discovering and
managing company skills.
> - This pull request adds the skills CLI, packaged skills catalog,
company skills APIs, and catalog-aware board UI.
> - The benefit is a more reusable Paperclip company setup where skills
are portable, auditable, and easier for operators and agents to manage.
## What Changed
- Added `paperclipai skills` CLI commands and coverage for catalog
listing, installing, resetting, and inspecting company skills.
- Added a packaged `@paperclipai/skills-catalog` workspace with bundled
and optional skill content plus validation/build tests.
- Added shared company-skill types and validators used across CLI,
server, and UI contracts.
- Added server catalog APIs/services for company skill catalog
operations, reset semantics, audit behavior, and portability provenance.
- Updated adapter skill handling so runtime/catalog provenance remains
explicit across local adapters.
- Added board UI support for browsing and managing catalog-backed
company skills.
- Updated docs for the skills CLI/catalog flow and the company skills
Paperclip skill reference.
- Rebased the branch onto current `paperclipai/paperclip:master`; no
`pnpm-lock.yaml`, `.github/workflows`, or migration files are included
in the final PR diff.
## Verification
- Passed: `pnpm run preflight:workspace-links && pnpm exec vitest run
cli/src/__tests__/skills.test.ts
packages/skills-catalog/src/catalog-builder.test.ts
packages/skills-catalog/src/shipped-catalog.test.ts
packages/shared/src/validators/company-skill.test.ts
packages/adapter-utils/src/server-utils.test.ts
packages/plugins/create-paperclip-plugin/src/entrypoints.test.ts
server/src/__tests__/company-skills-catalog-service.test.ts
server/src/__tests__/company-skills-routes.test.ts
server/src/__tests__/company-portability.test.ts`.
- Passed: `pnpm exec vitest run
server/src/__tests__/workspace-runtime.test.ts -t "default
branch|origin/master|symbolic-ref"`.
- Attempted: full `server/src/__tests__/workspace-runtime.test.ts`. Four
provisioning tests failed while seeding an isolated worktree database
from the local Paperclip instance because the local plugin schema dump
contains a duplicate-column foreign key
(`plugin_content_machine_18a7bc327b.content_case_signals`). The
default-branch tests touched by the rebase conflict passed in the
focused run above.
- Checked final diff: no `pnpm-lock.yaml`, no `.github/workflows`, and
no migration-file changes relative to `master`.
## Risks
- Medium: this is a broad skills/catalog change touching CLI, server
APIs, shared contracts, adapter skill sync, and UI.
- Catalog validation and reset semantics need careful reviewer attention
because they affect reusable company setup and portability.
- No database migrations are included in this PR, so there is no
migration ordering/idempotency risk in the final diff.
- No lockfile is included by design; dependency resolution will be
handled by the repository lockfile workflow.
## Model Used
- OpenAI Codex coding agent based on GPT-5, running in Paperclip via the
`codex_local` adapter with shell, git, GitHub CLI, and code-editing tool
access. Exact hosted model build/context-window metadata is not exposed
in this runtime.
## Checklist
- [x] I have included a thinking path that traces from project context
to this change
- [x] I have specified the model used (with version and capability
details)
- [x] I have checked ROADMAP.md and confirmed this PR does not duplicate
planned core work
- [x] I have run targeted tests locally and documented the local
workspace-runtime seed failure above
- [x] I have added or updated tests where applicable
- [x] If this change affects the UI, screenshots were intentionally
omitted per PAP-10124 instructions; UI behavior is covered by tests and
reviewer inspection
- [x] I have updated relevant documentation to reflect my changes
- [x] I have considered and documented any risks above
- [x] I will address all Greptile and reviewer comments before
requesting merge
---------
Co-authored-by: Paperclip <noreply@paperclip.ing>
## Thinking Path
> - Paperclip orchestrates AI agents for zero-human companies, so when
an adapter fails, the platform must surface enough detail for the next
agent (or human reviewer) to act
> - The `acpx_local` adapter wraps `claude-agent-acp`, which in turn
drives the Claude Code SDK — three layers, three different permission
and error-handling models
> - A user created a `Claude Local ACPX` agent in PAPA-387 and it failed
instantly with the generic `acpx.error / "Internal error"` log,
stranding the work and triggering an opaque `stranded_assigned_issue`
recovery to the CTO
> - Once the diagnostic blackbox was opened, the underlying cause turned
out to be two SDK-level mismatches: a model-name allowlist that rejects
bare IDs like `claude-opus-4-7`, and a Claude Code
permission/Read-sandbox configuration that silently denies every
non-allowlisted tool when the user's `~/.claude/settings.json` has
`defaultMode: "dontAsk"`
> - This pull request fixes both classes of failure in the adapter
itself so new ACPX agents work seamlessly without per-host
configuration, and widens the diagnostic surface so the *next* failure
of any kind is actionable
> - The benefit is that ACPX-Claude can join the regular agent roster —
verified end to end on PAPA-401, where the agent successfully reached
the Paperclip API, opened a worktree, surveyed existing notification
PRs, and posted a structured plan
## What Changed
- Widen ACPX failure diagnostics
(`packages/adapters/acpx-local/src/server/execute.ts`):
- Capture `err.name`, ACP code, `cause.message`, retryable flag, and a
5-frame stack preview into `errorMeta`.
- Promote phase-specific error codes: `ensure_session →
acpx_session_init_failed`, `configure_session →
acpx_session_config_failed`, `turn → acpx_turn_failed`, plus mapping for
`ACP_BACKEND_MISSING` / `ACP_BACKEND_UNAVAILABLE`.
- Set `verbose: true` on the ACPX runtime so its session-event log flows
through `ctx.onLog`.
- Capture child-process stderr via a wrapper-script tee into
`<stateDir>/run-stderr/<runId>.log`, inline the tail into the
`acpx.error` payload as `childStderrTail`, and forward it through
`ctx.onLog("stderr", …)` so it lands in the heartbeat `stderrExcerpt`
column (existing redaction applies).
- Set the model via `ANTHROPIC_MODEL` env for the `claude` agent instead
of `set_config_option(model, …)`. The ACP server's `set_config_option`
handler validates against an internal allowlist and rejects bare IDs
like `claude-opus-4-7`. `ANTHROPIC_MODEL` is read during initialization
and bypasses that check.
- Seed `<worktree>/.claude/settings.local.json` before spawning
`claude-agent-acp` (the seamless-API fix). Since `claude-agent-acp`
hard-codes `settingSources: ["user", "project", "local"]` and "local"
has the highest precedence:
- Set `permissions.defaultMode: "default"`, but **only** if the user's
value is missing or `"dontAsk"` (the broken case). Other modes like
`acceptEdits`/`plan` are preserved.
- Pre-allow Paperclip's Bash surface (`Bash(curl:*)`, `Bash(env:*)`,
`Bash(<cwd>/scripts/paperclip-issue-update.sh:*)`,
`Bash(<cwd>/scripts/paperclip:*)`).
- Widen `permissions.additionalDirectories` to include `stateDir`,
`agentHome`, and the per-company instance root
(`~/.paperclip/instances/<id>/companies/<companyId>`). Scoped to this
company only — does not expose other tenants.
- Existing user entries are merged, not replaced. The resolved roots are
folded into the session fingerprint so warm-session handles invalidate
when they change.
- Sync the existing server-side integration test
(`server/src/__tests__/acpx-local-execute.test.ts`) to assert
`acpx_session_init_failed` instead of the now-removed
`acpx_protocol_error` for `ACP_SESSION_INIT_FAILED` (a follow-up to
commit 1).
## Verification
- `pnpm --filter "@paperclipai/adapter-acpx-local" run typecheck` —
passes.
- `pnpm vitest run` in `packages/adapters/acpx-local` — 35/35 pass,
includes 4 new tests covering the settings.local.json write path (claude
only, merge with pre-existing content, `dontAsk` override, codex no-op).
- `pnpm vitest run src/__tests__/acpx-local-execute.test.ts` in
`server/` — 15/15 pass after the test-sync commit.
- End-to-end manual verification (PAPA-401): the `Claude Local ACPX`
agent that previously hit "restricted environment" now successfully
reaches the Paperclip API, opens its worktree, posts structured plan
comments, and flips the issue to `in_review` without any external
configuration.
## Risks
- **Low**, scoped to the `acpx_local` adapter. The settings.local.json
write is per-worktree (worktrees live under
`.paperclip/worktrees/<issue>/`) and only triggers when `acpxAgent ===
"claude"`. Existing user content is merged with `[...existing,
...paperclip]` and deduped — nothing is overwritten outright.
- The `defaultMode` override is intentionally narrow: it only flips
`"dontAsk"` (which silently denies every tool and is the root cause) to
`"default"`. Users who explicitly picked `acceptEdits`, `plan`, or any
other mode keep their choice.
- Stderr capture goes through the existing `log-redaction` pass before
persisting, so `PAPERCLIP_API_KEY` and similar secrets in the wrapper
env don't leak into heartbeat logs.
> For core feature work, check [`ROADMAP.md`](ROADMAP.md) first and
discuss it in `#dev` before opening the PR. Feature PRs that overlap
with planned core work may need to be redirected — check the roadmap
first. See `CONTRIBUTING.md`.
## Model Used
- Claude Opus 4.7 (`claude-opus-4-7`), running in the `claude_local`
adapter via Paperclip's harness. Extended thinking enabled, tool use
enabled.
## Checklist
- [x] I have included a thinking path that traces from project context
to this change
- [x] I have specified the model used (with version and capability
details)
- [x] I have checked ROADMAP.md and confirmed this PR does not duplicate
planned core work
- [x] I have run tests locally and they pass
- [x] I have added or updated tests where applicable
- [ ] If this change affects the UI, I have included before/after
screenshots — N/A (adapter-only)
- [ ] I have updated relevant documentation to reflect my changes — no
user-facing docs changed; internal commentary in the code change
explains the SDK constraints
- [x] I have considered and documented any risks above
- [ ] I will address all Greptile and reviewer comments before
requesting merge
---------
Co-authored-by: Paperclip <noreply@paperclip.ing>
## Thinking Path
> - Paperclip orchestrates AI agents for zero-human companies by keeping
task ownership, approvals, and operator control inside one control
plane.
> - Agent permissions and plugin-hosted company settings sit on the
boundary between autonomy and governance.
> - V1 needs scoped task assignment rules, plugin extension points, and
clearer company access surfaces without weakening company boundaries.
> - The branch builds the core authorization service, plugin SDK/host
APIs, and UI simplifications needed to support those controls.
> - Paperclip EE plugin surfaces were intentionally moved out of this
core PR per review direction, so this PR now carries only the public
core/plugin infrastructure work.
> - The latest updates preserve the PAP-9937 branch changes that belong
in this PR, remove the `design/` artifacts, and exclude the experimental
`plugin-briefs` package.
> - Greptile feedback was applied through the authorization/audit paths
and the final cleanup commit was re-reviewed at 5/5 with no unresolved
Greptile threads.
> - The benefit is safer assignment control with extension hooks for
richer permission products while preserving simple defaults for normal
operators.
## What Changed
- Added scoped task-assignment authorization decisions and routed
issue/agent assignment mutations through the authorization service.
- Added plugin SDK and host APIs for company settings slots,
authorization policy/grant management, assignment previews, and bridge
invocation scope propagation.
- Simplified core company access UI and moved advanced controls behind
plugin-provided settings surfaces.
- Added retry-now affordances for blocked issue next-step notices.
- Added protected-assignment enforcement for persisted
agent/project/issue policies, including explicit-grant fallback
behavior.
- Added incremental principal-access compatibility backfill for active
agent memberships and role-default human permission grants.
- Added the Markdown code block wrap action fix from the latest branch
changes.
- Removed `design/` artifacts from the PR and removed
`packages/plugins/plugin-briefs` from the final diff.
- Addressed Greptile feedback for plugin actor sanitization, legacy
membership handling, audit pagination, unknown grant-scope metadata, and
startup test mocks.
## Verification
- `pnpm exec vitest run server/src/__tests__/access-service.test.ts
server/src/__tests__/company-portability.test.ts` -> 2 files passed, 54
tests passed.
- `pnpm exec vitest run
server/src/__tests__/server-startup-feedback-export.test.ts
server/src/__tests__/access-service.test.ts
server/src/__tests__/company-portability.test.ts` -> 3 files passed, 62
tests passed.
- `pnpm exec vitest run
server/src/__tests__/authorization-service.test.ts
server/src/__tests__/plugin-access-authorization-host-services.test.ts
server/src/__tests__/server-startup-feedback-export.test.ts` -> 3 files
passed, 28 tests passed.
- `pnpm --filter @paperclipai/server typecheck` -> passed.
- `git diff --check` -> passed.
- `node ./scripts/check-docker-deps-stage.mjs` -> passed.
- `CI=true pnpm install --frozen-lockfile --ignore-scripts` -> passed
with no lockfile update.
- `pnpm exec vitest run
ui/src/components/MarkdownBody.interaction.test.tsx` -> 1 test passed.
- `git ls-files design packages/plugins/plugin-briefs | wc -l` -> 0.
- GitHub CI on `40cd83b53` -> all checks passed, merge state `CLEAN`.
- Greptile on `40cd83b53` -> 5/5, 102 files reviewed, 0
comments/annotations added, 0 unresolved review threads.
- Confirmed the PR diff contains no `design/`,
`packages/plugins/plugin-briefs`, `pnpm-lock.yaml`, or
`.github/workflows` changes.
## Risks
- Medium: task assignment authorization paths are behaviorally stricter
for protected/private policy data, so existing plugin-authored policies
may block assignment until explicit grants or approval flows are
configured.
- Medium: plugin-host authorization APIs expand the surface area
available to trusted plugins and need careful review for company
scoping.
- Low: startup now performs a principal-access compatibility backfill,
but the migration and runtime backfill use conflict-tolerant inserts.
> For core feature work, check [`ROADMAP.md`](ROADMAP.md) first and
discuss it in `#dev` before opening the PR. Feature PRs that overlap
with planned core work may need to be redirected — check the roadmap
first. See `CONTRIBUTING.md`.
## Model Used
- OpenAI Codex, GPT-5 coding agent, tool-enabled workflow with shell,
git, and GitHub CLI access.
## Checklist
- [x] I have included a thinking path that traces from project context
to this change
- [x] I have specified the model used (with version and capability
details)
- [x] I have checked ROADMAP.md and confirmed this PR does not duplicate
planned core work
- [x] I have run tests locally and they pass
- [x] I have added or updated tests where applicable
- [x] If this change affects the UI, I have included before/after
screenshots
- [x] I have updated relevant documentation to reflect my changes
- [x] I have considered and documented any risks above
- [x] I will address all Greptile and reviewer comments before
requesting merge
---------
Co-authored-by: Paperclip <noreply@paperclip.ing>
## Thinking Path
> - Paperclip is the control plane for autonomous AI companies, so agent
work needs visible ownership, recovery, and operator controls.
> - This local branch had accumulated several related control-plane
reliability and operator-experience fixes across recovery actions,
watchdog folding, model-profile defaults, mentions, markdown editing,
plugin launchers, and small UI polish.
> - The branch needed to be converted into a PR against the current
`origin/master` without losing dirty work or including lockfile/workflow
churn.
> - The safest standalone shape is a single rollup PR because the
recovery/server/UI files overlap heavily across the local commits and
splitting would create avoidable conflicts.
> - This pull request replays the local branch onto latest
`origin/master`, preserves the uncommitted work as logical commits, and
adds a Zod 4 validator compatibility fix found during verification.
> - The benefit is that the May 17 local branch can be reviewed and
merged as one coherent, conflict-free branch under the 100-file Greptile
limit.
## What Changed
- Rebased the local May 17 branch work onto current `origin/master` in a
dedicated worktree.
- Preserved and committed previously dirty changes for recovery retry
handling, plugin/sidebar launcher polish, and `.herenow` ignores.
- Added recovery-action behavior for returning source issues to `todo`
when retrying source-scoped recovery.
- Included the existing local recovery/liveness/watchdog fold, Codex
cheap-profile, markdown/mention, duplicate-agent, and UI polish commits
from the branch.
- Normalized shared validator `z.record(...)` schemas to explicit
string-key records for Zod 4 compatibility.
- Confirmed the PR has no `pnpm-lock.yaml` or `.github/workflows/*`
changes and stays below the 100-file Greptile limit.
## Verification
- `pnpm install --frozen-lockfile --ignore-scripts`
- `npm run install` in
`node_modules/.pnpm/sqlite3@5.1.7/node_modules/sqlite3` to build the
local native sqlite3 binding after installing with scripts disabled
- `pnpm exec vitest run packages/shared/src/validators/issue.test.ts
packages/shared/src/project-mentions.test.ts
packages/adapter-utils/src/server-utils.test.ts
server/src/__tests__/heartbeat-model-profile.test.ts
server/src/__tests__/issue-recovery-actions.test.ts
server/src/__tests__/issue-agent-mutation-ownership-routes.test.ts
server/src/__tests__/heartbeat-active-run-output-watchdog.test.ts
server/src/__tests__/plugin-local-folders.test.ts
ui/src/components/IssueRecoveryActionCard.test.tsx
ui/src/components/Sidebar.test.tsx
ui/src/components/SidebarAccountMenu.test.tsx
ui/src/components/IssueProperties.test.tsx
ui/src/components/MarkdownEditor.test.tsx
ui/src/components/MarkdownBody.test.tsx
ui/src/lib/duplicate-agent-payload.test.ts
ui/src/pages/Routines.test.tsx`
- First pass: 13 files passed with 201 passing tests; 3 server files
failed before sqlite3 native binding was built.
- After rebuilding sqlite3:
`server/src/__tests__/heartbeat-model-profile.test.ts`,
`server/src/__tests__/issue-recovery-actions.test.ts`, and
`server/src/__tests__/heartbeat-active-run-output-watchdog.test.ts`
passed/loaded; embedded Postgres tests were skipped by the local host
guard.
- `pnpm --filter @paperclipai/shared typecheck`
- `pnpm --filter @paperclipai/adapter-utils typecheck`
- `pnpm --filter @paperclipai/server typecheck`
- `pnpm --filter @paperclipai/ui typecheck`
## Risks
- Medium risk: this is a broad rollup PR across recovery semantics,
server tests, shared validators, and UI surfaces.
- Some embedded Postgres tests skipped locally due the host guard, so CI
should provide the stronger database-backed signal.
- UI changes were covered by component tests, but no browser screenshot
was captured in this PR creation pass.
- This branch may overlap with existing recovery/liveness PR work; merge
this PR independently or restack/close overlapping branches rather than
merging duplicate implementations together.
> For core feature work, check [`ROADMAP.md`](ROADMAP.md) first and
discuss it in `#dev` before opening the PR. Feature PRs that overlap
with planned core work may need to be redirected — check the roadmap
first. See `CONTRIBUTING.md`.
## Model Used
- OpenAI Codex, GPT-5-based coding agent, tool-enabled local repository
and GitHub workflow, medium reasoning effort.
## Checklist
- [x] I have included a thinking path that traces from project context
to this change
- [x] I have specified the model used (with version and capability
details)
- [x] I have checked ROADMAP.md and confirmed this PR does not duplicate
planned core work
- [x] I have run tests locally and they pass
- [x] I have added or updated tests where applicable
- [ ] If this change affects the UI, I have included before/after
screenshots
- [x] I have updated relevant documentation to reflect my changes
- [x] I have considered and documented any risks above
- [x] I will address all Greptile and reviewer comments before
requesting merge
---------
Co-authored-by: Paperclip <noreply@paperclip.ing>
## Thinking Path
> - Paperclip orchestrates AI agents for zero-human companies
> - The `grok-local` adapter streams reasoning text to the issue
"Working..." panel as the grok CLI runs
> - The `grok` CLI's `--output-format streaming-json` mode silently
drops the `\n` separator between reasoning turns around tool calls
> - Consecutive `thought` chunks (e.g. `` "`" `` followed by `"The"`)
arrive with no intervening whitespace event, so the UI's `delta: true`
concatenator merged them into run-on text like `"…planningGreat, now I
have the issue descriptionThe only co"`
> - This PR adds a small turn-boundary helper that detects sentence
boundaries in the upstream `thought` stream and inserts a single `\n`
only when the previous chunk ended with sentence punctuation (or a
balanced closing backtick) AND the next chunk begins a new uppercase
sentence
> - The benefit is readable streaming reasoning in the UI without
changing how completed messages are stored
## What Changed
- Added `packages/adapters/grok-local/src/shared/turn-boundary.ts` with
per-stream state (last chunk + backtick parity) and a
`restoreTurnBoundary()` helper that inserts `\n` only between balanced,
sentence-terminated `thought` chunks
- Wired the helper into `parseGrokJsonl` (server) and added a new
`createGrokStdoutParser` factory used by `grokLocalUIAdapter` for the
live "Working..." panel
- Added focused tests in `shared/turn-boundary.test.ts`, plus regression
assertions in `server/parse.test.ts` and `ui/parse-stdout.test.ts`
## Verification
- `pnpm --filter @paperclip/grok-local test` — 23/23 adapter tests pass
- `pnpm --filter @paperclip/grok-local typecheck` and UI typecheck —
clean
- Replayed an actual broken `grok 0.1.210` stream from the report;
previously-merged boundaries (`` `ls`The ``, `returned:Confirmed`) now
render with a separating newline; chunks inside un-closed backtick spans
are left alone
## Risks
- Low risk. Boundary insertion only fires when prev ends with
`.`/`!`/`?`/balanced `` ` `` and next begins with an uppercase ≥2-char
word, with no whitespace on either side. Worst case: a rare missed split
or a misplaced newline inside reasoning — both purely cosmetic and
confined to the live streaming panel.
## Model Used
- Claude Opus 4.7 (claude-opus-4-7), Anthropic, extended thinking + tool
use via Claude Code
## Checklist
- [x] I have included a thinking path that traces from project context
to this change
- [x] I have specified the model used (with version and capability
details)
- [x] I have checked ROADMAP.md and confirmed this PR does not duplicate
planned core work
- [x] I have run tests locally and they pass
- [x] I have added or updated tests where applicable
- [ ] If this change affects the UI, I have included before/after
screenshots
- [x] I have updated relevant documentation to reflect my changes
- [x] I have considered and documented any risks above
- [x] I will address all Greptile and reviewer comments before
requesting merge
---------
Co-authored-by: Paperclip <noreply@paperclip.ing>
## Thinking Path
> - Paperclip orchestrates AI agents for zero-human companies, so
adapter quality directly affects what runtimes the control plane can
supervise.
> - Local CLI adapters are one of the core execution surfaces because
they turn real coding tools into Paperclip-managed employees with
heartbeats, transcripts, and reviewability.
> - Grok Build was installed on the Paperclip host, but Paperclip had no
built-in `grok_local` adapter, so the runtime could not be configured
through the normal server/UI/CLI adapter path.
> - That gap needed to be closed with the same built-in registry,
environment diagnostics, transcript parsing, and skill/instructions
behavior that the other local adapters already rely on.
> - After the initial adapter landed, a real follow-up run showed that
Grok streaming text was being rendered one fragment per line, which made
transcripts harder to read even though the runtime itself was working.
> - This pull request adds the built-in `grok_local` adapter end-to-end
and then fixes the transcript parser so streamed Grok output is
coalesced into readable assistant/thinking blocks.
> - The benefit is that Grok Build becomes a first-class Paperclip
runtime with a usable operator experience instead of a partially wired
runtime with noisy transcript output.
## What Changed
- Added a new built-in `@paperclipai/adapter-grok-local` package with
server, UI, and CLI entrypoints.
- Implemented Grok execution, session handling, environment diagnostics,
config building, skill syncing, and parser coverage inside the new
adapter package.
- Registered `grok_local` across the built-in adapter inventories and
capability/display metadata in server, UI, CLI, and shared constants.
- Added adapter route coverage for the new built-in type.
- Fixed Grok transcript readability by emitting streamed `text` and
`thought` fragments as deltas so the shared transcript builder coalesces
them into readable message blocks.
- Added regression tests for the Grok parser and transcript coalescing
behavior.
## Verification
- `pnpm vitest run
packages/adapters/grok-local/src/ui/parse-stdout.test.ts
ui/src/adapters/transcript.test.ts`
- `pnpm --filter @paperclipai/adapter-grok-local build`
- Manual runtime verification on the Paperclip host during
implementation and follow-up review:
- confirmed the Grok CLI was installed and authenticated
- confirmed the worktree dev server could be restarted cleanly and
health-checked after the parser follow-up
- No screenshots attached. This change is primarily adapter plumbing
plus transcript formatting behavior; reviewers can verify via the
Grok-backed run surfaces directly.
## Risks
- This adds a new built-in adapter, so any missed registration surface
could create inconsistencies between server, UI, and CLI behavior.
- The adapter depends on Grok Build's current event/output shape; if
upstream Grok streaming JSON changes, transcript parsing or session
extraction may need follow-up updates.
- The transcript readability fix intentionally changes how Grok
fragments are grouped, so any downstream code that implicitly expected
one entry per fragment would behave differently.
> For core feature work, check [`ROADMAP.md`](ROADMAP.md) first and
discuss it in `#dev` before opening the PR. Feature PRs that overlap
with planned core work may need to be redirected — check the roadmap
first. See `CONTRIBUTING.md`.
## Model Used
- OpenAI Codex via Paperclip `codex_local` agent runtime.
- GPT-5-class coding model with tool use, shell execution, file editing,
and repo inspection enabled.
- Exact backend model ID/context window were not surfaced to the agent
in this Paperclip session.
## Checklist
- [x] I have included a thinking path that traces from project context
to this change
- [x] I have specified the model used (with version and capability
details)
- [x] I have checked ROADMAP.md and confirmed this PR does not duplicate
planned core work
- [x] I have run tests locally and they pass
- [x] I have added or updated tests where applicable
- [ ] If this change affects the UI, I have included before/after
screenshots
- [ ] I have updated relevant documentation to reflect my changes
- [x] I have considered and documented any risks above
- [x] I will address all Greptile and reviewer comments before
requesting merge
## Thinking Path
> - Paperclip orchestrates AI agents for zero-human companies.
> - Remote-managed adapters need sandbox/environment execution to behave
like real agent runs, not just local host probes.
> - The Cloudflare sandbox path was the weakest leg in the SSH +
Cloudflare QA matrix because bridge execution could truncate output,
time out long-running installs, and under-provision the worker instance.
> - That made several adapters fail for reasons unrelated to their
actual business logic, which blocks confidence in Paperclip's non-local
environment model.
> - This pull request hardens the Cloudflare bridge/runtime path and
adjusts sandbox probe budgets so adapter verification matches the
measured behavior of the fixed environment.
> - It also corrects the Pi sandbox install command so the QA matrix
exercises a real, supported install path.
> - The benefit is a materially more reliable SSH + Cloudflare adapter
matrix with fewer false negatives and clearer failure boundaries.
## What Changed
- Switched the Cloudflare bridge worker instance type to `standard-2`
for the QA-matrix execution path.
- Raised Cloudflare bridge/plugin-worker timeout budgets and added SSE
keepalives so long-running install/exec calls can complete instead of
dying at the transport layer.
- Fixed Cloudflare bridge-channel command handling to avoid dropped
final stdout chunks on short-lived execs.
- Made Claude, OpenCode, and Cursor sandbox probe timeouts
configurable/sandbox-aware, then tightened the defaults to the measured
post-fix range.
- Updated the Pi sandbox install command to use the package currently
installed by the official `pi.dev` installer, pinned to a specific npm
version.
- Added/updated tests around Cloudflare bridge behavior and adapter
sandbox probe paths.
## Verification
- `pnpm --filter @paperclipai/adapter-claude-local typecheck`
- `pnpm --filter @paperclipai/adapter-opencode-local typecheck`
- `pnpm --filter @paperclipai/adapter-cursor-local typecheck`
- `pnpm vitest run packages/adapters/cursor-local
packages/adapters/claude-local packages/adapters/opencode-local
packages/adapters/pi-local packages/plugins/sandbox-providers/cloudflare
server/src/services/__tests__/plugin-worker-manager.test.ts`
- Manual QA on the dedicated dev instance using the SSH + Cloudflare
environment matrix (`ENV-29` through `ENV-40`). Clean end-to-end passes:
SSH `claude_local`, `codex_local`, `cursor`, `gemini_local`; Cloudflare
`claude_local`, `codex_local`, `cursor`, `gemini_local`.
## Risks
- Cloudflare sandbox cost increases because the bridge worker now runs
on `standard-2` instead of `lite`.
- Higher timeout ceilings can delay surfacing truly hung Cloudflare
bridge calls, even though they remove transport-level false negatives.
- The manual heartbeat matrix still exposed follow-on
execution/sync/disposition bugs in `opencode_local` and `pi_local`;
those are not fixed by this PR.
## Model Used
- OpenAI `gpt-5.4` via Paperclip `codex_local`, reasoning effort `high`,
tool use enabled, repo search enabled.
## Checklist
- [x] I have included a thinking path that traces from project context
to this change
- [x] I have specified the model used (with version and capability
details)
- [x] I have checked ROADMAP.md and confirmed this PR does not duplicate
planned core work
- [x] I have run tests locally and they pass
- [x] I have added or updated tests where applicable
- [x] If this change affects the UI, I have included before/after
screenshots (not applicable)
- [x] I have updated relevant documentation to reflect my changes (not
applicable)
- [x] I have considered and documented any risks above
- [x] I will address all Greptile and reviewer comments before
requesting merge
---------
Co-authored-by: Paperclip <noreply@paperclip.ing>
## Thinking Path
> - Paperclip orchestrates AI agents through a control plane while
adapters execute work in local, remote, or sandboxed runtimes.
> - Remote sandbox execution depends on a strict host-versus-remote
workspace boundary: the host prepares/restores files, while the adapter
command runs inside the sandbox cwd.
> - Jannes' PR #5823 identified host-side failure modes that were not
covered by replacement PR #5822.
> - Persisting a remote pod cwd in session params could poison the next
host heartbeat resume and make Paperclip inspect or upload system temp
roots.
> - Plugin sandbox providers also need a narrow way to receive
model-provider API keys without exposing the full server environment to
every plugin worker.
> - This pull request ports the host-side fixes from #5823 in the
current codebase style, with focused regression coverage.
> - The benefit is safer remote sandbox resumes and plugin worker
environment handling without broadening core plugin privileges.
## What Changed
- Persist host workspace cwd, not remote sandbox cwd, in `claude_local`
session params while retaining remote execution identity metadata.
- Reject saved session cwds that point at system roots before heartbeat
falls back to agent home workspace.
- Skip sockets, FIFOs, devices, and other non-file entries during
workspace restore snapshot capture/comparison.
- Pass a small model-provider API-key allowlist only to plugins
declaring `environment.drivers.register`.
- Added focused regression tests for remote Claude session params,
unsafe session cwd detection, plugin worker env filtering, and non-file
snapshot entries.
Credits: ports host-side fixes from Jannes' #5823.
## Verification
- `pnpm vitest run
packages/adapter-utils/src/workspace-restore-merge.test.ts
server/src/services/session-workspace-cwd.test.ts
server/src/__tests__/claude-local-execute.test.ts
server/src/__tests__/plugin-database.test.ts` (25 passed, 7 skipped by
existing embedded-Postgres host guard)
- `pnpm --filter @paperclipai/adapter-utils typecheck`
- `pnpm --filter @paperclipai/adapter-claude-local typecheck`
- `pnpm --filter @paperclipai/server typecheck`
## Risks
- Low risk: changes are scoped to remote sandbox/session metadata,
workspace snapshot filtering, and plugin worker env setup.
- Sandbox-provider plugins now receive only the explicit model-provider
key allowlist; any provider needing another key name will need a
deliberate allowlist update.
> For core feature work, check [`ROADMAP.md`](ROADMAP.md) first and
discuss it in `#dev` before opening the PR. Feature PRs that overlap
with planned core work may need to be redirected — check the roadmap
first. See `CONTRIBUTING.md`.
## Model Used
- OpenAI Codex, GPT-5-based coding agent, tool-enabled local code
execution and repository editing.
## Checklist
- [x] I have included a thinking path that traces from project context
to this change
- [x] I have specified the model used (with version and capability
details)
- [x] I have checked ROADMAP.md and confirmed this PR does not duplicate
planned core work
- [x] I have run tests locally and they pass
- [x] I have added or updated tests where applicable
- [x] If this change affects the UI, I have included before/after
screenshots
- [x] I have updated relevant documentation to reflect my changes
- [x] I have considered and documented any risks above
- [x] I will address all Greptile and reviewer comments before
requesting merge
---------
Co-authored-by: Paperclip <noreply@paperclip.ing>
## Thinking Path
> - Paperclip orchestrates AI agents for zero-human companies, including
running adapter CLIs inside remote sandboxes
> - The QA matrix in PAPA-316 spins up local-runtime adapters
(claude/gemini/opencode) against both SSH and the new exe.dev sandbox
provider, and "Test" exercises the same install + probe path the real
runtime uses
> - On exe.dev the QA matrix failed at three different points:
SSH/sandbox secret refs would not resolve, gemini-local could not find
npm, and opencode-local installed a binary that was not on the
probe-shell PATH
> - These are all environment-shape issues the runtime should handle,
not regressions in any individual adapter, so they need to be fixed in
the shared install/resolve layer before the matrix can pass
> - This pull request wires the environment id through to secret-ref
resolution, bootstraps npm from a portable Node tarball when the sandbox
image lacks Node, and symlinks the opencode binary into a directory that
non-login shells see
> - The benefit is that the QA matrix passes end-to-end on exe.dev, and
any future sandbox provider that ships without Node or relies on rc-file
PATH wiring gets the same fixes for free
## What Changed
- `server/src/services/environment-execution-target.ts`: pass the
environment `id` into `resolveEnvironmentDriverConfigForRuntime` for
both the sandbox and SSH branches, so `privateKeySecretRef` /
sandbox-provider secret refs (e.g. exe.dev `apiKey`) can resolve against
the secret store at runtime instead of throwing `Runtime secret
resolution requires an environment id`.
- `packages/adapter-utils/src/sandbox-install-command.ts`: extend
`buildSandboxNpmInstallCommand` with an `ENSURE_NPM_PREAMBLE` that, when
`npm` is missing, downloads a portable Node v22 tarball into
`$HOME/.local` and sets `PAPERCLIP_NPM_BOOTSTRAPPED=1` so the install
step skips sudo (sudo's `secure_path` would lose the freshly-installed
`npm` in `$HOME/.local/bin`). Distro-packaged Node from apt-get is
intentionally avoided because it tends to be too old to parse modern JS
syntax used by `@google/gemini-cli`.
- `packages/adapters/gemini-local/src/index.ts`: switch the hardcoded
`npm install -g @google/gemini-cli` to `buildSandboxNpmInstallCommand`,
so gemini-local picks up the same sudo-aware + npm-bootstrap behavior as
the other local adapters.
- `packages/adapters/opencode-local/src/index.ts`: append a step to the
install command that symlinks `$HOME/.opencode/bin/opencode` into
`$HOME/.local/bin`. The upstream installer only adds `~/.opencode/bin`
to PATH via `~/.bashrc`, which non-login `sh -c` probe invocations do
not source.
- `packages/adapter-utils/src/sandbox-install-command.test.ts`: cover
the new preamble plus the unchanged root/sudo/user-prefix branches.
## Verification
- `cd packages/adapter-utils && npm test -- sandbox-install-command`
(passes; new "bootstraps npm from a portable Node tarball when missing"
case is included).
- Manual: ran the in-app `Test` action against the QA matrix dev
instance for `QA exe.dev Claude`, `QA exe.dev Gemini`, and `QA exe.dev
OpenCode` — all three now report `status=pass` including the hello
probe. `QA SSH Claude` also passes; without the environment-id fix, SSH
resolution threw before the wrapper / install fixes could run.
- Suggested reviewer check: re-run the matrix on a fresh exe.dev
environment and confirm the install step no longer hits `npm: command
not found` for gemini and the opencode probe no longer hits `opencode:
command not found`.
## Risks
- Low/medium. The npm bootstrap pins Node `v22.11.0` from
`nodejs.org/dist`; if that URL becomes unreachable the install will fail
with a clear `curl` error rather than corrupting state. The bootstrap
path is only taken when `npm` is genuinely missing, so existing sandbox
images that ship with Node are unaffected.
- The opencode symlink uses `ln -sf` into `$HOME/.local/bin`, which is
created with `mkdir -p`; idempotent on re-install.
- The `id` change is a strict additive: callers previously got
`undefined` and only the secret-ref code paths actually read it. No
behavior change for environments without secret refs.
## Model Used
- Claude (Anthropic), `claude-opus-4-7`, with extended thinking and tool
use enabled. Iterated through the Paperclip QA matrix harness; no other
model assisted.
## Checklist
- [x] I have included a thinking path that traces from project context
to this change
- [x] I have specified the model used (with version and capability
details)
- [x] I have checked ROADMAP.md and confirmed this PR does not duplicate
planned core work
- [x] I have run tests locally and they pass
- [x] I have added or updated tests where applicable
- [ ] If this change affects the UI, I have included before/after
screenshots (n/a — runtime/install path only)
- [x] I have considered and documented any risks above
- [x] I will address all Greptile and reviewer comments before
requesting merge
---------
Co-authored-by: Paperclip <noreply@paperclip.ing>
> _Stacked on top of #5685 (Harden remote sandbox runtime). Diff against
master includes commits from earlier PRs in the stack — review focuses
on the new commit only._
## Thinking Path
> - Paperclip orchestrates AI agents for zero-human companies
> - The cursor-local adapter wraps the Cursor Agent CLI so a Paperclip
workflow can drive it inside a sandbox
> - When the adapter runs in a remote sandbox, the Cursor Agent CLI
installs under `$HOME/.local/bin/cursor-agent` (or wherever
`$XDG_BIN_HOME` points), not on the global PATH
> - The existing post-install resolution assumed `cursor-agent` would
resolve via the sandbox's login shell PATH after `npm install -g`, which
fails on sandboxes where the install lands in a user-prefixed directory
that isn't on PATH at probe time
> - This pull request resolves the agent CLI from the cursor binary's
own directory (`dirname "$(command -v cursor)"`) so the install probe
and execute path agree on a real binary location
> - The benefit is that cursor-local works correctly on any sandbox
provider where `npm install` lands in a user-prefixed directory
## What Changed
- `packages/adapters/cursor-local/src/server/remote-command.ts`: resolve
the cursor-agent binary from the cursor bin directory after install,
instead of relying on PATH.
- `packages/adapters/cursor-local/src/server/test.ts`: corresponding
probe tweak.
- `packages/adapters/cursor-local/src/server/test.test.ts` (new) +
`remote-command.test.ts`: focused coverage that exercises the install +
resolve path against a sandbox runner that places the binary in a
user-prefixed directory.
## Verification
- `pnpm exec vitest run --no-coverage
packages/adapters/cursor-local/src/server/test.test.ts
packages/adapters/cursor-local/src/server/remote-command.test.ts
packages/adapters/cursor-local/src/server/execute.test.ts`
All passing locally.
## Risks
- Local cursor-local runs are unaffected — the resolution change only
kicks in for the sandbox install path.
- Low risk; isolated to one adapter.
## Model Used
- Provider: Anthropic
- Model: Claude Opus 4.7 (1M context)
- Capabilities used: tool use (Read/Edit/Bash), no code execution beyond
local repo commands
## Checklist
- [x] I have included a thinking path that traces from project context
to this change
- [x] I have specified the model used (with version and capability
details)
- [x] I have checked ROADMAP.md and confirmed this PR does not duplicate
planned core work
- [x] I have run tests locally and they pass
- [x] I have added or updated tests where applicable
- [ ] If this change affects the UI, I have included before/after
screenshots — N/A, no UI change
- [x] I have updated relevant documentation to reflect my changes
- [x] I have considered and documented any risks above
- [x] I will address all Greptile and reviewer comments before
requesting merge
Co-authored-by: Paperclip <noreply@paperclip.ing>
## Thinking Path
> - Paperclip orchestrates AI agents for zero-human companies
> - Each agent runs inside a sandbox environment so its CLI is isolated
from the host
> - Sandbox-backed adapter runs go through a small set of shared helpers
— `ensureAdapterExecutionTargetCommandResolvable`, the sandbox callback
bridge runner, and per-adapter `SANDBOX_INSTALL_COMMAND` strings
> - When standing up new sandbox provider plugins, the existing helpers
timed out, missed install fallbacks, or leaned on assumptions that only
held for E2B
> - Local adapters (`claude-local`, `codex-local`, `gemini-local`,
`opencode-local`) needed slightly hardened probes so they could install
themselves and validate inside *any* remote sandbox transport, not just
E2B
> - This pull request bundles those runtime fixes so future sandbox
provider plugins inherit a working baseline
> - The benefit is that adding a new sandbox provider plugin no longer
requires touching adapter-utils or each local-adapter probe — the
supporting infra is already correct
## What Changed
- `packages/adapter-utils/src/execution-target.ts`: introduce
`DEFAULT_REMOTE_SANDBOX_ADAPTER_TIMEOUT_SEC = 1800` and
`resolveAdapterExecutionTargetTimeoutSec(...)`. Local and SSH adapters
keep the historical "0 means no adapter timeout" behavior;
sandbox-backed runs without an explicit `timeoutSec` get an explicit
30-minute default so remote installs and warm-up don't time out at the
per-RPC default. Plumbed `timeoutSec` through
`ensureAdapterExecutionTargetCommandResolvable` so install probes inside
a sandbox honor adapter-level overrides instead of the bridge's 5-minute
default.
- `packages/adapters/opencode-local/src/index.ts`: switch
`SANDBOX_INSTALL_COMMAND` from `npm install -g opencode-ai` to `curl
-fsSL https://opencode.ai/install | bash`. The npm package reifies four
large prebuilt-binary subpackages in parallel even though only one
matches the host arch; on bandwidth-constrained sandboxes that blew
through the 240s install budget. The official installer fetches one
arch-specific binary and adds `$HOME/.opencode/bin` to PATH via
`~/.bashrc`, which the sandbox-callback-bridge login-shell script
already sources.
- `packages/adapters/{claude,codex,gemini,opencode}-local/`: harden
remote-target probes — pass `--skip-git-repo-check` for Codex when
probing outside a repo, normalize permission flags for Claude, and add
`*.remote.test.ts` coverage that exercises the remote-sandbox path
explicitly for each adapter.
- `packages/adapter-utils/src/sandbox-install-command.{ts,test.ts}`
(new): add `buildSandboxNpmInstallCommand` helper.
`server/src/adapters/registry.ts` + new
`server/src/__tests__/adapter-registry.test.ts`: wire adapter install
commands so they fall back to a writable `$HOME/.local` prefix when
global install isn't available.
- `server/src/__tests__/plugin-worker-manager.test.ts` + new
`server/src/__tests__/fixtures/plugin-worker-delayed.cjs`: pin per-call
timeout overrides so plugin worker exec calls honor the caller's timeout
instead of the worker's default.
## Verification
- `pnpm typecheck`
- `pnpm exec vitest run --no-coverage
packages/adapter-utils/src/execution-target-sandbox.test.ts
packages/adapter-utils/src/sandbox-install-command.test.ts`
- `pnpm exec vitest run --no-coverage
server/src/__tests__/plugin-worker-manager.test.ts
server/src/__tests__/adapter-registry.test.ts
server/src/__tests__/claude-local-adapter-environment.test.ts
server/src/__tests__/claude-local-execute.test.ts
server/src/__tests__/gemini-local-adapter-environment.test.ts`
- `pnpm exec vitest run --no-coverage
packages/adapters/codex-local/src/server/test.remote.test.ts
packages/adapters/opencode-local/src/server/test.remote.test.ts
packages/adapters/codex-local/src/server/codex-args.test.ts
packages/adapters/codex-local/src/server/execute.remote.test.ts
packages/adapters/gemini-local/src/server/execute.remote.test.ts`
All passing locally.
## Risks
- Touches shared `adapter-utils` and several `*-local` adapters. The
30-minute default applies only when both (a) the target is
`remote+sandbox` and (b) no `timeoutSec` is configured — local + SSH
paths are unchanged. New test coverage was added alongside each behavior
change to pin the contracts.
- Switching OpenCode's install command to the official installer is a
behavior change for any operator running OpenCode inside a remote
sandbox. Local installs are unaffected (the `SANDBOX_INSTALL_COMMAND`
only runs when an adapter is being installed inside a sandbox).
- Low risk overall — no migrations, no API surface change.
## Model Used
- Provider: Anthropic
- Model: Claude Opus 4.7 (1M context)
- Capabilities used: extended reasoning, tool use (Read/Edit/Bash/Grep),
no code execution beyond local repo commands
## Checklist
- [x] I have included a thinking path that traces from project context
to this change
- [x] I have specified the model used (with version and capability
details)
- [x] I have checked ROADMAP.md and confirmed this PR does not duplicate
planned core work
- [x] I have run tests locally and they pass
- [x] I have added or updated tests where applicable
- [ ] If this change affects the UI, I have included before/after
screenshots — N/A, no UI change
- [x] I have updated relevant documentation to reflect my changes
- [x] I have considered and documented any risks above
- [x] I will address all Greptile and reviewer comments before
requesting merge
Co-authored-by: Paperclip <noreply@paperclip.ing>
## Thinking Path
> - Paperclip orchestrates AI agents for zero-human companies
> - There are many adapter types, one per agent-runtime product (Claude,
Codex, OpenCode, Cursor local CLI, etc.)
> - Cursor shipped a public TypeScript SDK on 2026-04-29 that exposes
Cursor's full hosted-agent platform (cloud VMs, harness, MCP, skills,
hooks)
> - Paperclip had no first-class adapter for this — agents that wanted
to use Cursor's managed cloud runtime had to fall back to the local CLI
adapter, which loses the cloud session, streaming, and durable run model
> - This PR adds a new `cursor_cloud` adapter built directly on
`@cursor/sdk`, with Paperclip's heartbeat mapped to Cursor's
durable-agent + per-run model
> - The benefit is that any Paperclip agent can now drive a Cursor cloud
agent across heartbeats with native session reuse, streaming, and
cancellation, while Paperclip remains the source of truth for issue/task
state
## What Changed
- New built-in adapter package `packages/adapters/cursor-cloud` (15
files, ~1.7k LOC) backed by `@cursor/sdk` ^1.0.12
- `src/server/execute.ts` — SDK-first lifecycle: `Agent.create` /
`Agent.resume` / `Agent.getRun` / `agent.send` / `run.stream` /
`run.wait`, with session reuse keyed on the (runtime env type, env name,
repo set) tuple
- `src/server/session.ts` — codec for `cursorAgentId` + `latestRunId` +
repo metadata, persisted in `runtime.sessionParams`
- `src/server/test.ts` — environment probe via `Cursor.me()` and
optional model validation via `Cursor.models.list()`
- `src/ui/parse-stdout.ts` + `src/cli/format-event.ts` — normalize
Cursor SDK message types (`status`, `thinking`, `assistant`, `user`,
`tool_call`, `tool_result`, `result`) into Paperclip transcript events
for the UI and CLI
- Registrations: `packages/shared/src/constants.ts`,
`packages/adapter-utils/src/session-compaction.ts`,
`server/src/adapters/{registry,builtin-adapter-types}.ts`,
`ui/src/adapters/{registry,adapter-display-registry}.ts` +
`ui/src/adapters/cursor-cloud/index.ts`, `cli/src/adapters/registry.ts`,
plus workspace deps in `cli`/`server`/`ui` `package.json`
- `ui/src/components/AgentConfigForm.tsx` — hide local-Cursor
`mode`/thinking-effort field for `cursor_cloud` (different config
surface)
- 11 vitest tests covering execute paths (fresh create, matching-resume,
active-run reattach, non-finished result), session codec round-trip,
transcript parsing, and config building
## Verification
Reviewer steps:
```bash
pnpm install
pnpm --filter @paperclipai/adapter-cursor-cloud typecheck # → clean
pnpm vitest run packages/adapters/cursor-cloud # → 11/11 passing
```
End-to-end check against a real Cursor cloud agent (requires
`CURSOR_API_KEY` and Cursor GitHub-app install on the target repo):
1. Create a `cursor_cloud` agent in Paperclip with `repoUrl` set to the
test repo, `repoStartingRef: main`, and `env.CURSOR_API_KEY` set
2. Trigger a heartbeat → adapter calls `Agent.create({ cloud: { env: {
type: "cloud" }, repos: [...] } })`, streams events, terminates on
`finished`
3. Trigger a second heartbeat → adapter calls `Agent.resume` or
`agent.send` follow-up depending on prior-run state, reusing
`cursorAgentId`
4. The Paperclip UI/CLI transcript reflects Cursor `status` / `thinking`
/ `assistant` events as they stream
5. Cancellation from Paperclip maps to `run.cancel()` or Cloud API v1
`cancelRun` for cross-heartbeat cancellation
A direct-SDK smoke run against a real repo (devinfoley/my_test_project @
main) confirmed: `Cursor.me()` ok → `Agent.create` → `agent.send` →
`run.stream()` (30 events) → terminal status `finished` in ~11s.
## Risks
- **New adapter, additive only.** No existing adapter or registry is
replaced; current `cursor` local-CLI adapter is untouched. Default
behavior of any existing agent is unchanged.
- **External dependency on `@cursor/sdk`.** Cursor's SDK is v1.0.x and
may evolve. Mocked unit tests cover the public surface used here; if the
SDK breaks compatibility we update the adapter independently.
- **Cost/budget.** `cursor_cloud` runs on Cursor's billed cloud VMs;
operators must understand they are spending money outside Paperclip's
budget controls when they enable this adapter. Same shape as other
API-billed adapters.
- **No webhook support in V1.** The SDK already provides
stream/wait/cancel/reattach, so V1 does not require a public callback
URL. If a future use case needs out-of-band wakes, we add a Cloud API v1
webhook bridge as a separate change. This is called out in the issue
plan document.
- **Lockfile.** Per repo policy, `pnpm-lock.yaml` is intentionally not
in this PR — CI's lockfile workflow will update it on merge given the
manifest changes.
## Model Used
- Provider: Anthropic Claude (via Claude Code / Paperclip `claude_local`
adapter)
- Model: `claude-opus-4-7` (Claude Opus 4.7), knowledge cutoff January
2026
- Mode: standard tool-use with extended reasoning
- Context: ~200k token window
- Capabilities used: code generation, multi-file edits, shell/test
execution, GitHub PR workflow
## Checklist
- [x] I have included a thinking path that traces from project context
to this change
- [x] I have specified the model used (with version and capability
details)
- [x] I have checked ROADMAP.md and confirmed this PR does not duplicate
planned core work
- [x] I have run tests locally and they pass (11/11 in
`packages/adapters/cursor-cloud`)
- [x] I have added or updated tests where applicable (4 new test files,
11 cases)
- [ ] If this change affects the UI, I have included before/after
screenshots (the only UI change is hiding the local-Cursor mode field on
the `cursor_cloud` adapter — happy to attach a screenshot if the
reviewer wants one)
- [x] I have updated relevant documentation to reflect my changes (issue
plan document supersedes the pre-SDK design; tracked in PAPA-203)
- [x] I have considered and documented any risks above
- [x] I will address all Greptile and reviewer comments before
requesting merge
---------
Co-authored-by: Paperclip <noreply@paperclip.ing>
## Thinking Path
> - Paperclip orchestrates AI agents for zero-human companies.
> - The plugin system needs host contracts and runtime support before
large plugins can integrate cleanly.
> - The source branch mixed the LLM Wiki package with supporting
host/runtime work, managed plugin skills, root-level storage spaces, and
a bookmarks reference plugin.
> - [PAP-9173](/PAP/issues/PAP-9173) asked for the current branch to be
split by file boundary: plugin package separately from everything else.
> - [PAP-9188](/PAP/issues/PAP-9188) clarified that LLM Wiki may have
plugin-local spaces, but Paperclip core should not reorganize top-level
local storage into spaces.
> - Follow-up review clarified that the bookmarks example should not
ship in this PR either.
> - This pull request contains the
non-`packages/plugins/plugin-llm-wiki/` host/runtime work, keeps runtime
state under the selected Paperclip instance root, and no longer includes
the bookmarks example.
## What Changed
- Added/updated plugin host contracts, SDK types, worker RPC plumbing,
managed plugin skill support, and related server tests.
- Removed the bookmarks example plugin package and its
bundled-example/workspace references.
- Removed the root-level local spaces CLI/migration surface and restored
instance-root runtime defaults for config, db, logs, storage, secrets,
workspaces, projects, and adapter homes.
- Replaced shared root `space-paths` helpers with `home-paths` helpers
for core runtime storage.
- Tightened stranded recovery unique-conflict detection so concurrent
recovery scans reuse the raced recovery issue when Postgres errors are
wrapped.
- Kept `packages/plugins/plugin-llm-wiki/` out of this PR diff;
plugin-local spaces remain in the stacked plugin-only PR.
## Verification
- `pnpm exec vitest run cli/src/__tests__/data-dir.test.ts
cli/src/__tests__/home-paths.test.ts cli/src/__tests__/onboard.test.ts
packages/shared/src/home-paths.test.ts
packages/db/src/runtime-config.test.ts
server/src/__tests__/agent-instructions-service.test.ts
server/src/__tests__/claude-local-execute.test.ts
server/src/__tests__/codex-local-execute.test.ts`
- `pnpm exec vitest run packages/db/src/runtime-config.test.ts`
- `pnpm exec vitest run
server/src/__tests__/plugin-routes-authz.test.ts`
- `pnpm --filter @paperclipai/server typecheck`
- `pnpm exec vitest run
server/src/__tests__/heartbeat-process-recovery.test.ts -t "reuses the
raced stranded recovery issue"` skipped locally because embedded
Postgres did not initialize on this macOS temp host; the code path was
typechecked and is covered by Linux CI.
- Boundary check: no core references remain for `PAPERCLIP_SPACE_ID`,
`spaces migrate-default`, `@paperclipai/shared/space-paths`,
`registerSpacesCommands`, or the removed bookmarks example.
- Previous PR head `4f23e034` had green GitHub checks: `verify`, all
four serialized server shards, `e2e`, `Canary Dry Run`, `policy`, Snyk,
and `Greptile Review`. Current head `582f466d` is re-running checks
after the bookmarks deletion.
## Risks
- Plugin host changes touch shared runtime paths, so regressions would
most likely appear in adapter startup, plugin loading, or local dev path
defaults.
- Removing the bookmarks example also removes one demonstration of
plugin database namespaces plus local-folder persistence; remaining
plugin examples still cover bundled example discovery and plugin host
flows.
- The plugin package itself is intentionally deferred to the stacked
plugin-only PR, where LLM Wiki plugin-local spaces live.
- Existing installs that tested the transient root-level spaces CLI
should stop using it; this PR intentionally removes that unsupported
migration surface before merge.
> For core feature work, check [`ROADMAP.md`](ROADMAP.md) first and
discuss it in `#dev` before opening the PR. Feature PRs that overlap
with planned core work may need to be redirected — check the roadmap
first. See `CONTRIBUTING.md`.
## Model Used
- OpenAI GPT-5 Codex via Codex CLI, tool use and local code execution
enabled; context window not exposed.
## Checklist
- [x] I have included a thinking path that traces from project context
to this change
- [x] I have specified the model used (with version and capability
details)
- [x] I have checked ROADMAP.md and confirmed this PR does not duplicate
planned core work
- [x] I have run tests locally and they pass, except where noted above
for host-specific embedded Postgres initialization
- [x] I have added or updated tests where applicable
- [x] If this change affects the UI, I have included before/after
screenshots
- [x] I have updated relevant documentation to reflect my changes
- [x] I have considered and documented any risks above
- [x] I will address all Greptile and reviewer comments before
requesting merge
Stacked follow-up: PR #5592 contains only
`packages/plugins/plugin-llm-wiki/` and targets this branch.
---------
Co-authored-by: Paperclip <noreply@paperclip.ing>
## Thinking Path
> - Paperclip orchestrates AI agents for zero-human companies
> - The Cursor adapter spawns the Cursor CLI against local, SSH, and
sandbox execution targets; on a fresh sandbox lease, it has to resolve
where Cursor was installed
> - The previous resolver only looked for `~/.local/bin/cursor-agent`
even though the official installer (and the adapter's own
`SANDBOX_INSTALL_COMMAND`) sometimes lays the binary down as
`~/.local/bin/agent`, so a sandbox where the install ran successfully
would still fail to find the CLI
> - This pull request lets the resolver accept either basename and lets
the caller pass an optional `remoteSystemHomeDirHint` so a probe doesn't
pay the cost of a remote `printf $HOME` round-trip when the home
directory is already known
> - The benefit is sandboxed Cursor runs find the binary that the
install actually produced, and runtime probes are cheaper when the home
dir is already resolved
## What Changed
- `packages/adapters/cursor-local/src/server/remote-command.ts`: accept
either `agent` or `cursor-agent` as the preferred basename; new optional
`remoteSystemHomeDirHint` short-circuits the home-dir probe
- `packages/adapters/cursor-local/src/server/execute.ts`: thread the
home-dir hint through, prefer the resolved binary path, and shift the
effective execution cwd to the per-run managed subdirectory once the
runtime is prepared
- New `remote-command.test.ts` and `execute.test.ts` cover both
basenames, the hint short-circuit, and the cwd shift
- `packages/adapters/cursor-local/src/index.ts`: update doc string to
reflect the broader resolution
- `execute.remote.test.ts` updated to expect the managed-subdirectory
cwd shape introduced by the cwd shift
## Verification
- `pnpm vitest run --no-coverage --project
@paperclipai/adapter-cursor-local` — 6/6 passing
- `pnpm typecheck` clean
- Manual: a fresh sandbox lease with `npm install -g …`-installed Cursor
(binary lands as `~/.local/bin/agent`) now runs cleanly through the
adapter
## Risks
Low. Resolver is strictly broader (matches a superset of paths);
existing setups with `~/.local/bin/cursor-agent` continue to work. The
home-dir hint is opt-in; callers that don't pass it get the existing
probe behavior. Cursor's effective execution cwd now matches the rest of
the adapters (per-run managed subdirectory) — sessions previously rooted
at the workspace root will land in the new subdirectory.
## Model Used
Claude Opus 4.7 (1M context)
## Checklist
- [x] I have included a thinking path that traces from project context
to this change
- [x] I have specified the model used (with version and capability
details)
- [x] I have checked ROADMAP.md and confirmed this PR does not duplicate
planned core work
- [x] I have run tests locally and they pass
- [x] I have added or updated tests where applicable — new tests cover
both basenames + hint short-circuit + cwd shift
- [x] If this change affects the UI, I have included before/after
screenshots — N/A (no UI)
- [x] I have updated relevant documentation to reflect my changes
- [x] I have considered and documented any risks above
- [x] I will address all Greptile and reviewer comments before
requesting merge
---
> **Stacked PR.** Sits on top of #5445 (which sits on #5444). Cumulative
diff against `master` includes both of those PRs' content; the files
touched by *this* PR's commit are listed under "What Changed" above.
Will rebase onto `master` and force-push once the prerequisite PRs
merge.
## Thinking Path
> - Paperclip orchestrates AI agents for zero-human companies
> - Adapters expose a Test action that probes the configured runtime —
install, resolvability, hello — to give operators a fast yes/no on
whether an environment is healthy
> - The Codex test path was running its hello probe directly without
going through the managed-runtime preparation that production runs use,
so a healthy production setup could still report a probe failure
> - The plugin worker manager wasn't surfacing terminated workers
cleanly, leaving the runtime probe waiting on a dead worker until the
request timed out
> - This pull request routes the Codex test probe through
`prepareAdapterExecutionTargetRuntime` (so it sees the same managed
Codex home production sees), exposes `commandCwd` on
`createCommandManagedRuntimeClient` so callers can target a per-probe
directory without leaking the workspace `remoteCwd`, and propagates
plugin-worker termination as a usable error instead of a hang
> - The benefit is the Codex Test action mirrors production behavior
end-to-end, and probes against a terminated plugin worker fail fast
instead of timing out
## What Changed
- `packages/adapter-utils/src/command-managed-runtime.ts`: rename the
`remoteCwd` knob to `commandCwd` so callers can target a per-probe
directory without inheriting the workspace cwd; matching test coverage
in `command-managed-runtime.test.ts`
- `packages/adapter-utils/src/sandbox-callback-bridge.{ts,test.ts}`:
small fixes to keep callback bridge stop semantics deterministic
- `packages/adapters/codex-local/src/server/test.ts`: thread the Codex
hello probe through `prepareAdapterExecutionTargetRuntime` +
`prepareManagedCodexHome` so the probe sees the same managed home
production sees; new `test.remote.test.ts` covers the remote probe path
- `packages/adapters/cursor-local/src/server/execute.ts`: small
probe-side cleanup that aligns with the new commandCwd contract
- `server/src/services/plugin-worker-manager.ts`: surface plugin-worker
termination as a structured error so callers fail fast; new
`plugin-worker-terminated.cjs` fixture and
`plugin-worker-manager.test.ts` cases pin the behavior
## Verification
- `pnpm vitest run --no-coverage --project @paperclipai/adapter-utils
--project @paperclipai/adapter-codex-local --project
@paperclipai/adapter-cursor-local --project @paperclipai/server` —
1749/1750 passing (1 unrelated skip)
- `pnpm typecheck` clean
## Risks
Low–medium. The `remoteCwd → commandCwd` rename is a parameter renaming
on an internal helper used only by adapter test/execute paths in this
repo. The plugin-worker-terminated path was previously a hang; failing
fast may surface latent timeouts as explicit termination errors in
callers that already expected them.
## Model Used
Claude Opus 4.7 (1M context)
## Checklist
- [x] I have included a thinking path that traces from project context
to this change
- [x] I have specified the model used (with version and capability
details)
- [x] I have checked ROADMAP.md and confirmed this PR does not duplicate
planned core work
- [x] I have run tests locally and they pass
- [x] I have added or updated tests where applicable — new tests cover
commandCwd, plugin-worker termination, and Codex remote test path
- [x] If this change affects the UI, I have included before/after
screenshots — N/A (no UI)
- [x] I have updated relevant documentation to reflect my changes
- [x] I have considered and documented any risks above
- [x] I will address all Greptile and reviewer comments before
requesting merge
---
> **Stacked PR.** Sits on top of #5444 which adds the per-run runtime
API surface this PR builds on. Cumulative diff against `master` includes
that PR's content; the files touched by *this* PR's commit are listed
under "What Changed" above. Will rebase onto `master` and force-push
once #5444 merges.
## Thinking Path
> - Paperclip orchestrates AI agents for zero-human companies
> - When an agent runs against a remote target, Paperclip syncs the
workspace out to the remote at run start and restores changes back to
the local workspace at run end
> - The previous restore flow naïvely overwrote local files with
whatever the remote returned, so files that the remote run never touched
but had timestamp/mode drift could be needlessly rewritten — and a
single static `refs/paperclip/ssh-sync/imported` ref made concurrent SSH
workspace exports race on the same git ref
> - This pull request adds a `workspace-restore-merge` module that diffs
a pre-run snapshot against the post-run remote state and only writes
back files the remote actually changed; SSH workspace exports now use a
per-import unique ref so concurrent runs can't trample each other
> - Every adapter's execute path threads the snapshot through
`prepareAdapterExecutionTargetRuntime` so the merge has the baseline it
needs
> - The benefit is workspace restores no longer churn untouched files,
and concurrent SSH runs no longer collide on the import ref
## What Changed
- `packages/adapter-utils/src/workspace-restore-merge.{ts,test.ts}`: new
module — directory snapshot (kind/mode/sha256/symlink target) plus
snapshot-aware merge that writes only the files the remote changed
- `packages/adapter-utils/src/ssh.ts`: SSH workspace export uses a
per-import unique ref (`refs/paperclip/ssh-sync/imported/<uuid>`);
restore goes through the new merge helper; `ssh-fixture.test.ts` covers
the unique-ref + merge paths
- `packages/adapter-utils/src/sandbox-managed-runtime.ts` +
`remote-managed-runtime.ts`: thread the snapshot/merge through the
sandbox and SSH paths
- `packages/adapter-utils/src/server-utils.{ts,test.ts}` +
`execution-target.ts`: helpers for capturing the pre-run snapshot;
`prepareAdapterExecutionTargetRuntime` gains required `runId` and
optional `workspaceRemoteDir`, and returns the realized
`workspaceRemoteDir`
- Each adapter's `execute.ts` (acpx, claude, codex, cursor, gemini,
opencode, pi) takes the snapshot at run start and passes it through to
the runtime restore
- Remote execute test mocks updated to match the new
`prepareWorkspaceForSshExecution` return shape and the per-run
`${managedRemoteWorkspace}` cwd subdirectory
## Verification
- `pnpm vitest run --no-coverage --project @paperclipai/adapter-utils
--project @paperclipai/adapter-acpx-local --project
@paperclipai/adapter-claude-local --project
@paperclipai/adapter-codex-local --project
@paperclipai/adapter-cursor-local --project
@paperclipai/adapter-gemini-local --project
@paperclipai/adapter-opencode-local --project
@paperclipai/adapter-pi-local` — 196/196 passing
- `pnpm typecheck` clean across the workspace
## Risks
Medium. The restore path now writes a strict subset of what it
previously did — files the remote did not touch are no longer rewritten.
If any flow was relying on a touch-without-content-change being copied
back (timestamp or permission propagation only), that behavior is now
skipped. Snapshot capture adds an O(N-files-in-workspace) hash pass at
run start; the cost is bounded by the existing exclude list. The `runId`
parameter on `prepareAdapterExecutionTargetRuntime` is now required —
every in-tree caller is updated; out-of-tree adapter authors need to
pass it.
## Model Used
Claude Opus 4.7 (1M context)
## Checklist
- [x] I have included a thinking path that traces from project context
to this change
- [x] I have specified the model used (with version and capability
details)
- [x] I have checked ROADMAP.md and confirmed this PR does not duplicate
planned core work
- [x] I have run tests locally and they pass
- [x] I have added or updated tests where applicable — new module +
every adapter execute path covered
- [x] If this change affects the UI, I have included before/after
screenshots — N/A (no UI)
- [x] I have updated relevant documentation to reflect my changes
- [x] I have considered and documented any risks above
- [x] I will address all Greptile and reviewer comments before
requesting merge
## Thinking Path
> - Paperclip is a control plane for autonomous AI companies.
> - Issues are the core unit of work, and issue comments are how board
users and agents coordinate execution.
> - Some issue conversations need to produce plans and approvals instead
of immediate implementation work.
> - The existing issue contract did not distinguish standard execution
comments from planning-oriented issue work.
> - This pull request adds an issue work-mode contract and board UI
affordances for standard vs planning mode.
> - The benefit is that planning-mode issues can be created, displayed,
discussed, and carried through agent heartbeat context without losing
the normal issue workflow.
## What Changed
- Added `standard` / `planning` issue work-mode contracts across DB,
shared validators/types, server issue flows, plugin protocol, and
adapter heartbeat payloads.
- Added an idempotent `0081_optimal_dormammu` migration for
`issues.work_mode`, ordered after current `public-gh/master` migrations.
- Updated heartbeat/context summaries and issue-thread interaction
behavior so planning work mode is preserved when creating suggested
follow-up issues.
- Added UI support for planning-mode issue creation, issue rows, detail
composer styling, and composer work-mode toggles.
- Added focused server/shared/UI tests plus a Playwright visual
verification spec for planning-mode surfaces.
- Rebased the branch onto current `public-gh/master` and added durable
planning-mode screenshots under `doc/assets/pap-3368/`.
## Verification
- `pnpm --filter @paperclipai/db run check:migrations`
- `pnpm exec vitest run --project @paperclipai/shared
packages/shared/src/validators/issue.test.ts`
- `pnpm exec vitest run --project @paperclipai/server
server/src/__tests__/heartbeat-context-summary.test.ts
server/src/__tests__/issue-thread-interactions-service.test.ts
server/src/__tests__/issues-goal-context-routes.test.ts --pool=forks
--poolOptions.forks.isolate=true`
- `pnpm exec vitest run --project @paperclipai/ui
ui/src/components/IssueChatThread.test.tsx
ui/src/components/NewIssueDialog.test.tsx
ui/src/components/IssueRow.test.tsx ui/src/pages/IssueDetail.test.tsx`
- `pnpm exec vitest run --project @paperclipai/adapter-utils
packages/adapter-utils/src/server-utils.test.ts`
- `PAPERCLIP_E2E_SKIP_LLM=true npx playwright test --config
tests/e2e/playwright.config.ts
tests/e2e/planning-mode-visual-verification.spec.ts`
## Screenshots
Desktop planning detail:
Desktop planning row:
Desktop staged standard toggle:
Mobile planning detail:
Mobile planning row:
## Risks
- Medium migration risk: this adds a non-null issue column. The
migration uses `ADD COLUMN IF NOT EXISTS` so installations that applied
an older branch-local migration number can still apply the final
numbered migration safely.
- Medium contract risk: issue payloads, plugin payloads, and adapter
heartbeat payloads now include work mode; compatibility is handled by
defaulting missing values to `standard`.
- UI risk is moderate because composer controls changed; focused
component tests and visual e2e coverage exercise standard vs planning
display and toggle behavior.
> For core feature work, check [`ROADMAP.md`](ROADMAP.md) first and
discuss it in `#dev` before opening the PR. Feature PRs that overlap
with planned core work may need to be redirected — check the roadmap
first. See `CONTRIBUTING.md`.
## Model Used
- OpenAI Codex, GPT-5 coding agent in a local Paperclip worktree, with
shell/tool use. Exact context-window size is not exposed in this
runtime.
## Checklist
- [x] I have included a thinking path that traces from project context
to this change
- [x] I have specified the model used (with version and capability
details)
- [x] I have checked ROADMAP.md and confirmed this PR does not duplicate
planned core work
- [x] I have run tests locally and they pass
- [x] I have added or updated tests where applicable
- [x] If this change affects the UI, I have included before/after
screenshots
- [x] I have updated relevant documentation to reflect my changes
- [x] I have considered and documented any risks above
- [x] I will address all Greptile and reviewer comments before
requesting merge
---------
Co-authored-by: Paperclip <noreply@paperclip.ing>
## Thinking Path
> - Paperclip orchestrates AI agents across several adapter
implementations.
> - ACPX is a local adapter path that can proxy Claude and Codex-style
execution.
> - Its configuration needed stronger schema defaults, provider-aware
model handling, and better UI support.
> - Plugin authors also need clear docs for managed resources.
> - This pull request improves ACPX adapter configuration and documents
plugin-managed resources.
> - The benefit is a more predictable adapter setup path without
changing unrelated control-plane behavior.
## What Changed
- Improved ACPX config schema, execution config handling, UI build
config, and route coverage.
- Added ACPX model filtering support and tests.
- Updated the agent config form and storybook coverage for ACPX
model/provider behavior.
- Expanded plugin authoring documentation for managed resources.
## Verification
- `pnpm install --frozen-lockfile`
- `pnpm exec vitest run server/src/__tests__/acpx-local-execute.test.ts
server/src/__tests__/adapter-routes.test.ts
ui/src/lib/acpx-model-filter.test.ts`
## Risks
- Low-to-medium risk: adapter configuration behavior changes can affect
ACPX users, but the change is isolated to ACPX/plugin-doc surfaces and
covered by targeted adapter tests.
## Model Used
- OpenAI GPT-5 Codex via Paperclip `codex_local` adapter, with
shell/git/GitHub CLI tool use.
## Checklist
- [x] I have included a thinking path that traces from project context
to this change
- [x] I have specified the model used (with version and capability
details)
- [x] I have checked ROADMAP.md and confirmed this PR does not duplicate
planned core work
- [x] I have run tests locally and they pass
- [x] I have added or updated tests where applicable
- [x] If this change affects the UI, I have included before/after
screenshots
- [x] I have updated relevant documentation to reflect my changes
- [x] I have considered and documented any risks above
- [x] I will address all Greptile and reviewer comments before
requesting merge
---------
Co-authored-by: Paperclip <noreply@paperclip.ing>
## Thinking Path
> - Paperclip orchestrates AI agents for zero-human companies
> - Adapters spawn CLIs against local, SSH, and sandbox targets,
threading a runtime env through `runAdapterExecutionTargetProcess` and
the SSH/sandbox runners
> - Host identity vars (HOME, TMPDIR, XDG_*, NVM_DIR, PATH) routinely
leak into the env we send to remote targets — sometimes via test probes,
sometimes via runtime config — and break sandboxed/SSH'd CLIs whose own
profiles set those values correctly
> - The sanitization logic existed but lived alongside other helpers in
`server-utils.ts` and was applied piecemeal at adapter callsites, so it
was easy to bypass
> - This pull request lifts the sanitization into a standalone
`remote-execution-env.ts`, applies it at the SSH and sandbox runtime
boundary so every remote spawn goes through it, and removes the
duplicated callsite-level filtering
> - The benefit is identity-bound host env stops leaking across
SSH/sandbox transports regardless of which adapter calls in
## What Changed
- `packages/adapter-utils/src/remote-execution-env.ts`: new module —
single source of truth for which env keys are identity-bound and how to
strip them when the value matches the host's value
- `packages/adapter-utils/src/server-utils.ts`: remove the inline
sanitization (now in `remote-execution-env.ts`)
- `packages/adapter-utils/src/execution-target.ts`: apply sanitization
at the sandbox runtime boundary
- `packages/adapter-utils/src/ssh.ts`: apply sanitization at the SSH
spawn boundary
- `packages/adapters/opencode-local/src/server/test.ts`: drop
now-redundant callsite filtering
- `packages/adapters/pi-local/src/server/test.ts`: drop now-redundant
callsite filtering
- New tests `execution-target.test.ts` and
`execution-target-sandbox.test.ts` cover the sanitizer flow at both
transports, including positive cases (host-shaped path stripped) and
explicit-override preservation
## Verification
- `pnpm vitest run --no-coverage --project @paperclipai/adapter-utils
--project @paperclipai/adapter-opencode-local --project
@paperclipai/adapter-pi-local`
- `pnpm typecheck` clean
## Risks
Low–medium. The sanitization is now applied at one layer (boundary)
instead of N (callsites), so behavior is more consistent. Any adapter
that previously relied on a leaked host var landing on the remote shell
would now see it stripped — but those reliances were what this change
exists to fix.
## Model Used
Claude Opus 4.7 (1M context)
## Checklist
- [x] I have included a thinking path that traces from project context
to this change
- [x] I have specified the model used (with version and capability
details)
- [x] I have checked ROADMAP.md and confirmed this PR does not duplicate
planned core work
- [x] I have run tests locally and they pass
- [x] I have added or updated tests where applicable — new tests at both
transports
- [x] If this change affects the UI, I have included before/after
screenshots — N/A (no UI)
- [x] I have updated relevant documentation to reflect my changes
- [x] I have considered and documented any risks above
- [x] I will address all Greptile and reviewer comments before
requesting merge
## Thinking Path
> - Paperclip orchestrates AI agents for zero-human companies
> - The Gemini adapter's environment Test surfaces a hello probe so
operators can confirm the CLI runs end-to-end on the configured target
> - On SSH and E2B sandbox targets the round-trip cost (login-shell
sourcing, network, model warm-up) routinely exceeds the existing 10s
probe timeout, so the probe spuriously fails on environments that are
actually healthy
> - This pull request raises the gemini-local hello probe timeout to
60s, matching the timeout we use for slower-bootstrapping adapters
> - The benefit is the Gemini Test action no longer reports false
negatives on remote targets that need a longer first-run window
## What Changed
- `packages/adapters/gemini-local/src/server/test.ts`: hello probe
timeout raised from 10s to 60s
## Verification
- `pnpm vitest run --no-coverage --project
@paperclipai/adapter-gemini-local`
- Manual: SSH and E2B Gemini hello probes now complete cleanly without
spurious timeouts
## Risks
Low. A 60s ceiling on a non-blocking probe is consistent with sibling
adapters; the only behavior change is a longer worst-case wait when the
probe genuinely hangs.
## Model Used
Claude Opus 4.7 (1M context)
## Checklist
- [x] I have included a thinking path that traces from project context
to this change
- [x] I have specified the model used (with version and capability
details)
- [x] I have checked ROADMAP.md and confirmed this PR does not duplicate
planned core work
- [x] I have run tests locally and they pass
- [x] I have added or updated tests where applicable — N/A (one-line
timeout change)
- [x] If this change affects the UI, I have included before/after
screenshots — N/A (no UI)
- [x] I have updated relevant documentation to reflect my changes
- [x] I have considered and documented any risks above
- [x] I will address all Greptile and reviewer comments before
requesting merge
> **Stacked PR.** Sits on top of the e2b sandbox chain — #5278 (stdin
staging) and #5279 (honest-resolvability + login-profiles). The
cumulative diff against `master` includes both of those PRs' content;
the files touched by *this* PR's commit are the new
`maybeRunSandboxInstallCommand` helper in
`packages/adapter-utils/src/execution-target.ts` and the per-adapter
`index.ts`/`server/test.ts`/`server/execute.ts` wiring under
`packages/adapters/{claude,codex,cursor,gemini,opencode,pi}-local/`. The
honest resolvability check from #5279 is what gives this PR's install
command a meaningful "did it actually land on PATH" follow-up.
## Thinking Path
> - Paperclip orchestrates AI agents for zero-human companies
> - Sandbox execution targets are ephemeral — each fresh lease starts
from a template image that may or may not have the agent CLIs
preinstalled
> - When a CLI isn't preinstalled, the resolvability probe fails at
`command -v` and the hello probe never runs
> - There's no shared mechanism for "before you probe or provision,
install the CLI on this sandbox"
> - This pull request adds a `SANDBOX_INSTALL_COMMAND` constant per
adapter and a `maybeRunSandboxInstallCommand` helper that runs it via
the existing sandbox login shell, captures structured output, and never
throws (so the resolvability + hello probe still run after); each
adapter's `test()` and `execute()` share the constant so the two
callsites can't drift
> - The benefit is a fresh sandbox lease without a preinstalled CLI now
installs it once via `sh -lc` before the resolvability probe and before
managed-runtime provisioning, with a uniform
`<adapter>_install_command_run` check on the test report
## What Changed
- `packages/adapter-utils/src/execution-target.ts`: add
`AdapterSandboxInstallCommandCheck` and `maybeRunSandboxInstallCommand`
(runs the install via existing sandbox shell, captures
exit/stdout/stderr, returns a structured info/warn check, never throws)
- Add `SANDBOX_INSTALL_COMMAND` to each adapter's `index.ts` so `test()`
and `execute()` share a single source of truth
- Wire each of the 6 affected adapter `testEnvironment()`s to call
`maybeRunSandboxInstallCommand` before
`ensureAdapterExecutionTargetCommandResolvable`
- Pass `installCommand: SANDBOX_INSTALL_COMMAND` through
`prepareAdapterExecutionTargetRuntime` in each adapter's `execute()`
- Per-adapter install commands use npm globals where possible so
binaries land on a PATH segment the template already exports:
- claude → `npm install -g @anthropic-ai/claude-code`
- codex → `npm install -g @openai/codex`
- cursor → `curl https://cursor.com/install -fsS | bash`
- gemini → `npm install -g @google/gemini-cli`
- opencode → `npm install -g opencode-ai`
- pi → `npm install -g @mariozechner/pi-coding-agent`
SSH and local targets ignore `installCommand` (SSH runtime takes no such
param; local short-circuits before runtime prep), so this is a no-op for
non-sandbox environments.
## Verification
- `pnpm typecheck` clean
- `pnpm vitest run --no-coverage --project @paperclipai/adapter-utils`
and per-adapter projects pass
- Manual sandbox matrix (claude, codex, cursor, gemini, opencode, pi) —
each goes `install_command_run → resolvable → hello_probe_passed` (Codex
and Pi land on `hello_probe_auth_required`, which is the
configured-credentials problem, not an install issue)
- SSH no-regression: SSH Claude still passes; the helper short-circuits
on non-sandbox targets
## Risks
Medium — adds a network/CPU cost (npm install / curl) on every fresh
sandbox lease. Cost is bounded (one-time per lease, typically tens of
seconds for npm globals), and the helper never throws so a failing
install still lets the report run resolvability and hello probes. If a
sandbox image already has the CLI, the install is an idempotent
reinstall.
## Model Used
Claude Opus 4.7 (1M context)
## Checklist
- [x] I have included a thinking path that traces from project context
to this change
- [x] I have specified the model used (with version and capability
details)
- [x] I have checked ROADMAP.md and confirmed this PR does not duplicate
planned core work
- [x] I have run tests locally and they pass
- [x] I have added or updated tests where applicable
- [x] If this change affects the UI, I have included before/after
screenshots — N/A (no UI)
- [x] I have updated relevant documentation to reflect my changes
- [x] I have considered and documented any risks above
- [x] I will address all Greptile and reviewer comments before
requesting merge
## Thinking Path
> - Paperclip orchestrates AI agents for zero-human companies
> - The Codex adapter spawns the OpenAI Codex CLI to drive the model
> - Codex CLI 0.122 changed how it reads credentials: it ignores
`OPENAI_API_KEY` from the environment and reads only
`$CODEX_HOME/auth.json`
> - Without auth.json, Codex 0.122+ returns 401 "Missing bearer or basic
authentication" on `/v1/responses` even when `OPENAI_API_KEY` is
forwarded into the sandbox or remote shell
> - This pull request materializes an apikey-mode `auth.json` in the
managed Codex home (or per-run for the test probe) when an
`OPENAI_API_KEY` is configured
> - The benefit is configured Codex API keys authenticate correctly with
current Codex CLI versions across local, SSH, and sandbox targets
## What Changed
- `codex-home.ts`: add `writeApiKeyAuthJson()` and let
`prepareManagedCodexHome` accept an `apiKey` override that replaces the
symlinked host auth.json with an apikey-mode file
- `execute.ts`: pass `envConfig.OPENAI_API_KEY` into
`prepareManagedCodexHome` so the managed (and synced-to-remote) Codex
home authenticates via the configured key
- `test.ts`: when `OPENAI_API_KEY` is available, wrap the hello probe
with a small shell that materializes a per-run `$CODEX_HOME/auth.json`
before exec'ing codex; key content rides through env to avoid leaking
into process listings
- Update the `codex_hello_probe_auth_required` hint to explain Codex CLI
does not read `OPENAI_API_KEY` from env
## Verification
- `pnpm vitest run --no-coverage --project
@paperclipai/adapter-codex-local`
- `pnpm typecheck` clean
- Manual: Codex 0.122.0 with empty `CODEX_HOME` returns 401 with
env-only auth; with this change it authenticates cleanly
## Risks
Low risk — when no API key is configured, behavior is unchanged (no
auth.json written, existing chatgpt-mode flow preserved). Apikey-mode
`auth.json` is the upstream-supported format.
## Model Used
Claude Opus 4.7 (1M context)
## Checklist
- [x] I have included a thinking path that traces from project context
to this change
- [x] I have specified the model used (with version and capability
details)
- [x] I have checked ROADMAP.md and confirmed this PR does not duplicate
planned core work
- [x] I have run tests locally and they pass
- [x] I have added or updated tests where applicable
- [x] If this change affects the UI, I have included before/after
screenshots — N/A (no UI)
- [x] I have updated relevant documentation to reflect my changes
- [x] I have considered and documented any risks above
- [x] I will address all Greptile and reviewer comments before
requesting merge
## Thinking Path
> - Paperclip orchestrates AI agents for zero-human companies
> - The Pi adapter runs the pi-coding-agent CLI against local, SSH, and
sandbox execution targets
> - The Test path's hello probe spreads the host's `process.env` into
the remote process env, including the macOS PATH
> - The leaked Mac PATH overrides the nvm-sourced PATH set up by
`buildSshSpawnTarget`, so on a Linux SSH target `node` resolves to
system Node 18 instead of nvm's Node 20+
> - pi-coding-agent v0.68 / pi-tui then crashes at
`pi-tui/dist/utils.js:27` with `SyntaxError: Invalid regular expression
flags` on the `/v` unicode-sets regex (a Node 20+ feature)
> - This pull request stops the leak — same fix as the opencode SSH
probe — by passing only user-configured adapter env to the probe when
the target is remote
> - The benefit is the Pi hello probe now passes end-to-end against an
SSH target without the Node version mismatch
## What Changed
- `packages/adapters/pi-local/src/server/test.ts` passes only the
user-configured adapter env (`normalizeEnv(env)`) to
`runAdapterExecutionTargetProcess` when the target is remote
- Local probes still get the full `runtimeEnv` so headless permission
injection keeps working
## Verification
- `pnpm vitest run --no-coverage --project
@paperclipai/adapter-pi-local`
- `pnpm typecheck` clean
- Manual: Pi hello probe goes from `pi_hello_probe_failed` (Node 18
regex error) to `pi_hello_probe_passed` against an SSH target
## Risks
Low risk — same pattern shipped for opencode-local and consistent with
claude-local / codex-local / gemini-local.
## Model Used
Claude Opus 4.7 (1M context)
## Checklist
- [x] I have included a thinking path that traces from project context
to this change
- [x] I have specified the model used (with version and capability
details)
- [x] I have checked ROADMAP.md and confirmed this PR does not duplicate
planned core work
- [x] I have run tests locally and they pass
- [x] I have added or updated tests where applicable — pattern mirrors
sibling adapters
- [x] If this change affects the UI, I have included before/after
screenshots — N/A (no UI)
- [x] I have updated relevant documentation to reflect my changes
- [x] I have considered and documented any risks above
- [x] I will address all Greptile and reviewer comments before
requesting merge
## Thinking Path
> - Paperclip orchestrates AI agents for zero-human companies
> - The OpenCode adapter runs against local, SSH, and sandbox execution
targets
> - The Test path's hello probe spreads the Paperclip host's
`process.env` into the remote process env, which over SSH gets exported
on the remote shell
> - On a Linux SSH target, `HOME=/Users/...` and a host XDG_CONFIG_HOME
pointing at a macOS `/var/folders/...` temp dir cause OpenCode to walk a
host-only path and fail with `EACCES: permission denied, mkdir '/Users'`
> - This pull request stops the leak by passing only user-configured
adapter env to the probe when the target is remote, matching the pattern
already used by claude-local, codex-local, and gemini-local
> - The benefit is the OpenCode hello probe now passes end-to-end
against an SSH target without spurious filesystem errors
## What Changed
- `prepareOpenCodeRuntimeConfig` short-circuits when the target is
remote — the host-fs temp config dir is meaningless and harmful for a
remote target
- `test.ts` passes only the user-configured adapter env (no host
`process.env` spread) to `runAdapterExecutionTargetProcess` when
`targetIsRemote`
- Local probes still get the full `runtimeEnv` so headless permission
injection keeps working
## Verification
- `pnpm vitest run --no-coverage --project
@paperclipai/adapter-opencode-local`
- `pnpm typecheck` clean
- Manual: SSH OpenCode hello probe goes from `EACCES … mkdir '/Users'`
to `opencode_hello_probe_passed`
## Risks
Low risk — local probe behavior is unchanged; the change only narrows
the env passed to remote targets, matching the pattern already shipped
in sibling adapters.
## Model Used
Claude Opus 4.7 (1M context)
## Checklist
- [x] I have included a thinking path that traces from project context
to this change
- [x] I have specified the model used (with version and capability
details)
- [x] I have checked ROADMAP.md and confirmed this PR does not duplicate
planned core work
- [x] I have run tests locally and they pass
- [x] I have added or updated tests where applicable — pattern mirrors
existing sibling tests
- [x] If this change affects the UI, I have included before/after
screenshots — N/A (no UI)
- [x] I have updated relevant documentation to reflect my changes
- [x] I have considered and documented any risks above
- [x] I will address all Greptile and reviewer comments before
requesting merge
## Thinking Path
> - Paperclip orchestrates AI agents for zero-human companies
> - Each agent uses an adapter that drives a CLI (Claude, Gemini, Codex,
etc.)
> - The Gemini adapter parses a JSONL transcript stream the CLI emits to
learn what the model said
> - Gemini CLI v0.38 changed the transcript shape: assistant text now
comes through `type=message` with `role`/`content` and terminal status
comes through `type=status` / `type=stats`
> - The existing parser was written against the older `type=assistant` /
`type=result` shape, so post-v0.38 outputs left the parsed summary empty
and downgraded the SSH hello probe to "unexpected output"
> - This pull request updates every Gemini consumer (server parser, UI
parser, CLI formatter) to accept the v0.38 shape while keeping the
legacy shape working
> - The benefit is the Gemini adapter handles current upstream output
without losing backward compatibility, with explicit test coverage for
both shapes
## What Changed
- `packages/adapters/gemini-local/src/server/parse.ts` recognizes
`type=message` events with role/content and stops downgrading them
- `packages/adapters/gemini-local/src/ui/parse-stdout.ts` mirrors the
parser changes for the live UI transcript
- `packages/adapters/gemini-local/src/cli/format-event.ts` formats the
new event shape correctly for CLI output
- `parse.test.ts` and `parse-stdout.test.ts` add v0.38 coverage;
`gemini-local-adapter.test.ts` and `execute.remote.test.ts` switch
happy-path fixtures to the current real wire format and keep dedicated
tests for the older schema
## Verification
- `pnpm vitest run --no-coverage --project
@paperclipai/adapter-gemini-local` — full suite passes including new
v0.38 cases and preserved legacy cases
- `pnpm typecheck` clean
## Risks
Low risk — additive event handling. Legacy event shape path is preserved
with its own tests, so existing fixtures continue to parse identically.
## Model Used
Claude Opus 4.7 (1M context)
## Checklist
- [x] I have included a thinking path that traces from project context
to this change
- [x] I have specified the model used (with version and capability
details)
- [x] I have checked ROADMAP.md and confirmed this PR does not duplicate
planned core work
- [x] I have run tests locally and they pass
- [x] I have added or updated tests where applicable
- [x] If this change affects the UI, I have included before/after
screenshots — N/A (no UI)
- [x] I have updated relevant documentation to reflect my changes
- [x] I have considered and documented any risks above
- [x] I will address all Greptile and reviewer comments before
requesting merge
## Thinking Path
> - Paperclip orchestrates AI agents for zero-human companies, including
agents
> running the Gemini CLI (`gemini-local` adapter)
> - The Gemini CLI emits a JSONL event stream during a run that the
adapter
> parses to extract the assistant's response text, tool results, and
usage
> - Recent versions of the Gemini CLI emit assistant responses as
> `{ "type": "message", "role": "assistant", "content": ... }` events in
> addition to the previously-handled event shapes
> - The parser was not handling the new event type, so the assistant's
actual
> response text was being silently dropped from parsed output. Callers
ended
> up with empty assistant messages even when Gemini had successfully
> responded
> - This PR teaches the parser to recognize `{type: "message", role:
> "assistant"}` events and extract their content text via the same
> `collectMessageText` helper used for other message-shaped events
> - The benefit is that Gemini runs surface the assistant's real
response in
> downstream consumers (issue comments, run logs, downstream agent
context)
> instead of vanishing
## What Changed
- `packages/adapters/gemini-local/src/server/parse.ts`: in
`parseGeminiJsonl(...)`, add a branch for `event.type === "message"`
with
`role === "assistant"` that calls
`messages.push(...collectMessageText(event.content))`.
- `packages/adapters/gemini-local/src/server/parse.test.ts`: ~19 lines
of
coverage for the new branch.
## Verification
- `pnpm --filter @paperclipai/adapter-gemini-local test -- parse`
- Manual QA: run a Gemini agent on an issue, confirm the assistant's
response
appears as the issue comment / run output. Before this fix the comment
was
empty even when the run completed successfully.
## Risks
- Tightly scoped: 8 lines of production code in one parser branch. No
effect
on existing event shapes or other adapters.
- If the Gemini CLI changes its event schema again, this branch may need
to be
revisited — but adding it is strictly additive over current behaviour.
## Model Used
- OpenAI GPT-5.4 (reasoning effort: high) via Codex CLI
- Provider: OpenAI
- Used to author the code changes in this PR
## Checklist
- [x] I have included a thinking path that traces from project context
to this change
- [x] I have specified the model used (with version and capability
details)
- [x] I have checked ROADMAP.md and confirmed this PR does not duplicate
planned core work
- [x] I have run tests locally and they pass
- [x] I have added or updated tests where applicable
- [ ] If this change affects the UI, I have included before/after
screenshots — N/A
- [ ] I have updated relevant documentation to reflect my changes — N/A
- [x] I have considered and documented any risks above
- [x] I will address all Greptile and reviewer comments before
requesting merge
## Thinking Path
> - Paperclip orchestrates AI agents for zero-human companies, running
adapter
> commands like `claude`, `codex`, `pi` either locally or on remote
runtimes
> (SSH hosts, sandboxes, etc.)
> - On a fresh remote runtime — particularly an ephemeral sandbox — the
> adapter's CLI may not be installed yet. Today operators handle this
via
> external configuration (e.g. a project-level `provisionCommand` shell
> script) that has to know about every adapter the operator might want
to use
> - This means every adapter has its own well-known npm package, but
operators
> end up writing duplicate provision shell scripts that paste together
> `npm install -g @anthropic-ai/claude-code`, `npm install -g
@openai/codex`,
> etc. — knowledge the adapter itself already has
> - This PR moves that knowledge into the adapter modules: each adapter
declares
> how its runtime command should be detected and (if applicable)
installed
> via `getRuntimeCommandSpec(config)`. The execution path runs the
adapter's
> own install command on remote sandbox targets before launching, so a
fresh
> sandbox bootstraps itself instead of requiring a hand-written
provision script
> - The benefit is fewer footguns for operators provisioning remote
runtimes,
> and a clean place for new adapters to plug in their install recipe
## What Changed
- New types in `packages/adapter-utils/src/types.ts`:
- `AdapterRuntimeCommandSpec` describing `command`, optional
`detectCommand`, and optional `installCommand`
- Optional `getRuntimeCommandSpec(config)` on `ServerAdapterModule`
- Optional `runtimeCommandSpec` on `AdapterExecutionContext` so adapters
receive the resolved spec at execute time
- New helper `ensureAdapterExecutionTargetRuntimeCommandInstalled(...)`
in
`packages/adapter-utils/src/execution-target.ts` that runs the install
command
on remote targets when `transport === "sandbox"`. SSH and local targets
are
no-ops. Throws on timeout or non-zero exit so failures surface early.
- Each of `claude-local`, `codex-local`, `cursor-local`, `gemini-local`,
`opencode-local`, `pi-local`'s `execute.ts` now reads
`ctx.runtimeCommandSpec?.installCommand` and calls the helper before
launching
the adapter command.
- `server/src/adapters/registry.ts` declares `getRuntimeCommandSpec` for
each
adapter:
- claude/codex/gemini/opencode/pi-local: `npm install -g <package>`
recipe via
a shared `buildNpmRuntimeCommandSpec` helper, with a defensive guard
that
only auto-installs when the configured `command` matches the well-known
fallback (custom binaries are left alone).
- cursor-local: declares `command` only; no auto-install (no public npm
package), preserving the existing manual setup.
- `server/src/services/heartbeat.ts` resolves the spec via
`adapter.getRuntimeCommandSpec?.(runtimeConfig)` and passes it through
to
`AdapterExecutionContext`.
- Tests added in `execution-target.test.ts` (~75 lines), e2b
`plugin.test.ts` (~32 lines), and `environment-run-orchestrator.test.ts`
(~76 lines).
## Verification
- `pnpm --filter @paperclipai/adapter-utils test`
- `pnpm --filter @paperclipai/server test --
environment-run-orchestrator`
- `pnpm --filter @paperclipai/sandbox-providers-e2b test`
- Manual QA: run an adapter (claude/codex/etc.) against a fresh
sandbox-backed
environment that does NOT have the adapter CLI pre-installed. Confirm
the
install runs once at the start of the agent run and the adapter then
launches
successfully. Re-run on the same sandbox; confirm the install command is
idempotent and the second run starts faster.
- Confirm SSH and local execution paths are unaffected (gated by
`transport === "sandbox"`).
## Risks
- Behavioural shift on sandbox runs: a new install step now runs at the
start
of every sandbox agent run for adapters with `installCommand` set. The
install commands are idempotent (`if ! command -v X >/dev/null 2>&1;
then
npm install -g <pkg>; fi`), so this is fast on warm sandboxes. On a cold
sandbox, the first run takes longer.
- Operators who used the legacy project-level `provisionCommand` to
install
adapter CLIs can drop that part of their script; the adapter handles it
now.
Existing scripts continue to work — installs are idempotent.
- The cursor-local adapter has no auto-install (no public npm package).
Behaviour for cursor-local on sandboxes is unchanged.
- New optional surface on `ServerAdapterModule`. Plugins that don't
implement
`getRuntimeCommandSpec` retain previous behaviour (no auto-install).
## Model Used
- OpenAI GPT-5.4 (reasoning effort: high) via Codex CLI
- Provider: OpenAI
- Used to author the code changes in this PR
## Checklist
- [x] I have included a thinking path that traces from project context
to this change
- [x] I have specified the model used (with version and capability
details)
- [x] I have checked ROADMAP.md and confirmed this PR does not duplicate
planned core work
- [x] I have run tests locally and they pass
- [x] I have added or updated tests where applicable
- [ ] If this change affects the UI, I have included before/after
screenshots — N/A
- [ ] I have updated relevant documentation to reflect my changes — N/A
- [x] I have considered and documented any risks above
- [x] I will address all Greptile and reviewer comments before
requesting merge
> **Stacked PR (part 7 of 7).** Depends on:
- PR #5114
- PR #5115
- PR #5116
- PR #5117
- PR #5118
- PR #5119
> Diff against `master` includes commits from earlier PRs in the stack —
the new commit in this PR is the topmost one.
## Thinking Path
> - Paperclip orchestrates AI agents for zero-human companies
> - The Pi adapter persists a session jsonl per agent so subsequent runs
resume
> conversation context instead of starting cold
> - SSH testing reproduced a real failure: a verification issue reached
terminal
> `done` and the agent claimed success, but the proof artifact
> `manual-qa/environment-matrix/ssh/pi_local.md` was missing from the
realized
> SSH workspace on the QA target box
> - Root cause: the saved session header recorded a different cwd than
the new
> execution cwd, but the resume eligibility check only compared
session-params
> cwd via local-style `path.resolve` (which doesn't roundtrip on remote
POSIX
> paths). The stale session got resumed and writes landed in the wrong
cwd
> - This PR tightens resume eligibility for remote targets: it adds
remote-aware
> cwd normalisation, reads the first line of the session jsonl over SSH
(`head
> -n 1`) to verify the saved header cwd, and only resumes when both
> session-params cwd *and* the on-disk header cwd match the realised
execution
> cwd. Stale sessions are skipped silently and the run starts cold
> - The benefit is that Pi runs across cwd-changing environments stop
> accidentally resuming each other's sessions, and proof artifacts land
where
> reviewers expect them
## What Changed
- Added `normalizeExecutionCwd`, `executionCwdsMatch`,
`readSessionHeaderCwd`,
and `readSavedSessionCwd` helpers in `pi-local/src/server/execute.ts`
- `readSavedSessionCwd` reads the first line of the session jsonl —
locally via
`fs.readFile`, remotely via `runAdapterExecutionTargetShellCommand`
(`head -n 1`)
- Resume eligibility now requires:
1. Saved session id is non-empty
2. Execution target shape matches (existing check)
3. Session-params cwd matches the realised execution cwd
4. Session-header cwd (from the on-disk jsonl) matches the realised
execution cwd
- Stale sessions are skipped silently (run starts cold) instead of
resumed
- `execute.remote.test.ts` extended with: matching header → resume;
mismatched
header → start fresh; missing/unreadable header → start fresh; remote
head
command failure → start fresh
## Verification
- `pnpm --filter @paperclipai/adapter-pi-local test`
- `pnpm test -- pi-local`
- Manual QA: ran a Pi agent twice in two different remote cwds,
confirmed
the second run did not pick up the first run's session and that
subsequent
runs in the original cwd still resumed correctly
## Risks
- Adds a `head -n 1` shell call per Pi run on remote targets. Negligible
latency (single read of session jsonl), bounded by 15s timeout.
- If the `head` call fails for unrelated reasons (transient remote
unreachability), the run will start cold instead of resuming. This is
the
safe default but worth noting — operators may see one extra cold run if
a
remote glitches mid-session.
- No data is deleted or migrated; stale sessions remain on disk for
manual
inspection if desired.
## Model Used
- OpenAI GPT-5.4 (reasoning effort: high) via Codex CLI
- Provider: OpenAI
- Used to author the code changes in this PR
## Checklist
- [x] I have included a thinking path that traces from project context
to this change
- [x] I have specified the model used (with version and capability
details)
- [x] I have checked ROADMAP.md and confirmed this PR does not duplicate
planned core work
- [x] I have run tests locally and they pass
- [x] I have added or updated tests where applicable
- [ ] If this change affects the UI, I have included before/after
screenshots — N/A
- [ ] I have updated relevant documentation to reflect my changes — N/A
- [x] I have considered and documented any risks above
- [x] I will address all Greptile and reviewer comments before
requesting merge
> **Stacked PR (part 6 of 7).** Depends on:
- PR #5114
- PR #5115
- PR #5116
- PR #5117
- PR #5118
> Diff against `master` includes commits from earlier PRs in the stack —
the new commit in this PR is the topmost one.
## Thinking Path
> - Paperclip orchestrates AI agents for zero-human companies
> - The OpenCode adapter validates that its configured model exists
before letting
> a run start so misconfiguration fails fast with a clear error
> - SSH testing reproduced an OpenCode failure where issues stayed
`backlog`,
> timed out, and produced no comments. The root cause was in
> `packages/adapters/opencode-local/src/server/execute.ts`: the local
model
> guard `ensureOpenCodeModelConfiguredAndAvailable(...)` only ran when
execution
> was *not* remote, so SSH OpenCode bypassed it and failed silently
later
> - Subsequent testing surfaced a related remote-only failure where the
probe
> (when wired up naively) hits `EACCES: permission denied, mkdir
'/var/folders'`
> on the SSH box because of how OpenCode's runtime config picks a
tempdir
> - This PR runs the model probe on the actual execution target —
`opencode
> models` via `runAdapterExecutionTargetProcess` — instead of the local
CLI,
> parses the output with the shared `parseOpenCodeModelsOutput` helper,
and
> reports a concrete error naming the offending model and a sample of
available
> remote models when the configured model isn't present
> - The benefit is that mismatched OpenCode models surface as a clear
pre-flight
> error referencing the remote target instead of a silent run that never
leaves
> `backlog`
## What Changed
- Added `ensureRemoteOpenCodeModelConfiguredAndAvailable` in
`opencode-local/src/server/execute.ts` that runs `opencode models` via
`runAdapterExecutionTargetProcess` and validates the configured model is
in
the parsed output
- `models.ts` now exports `parseOpenCodeModelsOutput` and
`requireOpenCodeModelId`
so the remote path can reuse them
- `execute.ts` calls the remote variant when `executionTargetIsRemote`,
otherwise
the existing local `ensureOpenCodeModelConfiguredAndAvailable`
- Errors include the offending model id and a sample of available remote
models
so the operator knows exactly what's missing
- `execute.remote.test.ts` extended with cases for: probe timeout, probe
non-zero exit, empty model list, and missing-model error
## Verification
- `pnpm --filter @paperclipai/adapter-opencode-local test`
- `pnpm test -- opencode-local`
- Manual QA: configured an OpenCode agent with a model that exists
locally but
not in the remote sandbox, and confirmed the new error fires before the
run
starts and references the remote target
## Risks
- New behaviour: remote model validation adds a `~20s timeout` `opencode
models`
call on every remote run start. For most environments this is fast, but
a
network-slow sandbox could see startup latency rise. Timeout is bounded.
- If the remote CLI is missing or misconfigured, the new error replaces
the old
generic startup failure — clearer message, but the failure point shifts
earlier. Monitor for any QA flows that relied on the old failure shape.
## Model Used
- OpenAI GPT-5.4 (reasoning effort: high) via Codex CLI
- Provider: OpenAI
- Used to author the code changes in this PR
## Checklist
- [x] I have included a thinking path that traces from project context
to this change
- [x] I have specified the model used (with version and capability
details)
- [x] I have checked ROADMAP.md and confirmed this PR does not duplicate
planned core work
- [x] I have run tests locally and they pass
- [x] I have added or updated tests where applicable
- [ ] If this change affects the UI, I have included before/after
screenshots — N/A
- [ ] I have updated relevant documentation to reflect my changes — N/A
- [x] I have considered and documented any risks above
- [x] I will address all Greptile and reviewer comments before
requesting merge
> **Stacked PR (part 5 of 7).** Depends on:
- PR #5114
- PR #5115
- PR #5116
- PR #5117
> Diff against `master` includes commits from earlier PRs in the stack —
the new commit in this PR is the topmost one.
## Thinking Path
> - Paperclip orchestrates AI agents for zero-human companies
> - Agents run with a Paperclip-shaped environment
(`PAPERCLIP_WORKSPACE_CWD`,
> worktree path, `PAPERCLIP_WORKSPACES_JSON` hints) so the CLI can
locate the
> correct project tree
> - SSH testing reproduced a real failure: a Codex SSH run wrote to
> `/tmp/paperclip-env-matrix-...` (the *host* path) instead of the
realized
> remote workspace at `/home/<user>/paperclip-env-matrix-ssh-claude/...`
> because the adapter injected `PAPERCLIP_WORKSPACE_CWD=/tmp/...` into
the
> remote env
> - Code review on the initial codex-only fix asked to roll the same
approach
> into every other SSH-capable adapter (claude, acpx, cursor, opencode,
gemini,
> pi) via a shared helper rather than duplicating per-adapter
> - This PR adds `shapePaperclipWorkspaceEnvForExecution` in
adapter-utils that,
> when the execution target is remote: replaces local cwd with the
realized
> execution cwd, nulls out worktree path (which has no remote meaning),
and
> rewrites/strips `cwd` entries in workspace hints based on what was
actually
> synced. Every adapter calls it before invoking the remote runner
> - The benefit is that remote runs see the realized remote workspace,
host-local
> paths stop leaking into remote env, and the rule is unit-tested in one
place
## What Changed
- Added `shapePaperclipWorkspaceEnvForExecution` to
`packages/adapter-utils/src/server-utils.ts` with full unit coverage
(`server-utils.test.ts`)
- Each of acpx-local, claude-local, codex-local, cursor-local,
gemini-local,
opencode-local, pi-local now calls the new shaper before issuing the
remote
command and feeds the shaped values into `applyPaperclipWorkspaceEnv`
- Per-adapter `execute.remote.test.ts` files extended to cover the new
shaping
behaviour: localhost paths replaced with remote cwd, foreign-cwd hints
stripped, worktree path nulled out for remote targets
- `acpx-local/src/server/execute.test.ts` extended with shaping coverage
## Verification
- `pnpm test -- server-utils execute.remote`
- `pnpm --filter @paperclipai/adapter-acpx-local test`
- Manual QA reproducing the original failure:
1. Provision an E2B sandbox environment for the Paperclip QA company
2. Assign an issue to a remote-targeted claude-local agent and confirm
the
run starts in the correct remote cwd (no `/Users/...` path leakage in
the
run logs)
3. Repeat for opencode-local and pi-local
## Risks
- Behavioural shift: hints whose `cwd` doesn't match the workspace cwd
are now
stripped on remote targets. If any adapter relied on a leaked local hint
cwd,
it will see a missing `cwd` instead. Reviewed all current callers — none
do.
- Adds a small per-run cost (path resolve + string normalisation) on
every remote
execution. Negligible.
- Worktree path is now nulled out on remote (it has no meaning there).
Adapters
that previously read the value defensively will continue to work.
## Model Used
- OpenAI GPT-5.4 (reasoning effort: high) via Codex CLI
- Provider: OpenAI
- Used to author the code changes in this PR
## Checklist
- [x] I have included a thinking path that traces from project context
to this change
- [x] I have specified the model used (with version and capability
details)
- [x] I have checked ROADMAP.md and confirmed this PR does not duplicate
planned core work
- [x] I have run tests locally and they pass
- [x] I have added or updated tests where applicable
- [ ] If this change affects the UI, I have included before/after
screenshots — N/A
- [ ] I have updated relevant documentation to reflect my changes — N/A
- [x] I have considered and documented any risks above
- [x] I will address all Greptile and reviewer comments before
requesting merge
> **Stacked PR (part 4 of 7).** Depends on:
- PR #5114
- PR #5115
- PR #5116
> Diff against `master` includes commits from earlier PRs in the stack —
the new commit in this PR is the topmost one.
## Thinking Path
> - Paperclip orchestrates AI agents for zero-human companies
> - When creating an OpenCode-local agent, Paperclip currently validates
> `adapterConfig.model` against the *Paperclip host's* `opencode models`
output
> - SSH testing surfaced that this blocks creating an OpenCode agent for
an SSH
> environment: the model that exists on the SSH target isn't visible to
the
> host, so creation fails with "OpenCode requires `adapterConfig.model`
in
> provider/model format" even when the operator picked a real remote
model
> - The initial direction was environment-aware model discovery; the
final
> decision was to keep OpenCode on the same explicit-model pattern as
other
> adapters (default + curated list + manual override) and stop blocking
> creation on host-side discovery
> - This PR does both: the adapter-models endpoint now accepts
`environmentId` and
> probes against the target environment, and the create-time hard gate
is
> replaced by `requireOpenCodeModelId` which validates `provider/model`
*format*
> without requiring host-local discovery. Test/run-time still surfaces
real
> auth/availability problems
> - The benefit is that operators can create OpenCode agents for remote
> environments without out-of-band setup, and the model picker in the UI
> reflects the actually-targeted environment
## What Changed
- Added `requireOpenCodeModelId(input)` in
`opencode-local/src/server/models.ts`,
exported it from the adapter index
- `ensureOpenCodeModelConfiguredAndAvailable` now delegates the format
check to
`requireOpenCodeModelId`
- `agentsApi.adapterModels(companyId, adapterType, { environmentId })`
now accepts
an environment ID and passes it as a query parameter
- `queryKeys.agents.adapterModels` now keys on `(companyId, adapterType,
environmentId)`
- `server/src/routes/agents.ts` reads and validates the new query
parameter,
forwarding it to the adapter's model probe
- `AgentConfigForm.tsx` and `OnboardingWizard.tsx` build the model query
key from
the currently selected default environment ID and disable autodetect for
`opencode_local` (model selection is explicit)
- `NewAgent.tsx` simplified — no longer special-cases OpenCode
autodetect
- `company-portability.ts` no longer needs OpenCode-specific autodetect
handling
- Tests added/updated:
`adapter-model-refresh-routes.test.ts`, `adapter-models.test.ts`,
`agent-permissions-routes.test.ts`,
`opencode-local/src/server/models.test.ts`
## Verification
- `pnpm --filter @paperclipai/server test -- adapter-models
adapter-model-refresh agent-permissions`
- `pnpm --filter @paperclipai/adapter-opencode-local test`
- `pnpm --filter @paperclipai/ui test -- AgentConfigForm
OnboardingWizard NewAgent`
- Manual QA in browser:
1. Boot Paperclip on Tailscale-bound port (so it's reachable from
another
machine), create an OpenCode-local agent, switch the default environment
between two installed sandboxes, and confirm the model list refreshes
per-environment
2. Submit with a malformed `provider/model` string and verify the new
`requireOpenCodeModelId` error surfaces
- Before/after screenshots attached for `AgentConfigForm` model picker
## Risks
- Behavioural shift: switching default environment now triggers a model
refetch.
Should be cheap but introduces a new UI loading state for OpenCode
users.
- Removing dynamic autodetect for OpenCode: if any user configured an
agent
without specifying `model` and relied on autodetect populating it, that
agent
will now fail at submit time. Mitigation: validation error is explicit
and
actionable.
- New query string parameter on `/api/companies/:id/adapter-models` —
older
clients that omit it still work (parameter is optional and defaults to
null).
## Model Used
- OpenAI GPT-5.4 (reasoning effort: high) via Codex CLI
- Provider: OpenAI
- Used to author the code changes in this PR
## Checklist
- [x] I have included a thinking path that traces from project context
to this change
- [x] I have specified the model used (with version and capability
details)
- [x] I have checked ROADMAP.md and confirmed this PR does not duplicate
planned core work
- [x] I have run tests locally and they pass
- [x] I have added or updated tests where applicable
- [x] If this change affects the UI, I have included before/after
screenshots
- [ ] I have updated relevant documentation to reflect my changes — N/A
- [x] I have considered and documented any risks above
- [x] I will address all Greptile and reviewer comments before
requesting merge
> **Stacked PR (part 3 of 7).** Depends on:
- PR #5114
- PR #5115
> Diff against `master` includes commits from earlier PRs in the stack —
the new commit in this PR is the topmost one.
## Thinking Path
> - Paperclip orchestrates AI agents for zero-human companies
> - Agents executing on a remote SSH-backed environment need a way to
call back into
> the Paperclip control plane (run events, log streaming, signals)
> - When the SSH host can't reach the Paperclip host (NAT, firewalls, or
simply not
> on the same network), the run silently fails or hangs — a recurring
class of
> failure during SSH testing
> - In sandboxed environments we already solved this with a callback
bridge that
> tunnels back through the existing connection; SSH was the odd one out
> - This PR migrates SSH execution to use the same callback bridge, so
every
> adapter's remote run uses one consistent reverse-channel. Per-adapter
SSH glue
> is deleted in favour of a shared `CommandManagedRuntimeRunner` built
from the
> SSH spec
> - The benefit is fewer SSH-specific failure modes, a smaller code
surface, and
> one place to evolve the callback contract going forward
## What Changed
- Added `createSshCommandManagedRuntimeRunner` in
`packages/adapter-utils/src/ssh.ts` that adapts an SSH spec into a
generic
command-managed-runtime runner (with cwd, env, and timeout handling)
- Removed `paperclipApiUrl` from `SshRemoteExecutionSpec`; the bridge
URL now flows
through the shared runner
- Reworked `execution-target.ts` to use the SSH runner alongside sandbox
runners
via a unified `CommandManagedRuntimeRunner` interface
- Simplified `remote-managed-runtime.ts` and
`sandbox-managed-runtime.ts` to consume
the shared runner abstraction
- Deleted per-adapter SSH callback wiring from claude-local,
codex-local,
cursor-local, gemini-local, opencode-local, pi-local execute.ts files
- Removed `environment-runtime-driver-contract.test.ts` (the contract is
now
enforced by `environment-execution-target.test.ts`)
- Added/updated `execute.remote.test.ts` cases for each adapter to cover
the SSH
runner path
## Verification
- `pnpm --filter @paperclipai/adapter-utils test`
- `pnpm test -- execute.remote` (covers all six local adapters' SSH
paths)
- Manual QA: ran a claude-local agent against an SSH-backed environment,
confirmed
the agent successfully called back to `/api/agent-callback/*` endpoints
during
the run
## Risks
- Refactor touches all six local adapters. If any adapter had subtle
SSH-specific
behaviour that wasn't captured in tests, it could regress. Mitigation:
each
adapter's `execute.remote.test.ts` was extended.
- `paperclipApiUrl` removal from `SshRemoteExecutionSpec` is a breaking
type change
for any internal consumer. Verified no external plugins consume this
type.
- The new `CommandManagedRuntimeRunner` shape is a public surface in
`@paperclipai/adapter-utils`; downstream plugins implementing custom
runners may
need updates, but no such plugins exist in this repo.
## Model Used
- OpenAI GPT-5.4 (reasoning effort: high) via Codex CLI
- Provider: OpenAI
- Used to author the code changes in this PR
## Checklist
- [x] I have included a thinking path that traces from project context
to this change
- [x] I have specified the model used (with version and capability
details)
- [x] I have checked ROADMAP.md and confirmed this PR does not duplicate
planned core work
- [x] I have run tests locally and they pass
- [x] I have added or updated tests where applicable
- [ ] If this change affects the UI, I have included before/after
screenshots — N/A
- [ ] I have updated relevant documentation to reflect my changes — N/A
- [x] I have considered and documented any risks above
- [x] I will address all Greptile and reviewer comments before
requesting merge
## Thinking Path
> - Paperclip orchestrates AI agents for autonomous companies, and
heartbeat execution is the control-plane loop that keeps assigned work
moving.
> - Max-turn exhaustion is a recoverable local-adapter stop condition
for Claude and Gemini agents when a run needs another heartbeat to
continue safely.
> - The previous behavior could leave max-turn continuation details hard
to inspect, and duplicate/stale continuation wakes could keep running
after issue state changed.
> - The adapter layer also needed to avoid trusting arbitrary
stdout/stderr text as scheduler control metadata.
> - This pull request adds bounded max-turn continuation scheduling,
visible retry state, structured stop metadata handling, and
stale/duplicate continuation guards.
> - The benefit is safer automatic continuation after max-turn stops,
clearer operator visibility, and fewer duplicate or stale agent runs.
## What Changed
- Replaces closed PR #4952, whose head repository was deleted.
- Rebases the recovered max-turn continuation branch onto current
`paperclipai/paperclip:master`.
- Adds max-turn continuation scheduling and retry-state plumbing for
heartbeat runs.
- Adds stale/duplicate continuation suppression when issue status,
ownership, or execution locks change.
- Normalizes Claude/Gemini max-turn detection around structured stop
metadata instead of unstructured stdout/stderr text.
- Surfaces max-turn continuation settings and retry visibility in the
board UI.
- Adds focused server, adapter, and UI tests for max-turn stop metadata,
retry scheduling, stale queued-run invalidation, adapter
parsing/execution, run ledger display, and agent config patching.
## Verification
- `pnpm install --no-frozen-lockfile` to refresh local dependencies
after rebasing onto current `master`.
- `pnpm run preflight:workspace-links && pnpm exec vitest run
server/src/__tests__/claude-local-adapter.test.ts
server/src/__tests__/claude-local-execute.test.ts
server/src/__tests__/gemini-local-adapter.test.ts
server/src/__tests__/gemini-local-execute.test.ts
server/src/__tests__/heartbeat-retry-scheduling.test.ts
server/src/__tests__/heartbeat-stale-queue-invalidation.test.ts
server/src/services/heartbeat-stop-metadata.test.ts
ui/src/components/IssueRunLedger.test.tsx
ui/src/lib/agent-config-patch.test.ts ui/src/lib/runRetryState.test.ts
--testTimeout=20000`
- `pnpm --filter @paperclipai/adapter-claude-local typecheck && pnpm
--filter @paperclipai/adapter-gemini-local typecheck && pnpm --filter
@paperclipai/server typecheck && pnpm --filter @paperclipai/ui
typecheck`
- UI screenshot note: the UI changes are limited to config/ledger state
rendering rather than layout changes; component/unit coverage above
verifies the rendered behavior.
## Risks
- Medium behavior risk: heartbeat retry gating now suppresses max-turn
continuations when issue state or execution locks drift, so any callers
that relied on stale continuations running will now see cancellation
instead.
- Low adapter risk: Claude/Gemini unstructured text no longer triggers
max-turn scheduler metadata, so only structured stop signals and Gemini
exit code 53 are trusted.
- No database migrations.
> For core feature work, check [`ROADMAP.md`](ROADMAP.md) first and
discuss it in `#dev` before opening the PR. Feature PRs that overlap
with planned core work may need to be redirected — check the roadmap
first. See `CONTRIBUTING.md`.
## Model Used
- OpenAI Codex coding agent, GPT-5-class model, tool-enabled local
repository editing and command execution.
## Checklist
- [x] I have included a thinking path that traces from project context
to this change
- [x] I have specified the model used (with version and capability
details)
- [x] I have checked ROADMAP.md and confirmed this PR does not duplicate
planned core work
- [x] I have run tests locally and they pass
- [x] I have added or updated tests where applicable
- [x] If this change affects the UI, I have included before/after
screenshots (not applicable: state/default rendering only; covered by
component/unit tests)
- [x] I have updated relevant documentation to reflect my changes (not
applicable: no user-facing command or docs contract changed)
- [x] I have considered and documented any risks above
- [x] I will address all Greptile and reviewer comments before
requesting merge
---------
Co-authored-by: Paperclip <noreply@paperclip.ing>
## Thinking Path
> - Paperclip’s PR workflow is part of the control-plane safety surface
because it decides whether a branch is allowed to merge.
> - This issue started in that workflow: the lockfile and manifest
policy checks were diffing `base.sha..head.sha`, which incorrectly
treated unrelated `master` commits as if they belonged to the PR branch.
> - The right fix there is to diff from the PR merge base
(`base...head`) so policy checks only evaluate files introduced by the
branch itself.
> - Once that workflow fix was in place, `/checkpr` exposed a second
blocker on the PR merge ref: `verify` was failing in newer `master`-side
tests that were not part of the original branch diff.
> - The actionable repeated failure came from the ACPX local adapter
test suite, where a test hard-coded the managed Codex home under
`instances/default` even though the stable Vitest runner sets a
non-default `PAPERCLIP_INSTANCE_ID`.
> - This pull request now includes both the original CI diff-scope fix
and the targeted ACPX test fix so the PR’s actual checks align with
current base-branch execution.
> - The benefit is that the original false-positive lockfile failure is
removed, and the merge-ref verify path is hardened against the
instance-id isolation used in CI.
## What Changed
- Updated `.github/workflows/pr.yml` so the lockfile policy and manifest
policy steps diff `pull_request.base.sha...pull_request.head.sha` from
the merge base instead of using a two-dot base/head diff.
- Added an inline workflow comment explaining why the three-dot diff is
required for PR-scoped file detection.
- Updated `packages/adapters/acpx-local/src/server/execute.test.ts` so
the managed Codex home assertion uses a test-specific
`PAPERCLIP_INSTANCE_ID` instead of hard-coding `default`.
- Restored `PAPERCLIP_INSTANCE_ID` after that ACPX test finishes so the
test remains isolated and does not leak process env changes.
## Verification
- Reproduced the original false positive locally by comparing PR heads
`#4901` and `#4902` with the old `base..head` logic; both incorrectly
included `pnpm-lock.yaml` from unrelated `master` commits.
- Verified the new `base...head` logic reduces those PRs to only their
actual changed files and excludes `pnpm-lock.yaml`.
- Verified a real manifest-changing PR (`#4893`) still reports
`package.json` changes under the new logic.
- Ran `pnpm -r typecheck` successfully.
- Ran `pnpm vitest run
packages/adapters/acpx-local/src/server/execute.test.ts` successfully
after the ACPX test fix.
- Ran `pnpm vitest run packages/db/src/backup-lib.test.ts` successfully
against the merge-ref-related DB failure path observed during
`/checkpr`.
- Pushed commit `9520a976` and allowed PR `#4903` checks to rerun on the
updated branch.
## Risks
- Low risk: the workflow change only affects how PR policy checks
determine the changed file set.
- Low risk: the ACPX change is test-only and aligns the test with the
instance-isolation behavior already used by
`scripts/run-vitest-stable.mjs` in CI.
- The remaining operational risk is limited to other unrelated
merge-ref-only failures that were not reproduced in the targeted local
verification above.
> For core feature work, check [`ROADMAP.md`](ROADMAP.md) first and
discuss it in `#dev` before opening the PR. Feature PRs that overlap
with planned core work may need to be redirected — check the roadmap
first. See `CONTRIBUTING.md`.
## Model Used
- OpenAI Codex, `gpt-5-codex`, via the Codex local adapter in Paperclip.
- Tool-using coding model with shell execution, git, GitHub CLI, and
repository inspection in a local worktree.
- Context included the current repo, the Paperclip task thread, PR check
output, and the isolated execution workspace.
## Checklist
- [x] I have included a thinking path that traces from project context
to this change
- [x] I have specified the model used (with version and capability
details)
- [x] I have checked ROADMAP.md and confirmed this PR does not duplicate
planned core work
- [x] I have run tests locally and they pass
- [x] I have added or updated tests where applicable
- [ ] If this change affects the UI, I have included before/after
screenshots
- [ ] I have updated relevant documentation to reflect my changes
- [x] I have considered and documented any risks above
- [x] I will address all Greptile and reviewer comments before
requesting merge
## Thinking Path
> - Paperclip orchestrates AI-agent companies through a control plane
that can start, supervise, and recover agent runs.
> - Local adapters are the bridge between Paperclip issues and concrete
agent runtimes such as Claude, Codex, and other ACP-compatible tools.
> - The roadmap calls out broader “bring your own agent” and claw-style
agent support, and ACPX gives Paperclip one path to normalize multiple
ACP agents behind a single adapter.
> - The branch needed to become one reviewable PR against current
`paperclipai/paperclip:master`, without carrying stale base conflicts or
generated lockfile churn.
> - This pull request adds an experimental built-in `acpx_local`
adapter, integrates it through the server/CLI/UI adapter surfaces, and
adds regression coverage for runtime execution, skill sync, stream
parsing, diagnostics, and log redaction.
> - The benefit is that Paperclip can run Claude/Codex/custom ACP agents
through ACPX while keeping operator configuration, skills, logging, and
transcript rendering inside the existing adapter model.
## What Changed
- Added `@paperclipai/adapter-acpx-local` with server execution, config
schema, ACPX session handling, CLI formatting, UI config helpers, and
stdout parsing.
- Registered `acpx_local` across CLI, server, shared constants, UI
adapter metadata, adapter capabilities, and agent creation/editing
surfaces.
- Added ACPX runtime execution support with persistent sessions,
local-agent JWT environment handling, skill snapshots, runtime skill
materialization, and isolation/security regressions.
- Added ACPX adapter diagnostics and marked the adapter experimental in
the UI.
- Added command/env secret redaction for resolved command metadata in
adapter-utils, server event storage, and the Agent Detail invocation UI.
- Added Storybook coverage for ACPX config, transcript rendering, and
skill states, plus PR screenshots under `docs/pr-screenshots/pap-2944/`.
- Rebased the branch onto current `public-gh/master`; `pnpm-lock.yaml`
is intentionally not included and there are no migration/schema changes.
## Verification
- `pnpm exec vitest run
packages/adapters/acpx-local/src/server/execute.test.ts
packages/adapters/acpx-local/src/server/test.test.ts
packages/adapters/acpx-local/src/cli/format-event.test.ts
packages/adapters/acpx-local/src/ui/parse-stdout.test.ts
packages/adapter-utils/src/server-utils.test.ts
server/src/__tests__/redaction.test.ts
server/src/__tests__/acpx-local-execute.test.ts
server/src/__tests__/acpx-local-skill-sync.test.ts
server/src/__tests__/acpx-local-adapter-environment.test.ts
server/src/__tests__/adapter-routes.test.ts
server/src/__tests__/agent-skills-routes.test.ts
ui/src/adapters/metadata.test.ts` — 12 files, 87 tests passed.
- `pnpm --filter @paperclipai/adapter-acpx-local typecheck` — passed.
- `pnpm --filter @paperclipai/server typecheck` — passed.
- `pnpm --filter @paperclipai/ui typecheck` — passed.
- Confirmed PR diff does not include `pnpm-lock.yaml`, database schema
files, or migrations.
Screenshots:
## Risks
- Medium risk: this introduces a new built-in adapter package and
touches runtime execution, adapter registration, agent config, skills,
and transcript rendering.
- ACPX and ACP agent behavior can vary by installed tool versions; the
adapter is marked experimental to set operator expectations.
- `pnpm-lock.yaml` is excluded per repository PR policy, so dependency
lock refresh must be handled by the repo’s automation or maintainers.
- No database migration risk: no schema or migration files changed.
> For core feature work, check [`ROADMAP.md`](ROADMAP.md) first and
discuss it in `#dev` before opening the PR. Feature PRs that overlap
with planned core work may need to be redirected — check the roadmap
first. See `CONTRIBUTING.md`.
## Model Used
- OpenAI Codex coding agent based on GPT-5, with repository tool use,
shell execution, git operations, and local verification. Exact hosted
context window was not exposed in this environment.
## Checklist
- [x] I have included a thinking path that traces from project context
to this change
- [x] I have specified the model used (with version and capability
details)
- [x] I have checked ROADMAP.md and confirmed this PR does not duplicate
planned core work
- [x] I have run tests locally and they pass
- [x] I have added or updated tests where applicable
- [x] If this change affects the UI, I have included before/after
screenshots
- [x] I have updated relevant documentation to reflect my changes
- [x] I have considered and documented any risks above
- [x] I will address all Greptile and reviewer comments before
requesting merge
---------
Co-authored-by: Paperclip <noreply@paperclip.ing>
## Thinking Path
> - Paperclip is a control plane for autonomous AI companies, where
adapters are the boundary between the board, agents, and execution
runtimes.
> - Local adapters currently expose a primary runtime configuration, but
operators often need a cheaper model lane for routine or low-risk work.
> - That cheap lane has to stay adapter-owned: runtime profile settings
should not mutate the primary adapter config or bypass existing
auth/secret mediation.
> - Issue creation also needs an ergonomic way to request primary,
cheap, or custom model behavior for a selected assignee.
> - This pull request adds a first-class `cheap` model profile contract
across adapter capabilities, heartbeat config resolution, agent
configuration, and issue creation.
> - The benefit is cheaper task execution can be configured and
requested explicitly while preserving adapter boundaries, secret
handling, and audit visibility.
## What Changed
- Added adapter model-profile capability metadata and a `cheap` profile
contract for supported local adapters.
- Applied `runtimeConfig.modelProfiles.cheap.adapterConfig` during
heartbeat config resolution, including requested/applied/fallback run
metadata.
- Added agent configuration UI for cheap model profile settings without
writing those settings into primary `adapterConfig`.
- Added New Issue assignee model lane controls for Primary / Cheap /
Custom and request payload handling.
- Added run ledger profile badges and Storybook stories for the new
cheap-lane UI states.
- Added tests for validators, heartbeat model profile application,
permission/secret mediation, UI payload helpers, and run ledger
rendering.
- Added committed UI verification screenshots under
`docs/pr-screenshots/pap-2837/`.
- Addressed Greptile review feedback around cheap-profile defaults,
shared profile types, and fallback test data.
## Verification
Local:
- `pnpm exec vitest run packages/shared/src/validators/issue.test.ts
server/src/__tests__/adapter-registry.test.ts
server/src/__tests__/agent-permissions-routes.test.ts
server/src/__tests__/heartbeat-model-profile.test.ts
ui/src/components/IssueRunLedger.test.tsx
ui/src/lib/agent-config-patch.test.ts
ui/src/lib/issue-assignee-overrides.test.ts
ui/src/lib/new-agent-runtime-config.test.ts` — passed, 8 files / 103
tests.
- `pnpm exec vitest run ui/src/lib/new-agent-runtime-config.test.ts
ui/src/components/IssueRunLedger.test.tsx` — passed after
Greptile/rebase follow-up, 2 files / 17 tests.
- `pnpm --filter @paperclipai/ui typecheck` — passed after
Greptile/rebase follow-up.
- `pnpm -r typecheck` — passed.
- `pnpm build` — passed.
- `pnpm test:run` — did not complete successfully in this local
worktree: it stopped in pre-existing `@paperclipai/adapter-utils`
sandbox/SSH fixture suites outside this PR diff. Failures were 5s local
timeouts plus `git init -b` unsupported by this machine's Git 2.21.0.
The branch-specific targeted suites above passed.
- Branch was fetched/rebased onto `public-gh/master`; `git rev-list
--left-right --count public-gh/master...HEAD` reports `0 9`.
Remote PR checks on latest head
`e30bf399146451c86cee98ed528d51d33fa5af5a`:
- `policy` — passed.
- `verify` — passed.
- `e2e` — passed.
- `Greptile Review` — passed, confidence score 5/5; Greptile review
threads resolved.
- `security/snyk (cryppadotta)` — passed.
Screenshots:
- [New issue cheap lane
desktop](https://github.com/paperclipai/paperclip/blob/PAP-2837-plan-cheap-model-for-adapters-that-can-support-it/docs/pr-screenshots/pap-2837/newissue-cheap-desktop.png)
- [New issue custom lane
desktop](https://github.com/paperclipai/paperclip/blob/PAP-2837-plan-cheap-model-for-adapters-that-can-support-it/docs/pr-screenshots/pap-2837/newissue-custom-desktop.png)
- [New issue unsupported adapter
desktop](https://github.com/paperclipai/paperclip/blob/PAP-2837-plan-cheap-model-for-adapters-that-can-support-it/docs/pr-screenshots/pap-2837/newissue-unsupported-desktop.png)
- [Run ledger model profile badges
desktop](https://github.com/paperclipai/paperclip/blob/PAP-2837-plan-cheap-model-for-adapters-that-can-support-it/docs/pr-screenshots/pap-2837/runledger-profile-badges-desktop.png)
- Mobile variants are also in `docs/pr-screenshots/pap-2837/`.
## Risks
- Medium: heartbeat config mediation now merges runtime model profiles
into adapter configs, so adapter secret normalization and host-command
restrictions must keep covering nested config paths.
- Medium: the UI adds another issue creation choice; unsupported
adapters must keep hiding the cheap lane and preserve primary behavior.
- Low migration risk: no database migration is included.
> For core feature work, check [`ROADMAP.md`](ROADMAP.md) first and
discuss it in `#dev` before opening the PR. Feature PRs that overlap
with planned core work may need to be redirected — check the roadmap
first. See `CONTRIBUTING.md`.
## Model Used
OpenAI Codex coding agent using GPT-5-class reasoning with repo tool use
and command execution. Exact served model/context window was not exposed
by the runtime.
## Checklist
- [x] I have included a thinking path that traces from project context
to this change
- [x] I have specified the model used (with version and capability
details)
- [x] I have checked ROADMAP.md and confirmed this PR does not duplicate
planned core work
- [ ] I have run tests locally and they pass
- [x] I have added or updated tests where applicable
- [x] If this change affects the UI, I have included before/after
screenshots
- [x] I have updated relevant documentation to reflect my changes
- [x] I have considered and documented any risks above
- [x] I will address all Greptile and reviewer comments before
requesting merge
---------
Co-authored-by: Paperclip <noreply@paperclip.ing>
Co-authored-by: Claude Opus 4.7 <noreply@anthropic.com>
## Thinking Path
> - Paperclip's board UI and bundled skills are the operator layer for
configuring agents, routines, issue workflows, and local troubleshooting
loops.
> - The prior rollup mixed this operator polish with database backups,
backend reliability, thread scale, and cost/workflow primitives.
> - This pull request isolates the remaining board QoL, settings,
issue-detail integration, adapter config cleanup, and skills smoke
tooling.
> - It includes some integration-level overlap with the thread and
workflow slices so this branch can run from `origin/master` while still
preserving the full original work.
> - Preferred merge order is the narrower primitives first, then this
integration PR last.
> - The benefit is that reviewers can inspect the user-facing
board/settings/skills layer separately from backend infrastructure
changes.
## What Changed
- Added board/settings polish for agents, routines, company settings,
project workspace detail, and issue detail controls.
- Added agent/routine UI regression tests and New Issue dialog coverage.
- Integrated issue-detail activity/cost/interaction surfaces and leaf
work pause/resume controls.
- Cleaned bundled adapter UI config defaults and onboarding copy.
- Added terminal-bench loop and work-stoppage diagnosis skills plus a
smoke test script.
- Updated attachment type handling and Paperclip skill/API guidance.
## Verification
- `pnpm install --frozen-lockfile`
- `pnpm exec vitest run ui/src/pages/Agents.test.tsx
ui/src/pages/Routines.test.tsx ui/src/components/NewIssueDialog.test.tsx
ui/src/pages/IssueDetail.test.tsx
server/src/__tests__/costs-service.test.ts
server/src/__tests__/issue-thread-interaction-routes.test.ts
server/src/__tests__/issue-thread-interactions-service.test.ts`
- Result: 7 test files passed, 54 tests passed.
- `pnpm run smoke:terminal-bench-loop-skill`
- Result: JSON output included `"ok": true` and `"cleanup": true`.
- UI screenshots not included because verification is focused
component/page coverage for the changed board surfaces.
## Risks
- This is the integration-heavy PR in the split and intentionally
overlaps some component/API primitives with the issue-thread and
workflow PRs so it can run from `origin/master`.
- Preferred merge order: #4859, #4860, #4861, #4862, then this PR last.
If earlier branches merge first, this PR may need a straightforward
conflict refresh in shared UI files.
- The terminal-bench smoke script creates temporary mock issues and
relies on cleanup; the verified run returned `cleanup: true`.
> For core feature work, check [`ROADMAP.md`](ROADMAP.md) first and
discuss it in `#dev` before opening the PR. Feature PRs that overlap
with planned core work may need to be redirected — check the roadmap
first. See `CONTRIBUTING.md`.
## Model Used
- OpenAI Codex, GPT-5.5, code execution and GitHub CLI tool use, medium
reasoning effort.
## Checklist
- [x] I have included a thinking path that traces from project context
to this change
- [x] I have specified the model used (with version and capability
details)
- [x] I have checked ROADMAP.md and confirmed this PR does not duplicate
planned core work
- [x] I have run tests locally and they pass
- [x] I have added or updated tests where applicable
- [x] If this change affects the UI, I have included before/after
screenshots
- [x] I have updated relevant documentation to reflect my changes
- [x] I have considered and documented any risks above
- [x] I will address all Greptile and reviewer comments before
requesting merge
---------
Co-authored-by: Paperclip <noreply@paperclip.ing>
## Thinking Path
> - Paperclip orchestrates AI agents for zero-human companies
> - Agents can run inside sandboxed environments like E2B, which are
isolated from the host network
> - Sandboxed agents need to call back to the Paperclip API to report
progress, post comments, and update issue status
> - But sandbox environments cannot reach the Paperclip server directly
because they run in isolated network namespaces
> - This PR adds a callback bridge that proxies API requests from the
sandbox to the Paperclip server, running as a local HTTP server on the
host that forwards authenticated requests
> - The bridge is started automatically when an adapter launches a
sandbox execution, and torn down when the run completes
> - The benefit is sandboxed agents can interact with the Paperclip API
without requiring network-level access to the host, enabling E2B and
similar providers to work end-to-end
## What Changed
- Added `sandbox-callback-bridge.ts` in `packages/adapter-utils/` — a
lightweight HTTP bridge server that accepts requests from sandbox
environments and proxies them to the Paperclip API with authentication
- Added request validation and security policy: the bridge only forwards
requests to the configured API URL, validates content types, enforces
size limits, and rejects non-API paths
- Wired the bridge into all remote adapter execute paths (claude, codex,
cursor, gemini, pi) — the bridge starts before the agent process and the
bridge URL is passed via environment variables
- Updated `environment-execution-target.ts` to prefer the explicit API
URL from environment lease metadata for sandbox callback routing
- Fixed Claude sandbox runtime setup to work with the bridge
configuration
- Added comprehensive test coverage for bridge request handling, policy
enforcement, and sandbox execution integration
- Fixed browser bundling — the bridge module is excluded from the
frontend bundle via the adapter-utils index export
## Verification
- `pnpm test` — all existing and new tests pass, including bridge unit
tests and sandbox execution integration tests
- `pnpm typecheck` — clean
- Manual: configure an E2B environment, run an agent task, verify the
agent can post comments and update issue status through the bridge
## Risks
- Medium. This is a new network-facing component (HTTP server on
localhost). The security policy restricts forwarding to the configured
API URL only and validates all requests, but any proxy introduces attack
surface. The bridge binds to localhost only and is scoped to the
lifetime of a single agent run.
## Model Used
Codex GPT 5.4 high via Paperclip.
## Checklist
- [x] I have included a thinking path that traces from project context
to this change
- [x] I have specified the model used (with version and capability
details)
- [x] I have checked ROADMAP.md and confirmed this PR does not duplicate
planned core work
- [x] I have run tests locally and they pass
- [x] I have added or updated tests where applicable
- [ ] If this change affects the UI, I have included before/after
screenshots
- [x] I have updated relevant documentation to reflect my changes
- [x] I have considered and documented any risks above
- [x] I will address all Greptile and reviewer comments before
requesting merge
## Thinking Path
> - Paperclip orchestrates AI agents for zero-human companies
> - Agents can run inside sandboxed environments like E2B, or on remote
hosts via SSH
> - The cursor adapter needs to resolve `cursor-agent` inside sandbox
environments where it's installed in `~/.local/bin`
> - But when using the default `agent` command on a sandbox target, the
adapter didn't know to look in `~/.local/bin/cursor-agent`, causing
"command not found" failures
> - Additionally, repeated SSH runs failed because `git checkout` during
workspace sync conflicted with leftover `.paperclip-runtime` files from
previous runs
> - This PR adds sandbox-aware command resolution for cursor and fixes
the SSH workspace sync conflict
> - The benefit is cursor works in E2B sandboxes out of the box, and
repeated SSH runs don't fail on workspace sync
## What Changed
- `cursor-local`: Added `prepareCursorSandboxCommand` — on sandbox
targets, reads the remote `$HOME`, prepends `~/.local/bin` to PATH, and
prefers `~/.local/bin/cursor-agent` when the default command is
requested; tightened the sandbox command probe to validate the binary
exists before launching; preserves explicit custom command overrides
- `adapter-utils/ssh.ts`: Added `--force` to git checkout in SSH
workspace sync to handle `.paperclip-runtime` untracked file conflicts
from previous runs
## Verification
- `pnpm test` — all existing and new tests pass, including cursor
sandbox probe, sandbox execution, and custom command override tests
- `pnpm typecheck` — clean
- Manual: configure an E2B environment, run a cursor-local task, verify
it resolves cursor-agent from the sandbox install path
## Risks
- Low-medium. The `--force` flag on git checkout could discard
uncommitted changes in the remote workspace, but the workspace is
managed by Paperclip and should not contain user edits.
## Model Used
Codex GPT 5.4 high via Paperclip.
## Checklist
- [x] I have included a thinking path that traces from project context
to this change
- [x] I have specified the model used (with version and capability
details)
- [x] I have checked ROADMAP.md and confirmed this PR does not duplicate
planned core work
- [x] I have run tests locally and they pass
- [x] I have added or updated tests where applicable
- [ ] If this change affects the UI, I have included before/after
screenshots
- [x] I have updated relevant documentation to reflect my changes
- [x] I have considered and documented any risks above
- [x] I will address all Greptile and reviewer comments before
requesting merge
## Thinking Path
> - Paperclip orchestrates AI agents for zero-human companies
> - Agents run inside environments (local, SSH, E2B sandbox)
> - Operators need to configure and manage these environments
> - But environment settings were buried inside the general company
settings page, making them hard to find
> - Additionally, when testing an agent from the configuration form, the
test always ran locally regardless of which environment was selected
> - This PR moves environments into a dedicated top-level company
settings section and wires the "Test Environment" button to run inside
the selected environment
> - The benefit is operators can find and manage environments more
easily, and the test button now validates the actual environment the
agent will use
## What Changed
- Added a dedicated `CompanyEnvironments` settings page with its own
route and sidebar entry
- Updated `CompanySettingsSidebar` and `CompanySettingsNav` to include
the new environments section
- Modified the agent test route (`POST /agents/:id/test`) to accept an
optional `environmentId` parameter
- Updated all adapter `test.ts` handlers to resolve and use the
specified execution target environment
- Added `resolveTestExecutionTarget` to `execution-target.ts` for remote
environment test resolution with cwd fallback
- Moved the "Test Environment" button and its feedback display into the
`NewAgent` page footer for better UX flow
## Verification
- `pnpm test` — all existing and new tests pass
- `pnpm typecheck` — clean
- Manual: navigate to Company Settings, confirm "Environments" appears
as a top-level section
- Manual: configure an agent with a non-local environment, click "Test
Environment", confirm the test runs inside that environment
## Risks
- Low risk. UI-only routing change for the settings page. The
test-in-environment change adds an optional parameter with a local
fallback, so existing behavior is preserved when no environment is
specified.
## Model Used
Codex GPT 5.4 high via Paperclip.
## Checklist
- [x] I have included a thinking path that traces from project context
to this change
- [x] I have specified the model used (with version and capability
details)
- [x] I have checked ROADMAP.md and confirmed this PR does not duplicate
planned core work
- [x] I have run tests locally and they pass
- [x] I have added or updated tests where applicable
- [x] If this change affects the UI, I have included before/after
screenshots
- [x] I have updated relevant documentation to reflect my changes
- [x] I have considered and documented any risks above
- [x] I will address all Greptile and reviewer comments before
requesting merge
## Thinking Path
> - Paperclip orchestrates AI agents for zero-human companies.
> - The local adapter layer is responsible for turning Paperclip runtime
context into the environment seen by the child agent process.
> - The CEO onboarding bundle tells the agent where to read and write
its persistent memory and fact files.
> - That bundle was using `./memory/...` and `./life/...`, which only
works when the process cwd happens to equal the agent home directory.
> - At the same time, six local adapters each duplicated the same
workspace-env propagation logic, including `AGENT_HOME`, which makes
this contract easy to drift.
> - This pull request fixes the CEO instructions to use
`$AGENT_HOME/...` and centralizes workspace-env propagation in one
shared helper with shared tests.
> - The benefit is a real bug fix for agent memory paths plus a single
tested contract that makes future built-in adapter work less likely to
forget `AGENT_HOME`.
## What Changed
- Updated `server/src/onboarding-assets/ceo/HEARTBEAT.md` to use
`$AGENT_HOME/memory/...` and `$AGENT_HOME/life/...` instead of
cwd-relative `./memory/...` and `./life/...`.
- Added `applyPaperclipWorkspaceEnv(...)` in
`packages/adapter-utils/src/server-utils.ts` to centralize
`PAPERCLIP_WORKSPACE_*` and `AGENT_HOME` propagation.
- Added shared helper coverage in
`packages/adapter-utils/src/server-utils.test.ts` for both populated and
skip-empty cases.
- Switched the built-in local adapters (`claude_local`, `codex_local`,
`cursor_local`, `gemini_local`, `opencode_local`, `pi_local`) over to
the shared helper instead of inline env assignment blocks.
## Verification
- `pnpm install`
- `pnpm exec vitest run packages/adapter-utils/src/server-utils.test.ts
packages/adapters/claude-local/src/server/execute.remote.test.ts
packages/adapters/codex-local/src/server/execute.remote.test.ts
packages/adapters/cursor-local/src/server/execute.remote.test.ts
packages/adapters/gemini-local/src/server/execute.remote.test.ts
packages/adapters/opencode-local/src/server/execute.remote.test.ts
packages/adapters/pi-local/src/server/execute.remote.test.ts`
- Result: 7 test files passed, 31 tests passed, 0 failures.
## Risks
- Low risk.
- The only behavioral surface is the shared env propagation refactor
across six adapters; if the helper diverged from prior semantics, an
adapter could miss a workspace env var.
- The shared helper test plus the affected adapter execute tests reduce
that risk, and the helper preserves the prior "set only non-empty
strings" behavior.
## Model Used
- OpenAI Codex via Paperclip `codex_local` agent runtime; tool-assisted
coding workflow with shell execution, file patching, git operations, and
API interaction. The exact backend model identifier and context window
are not surfaced by this local runtime.
## Checklist
- [x] I have included a thinking path that traces from project context
to this change
- [x] I have specified the model used (with version and capability
details)
- [x] I have checked ROADMAP.md and confirmed this PR does not duplicate
planned core work
- [x] I have run tests locally and they pass
- [x] I have added or updated tests where applicable
- [x] If this change affects the UI, I have included before/after
screenshots
- [x] I have updated relevant documentation to reflect my changes
- [x] I have considered and documented any risks above
- [x] I will address all Greptile and reviewer comments before
requesting merge
## Thinking Path
> - Paperclip orchestrates AI agents for zero-human companies
> - A fast-moving control plane needs stable local tests and repeatable
local maintenance tools so contributors can safely split and review work
> - Several route suites needed stronger isolation, Codex manual model
selection needed a faster-mode option, and local browser cleanup missed
Playwright's headless shell binary
> - Storybook static output also needed to be preserved as a generated
review artifact from the working branch
> - This pull request groups the test/local-dev maintenance pieces so
they can be reviewed separately from product runtime changes
> - The benefit is more predictable contributor verification and cleaner
local maintenance without mixing these changes into feature PRs
## What Changed
- Added stable Vitest runner support and serialized route/authz test
isolation.
- Fixed workspace runtime authz route mocks and stabilized
Claude/company-import related assertions.
- Allowed Codex fast mode for manually selected models.
- Broadened the agent browser cleanup script to detect
`chrome-headless-shell` as well as Chrome for Testing.
- Preserved generated Storybook static output from the source branch.
## Verification
- `pnpm exec vitest run
src/__tests__/workspace-runtime-routes-authz.test.ts
src/__tests__/claude-local-execute.test.ts --config vitest.config.ts`
from `server/` passed: 2 files, 19 tests.
- `pnpm exec vitest run src/server/codex-args.test.ts --config
vitest.config.ts` from `packages/adapters/codex-local/` passed: 1 file,
3 tests.
- `bash -n scripts/kill-agent-browsers.sh &&
scripts/kill-agent-browsers.sh --dry` passed; dry-run detected
`chrome-headless-shell` processes without killing them.
- `test -f ui/storybook-static/index.html && test -f
ui/storybook-static/assets/forms-editors.stories-Dry7qwx2.js` passed.
- `git diff --check public-gh/master..pap-2228-test-local-maintenance --
. ':(exclude)ui/storybook-static'` passed.
- `pnpm exec vitest run
cli/src/__tests__/company-import-export-e2e.test.ts --config
cli/vitest.config.ts` did not complete in the isolated split worktree
because `paperclipai run` exited during build prep with `TS2688: Cannot
find type definition file for 'react'`; this appears to be caused by the
worktree dependency symlink setup, not the code under test.
- Confirmed this PR does not include `pnpm-lock.yaml`.
## Risks
- Medium risk: the stable Vitest runner changes how route/authz tests
are scheduled.
- Generated `ui/storybook-static` files are large and contain minified
third-party output; `git diff --check` reports whitespace inside those
generated assets, so reviewers may choose to drop or regenerate that
artifact before merge.
- No database migrations.
> For core feature work, check [`ROADMAP.md`](ROADMAP.md) first and
discuss it in `#dev` before opening the PR. Feature PRs that overlap
with planned core work may need to be redirected — check the roadmap
first. See `CONTRIBUTING.md`.
## Model Used
- OpenAI Codex coding agent based on GPT-5, with shell, git, Paperclip
API, and GitHub CLI tool use in the local Paperclip workspace.
## Checklist
- [x] I have included a thinking path that traces from project context
to this change
- [x] I have specified the model used (with version and capability
details)
- [x] I have checked ROADMAP.md and confirmed this PR does not duplicate
planned core work
- [x] I have run tests locally and they pass
- [x] I have added or updated tests where applicable
- [x] If this change affects the UI, I have included before/after
screenshots
- [x] I have updated relevant documentation to reflect my changes
- [x] I have considered and documented any risks above
- [x] I will address all Greptile and reviewer comments before
requesting merge
Note: screenshot checklist item is not applicable to source UI behavior;
the included Storybook static output is generated artifact preservation
from the source branch.
---------
Co-authored-by: Paperclip <noreply@paperclip.ing>
## Thinking Path
> - Paperclip orchestrates AI agents for zero-human companies
> - Adapter execution and retry classification decide whether agent work
pauses, retries, or recovers automatically
> - Transient provider failures need to be classified precisely so
Paperclip does not convert retryable upstream conditions into false hard
failures
> - At the same time, operators need an up-to-date model list for
Codex-backed agents and prompts should nudge agents toward targeted
verification instead of repo-wide sweeps
> - This pull request tightens transient recovery classification for
Claude and Codex, updates the agent prompt guidance, and adds Codex
model refresh support end-to-end
> - The benefit is better automatic retry behavior plus fresher
operator-facing model configuration
## What Changed
- added Codex usage-limit retry-window parsing and Claude extra-usage
transient classification
- normalized the heartbeat transient-recovery contract across adapter
executions and heartbeat scheduling
- documented that deferred comment wakes only reopen completed issues
for human/comment-reopen interactions, while system follow-ups leave
closed work closed
- updated adapter-utils prompt guidance to prefer targeted verification
- added Codex model refresh support in the server route, registry,
shared types, and agent config form
- added adapter/server tests covering the new parsing, retry scheduling,
and model-refresh behavior
## Verification
- `pnpm exec vitest run --project @paperclipai/adapter-utils
packages/adapter-utils/src/server-utils.test.ts`
- `pnpm exec vitest run --project @paperclipai/adapter-claude-local
packages/adapters/claude-local/src/server/parse.test.ts`
- `pnpm exec vitest run --project @paperclipai/adapter-codex-local
packages/adapters/codex-local/src/server/parse.test.ts`
- `pnpm exec vitest run --project @paperclipai/server
server/src/__tests__/adapter-model-refresh-routes.test.ts
server/src/__tests__/adapter-models.test.ts
server/src/__tests__/claude-local-execute.test.ts
server/src/__tests__/codex-local-execute.test.ts
server/src/__tests__/heartbeat-process-recovery.test.ts
server/src/__tests__/heartbeat-retry-scheduling.test.ts`
## Risks
- Moderate behavior risk: retry classification affects whether runs
auto-recover or block, so mistakes here could either suppress needed
retries or over-retry real failures
- Low workflow risk: deferred comment wake reopening is intentionally
scoped to human/comment-reopen interactions so system follow-ups do not
revive completed issues unexpectedly
> For core feature work, check [`ROADMAP.md`](ROADMAP.md) first and
discuss it in `#dev` before opening the PR. Feature PRs that overlap
with planned core work may need to be redirected — check the roadmap
first. See `CONTRIBUTING.md`.
## Model Used
- OpenAI Codex GPT-5-based coding agent with tool use and code execution
in the Codex CLI environment
## Checklist
- [x] I have included a thinking path that traces from project context
to this change
- [x] I have specified the model used (with version and capability
details)
- [x] I have checked ROADMAP.md and confirmed this PR does not duplicate
planned core work
- [x] I have run tests locally and they pass
- [x] I have added or updated tests where applicable
- [ ] If this change affects the UI, I have included before/after
screenshots
- [ ] I have updated relevant documentation to reflect my changes
- [x] I have considered and documented any risks above
- [x] I will address all Greptile and reviewer comments before
requesting merge
---------
Co-authored-by: Paperclip <noreply@paperclip.ing>
## Thinking Path
> - Paperclip orchestrates AI agents for zero-human companies
> - The environments subsystem already models execution environments,
but before this branch there was no end-to-end SSH-backed runtime path
for agents to actually run work against a remote box
> - That meant agents could be configured around environment concepts
without a reliable way to execute adapter sessions remotely, sync
workspace state, and preserve run context across supported adapters
> - We also need environment selection to participate in normal
Paperclip control-plane behavior: agent defaults, project/issue
selection, route validation, and environment probing
> - Because this capability is still experimental, the UI surface should
be easy to hide and easy to remove later without undoing the underlying
implementation
> - This pull request adds SSH environment execution support across the
runtime, adapters, routes, schema, and tests, then puts the visible
environment-management UI behind an experimental flag
> - The benefit is that we can validate real SSH-backed agent execution
now while keeping the user-facing controls safely gated until the
feature is ready to come out of experimentation
## What Changed
- Added SSH-backed execution target support in the shared adapter
runtime, including remote workspace preparation, skill/runtime asset
sync, remote session handling, and workspace restore behavior after
runs.
- Added SSH execution coverage for supported local adapters, plus remote
execution tests across Claude, Codex, Cursor, Gemini, OpenCode, and Pi.
- Added environment selection and environment-management backend support
needed for SSH execution, including route/service work, validation,
probing, and agent default environment persistence.
- Added CLI support for SSH environment lab verification and updated
related docs/tests.
- Added the `enableEnvironments` experimental flag and gated the
environment UI behind it on company settings, agent configuration, and
project configuration surfaces.
## Verification
- `pnpm exec vitest run
packages/adapters/claude-local/src/server/execute.remote.test.ts
packages/adapters/cursor-local/src/server/execute.remote.test.ts
packages/adapters/gemini-local/src/server/execute.remote.test.ts
packages/adapters/opencode-local/src/server/execute.remote.test.ts
packages/adapters/pi-local/src/server/execute.remote.test.ts`
- `pnpm exec vitest run server/src/__tests__/environment-routes.test.ts`
- `pnpm exec vitest run
server/src/__tests__/instance-settings-routes.test.ts`
- `pnpm exec vitest run ui/src/lib/new-agent-hire-payload.test.ts
ui/src/lib/new-agent-runtime-config.test.ts`
- `pnpm -r typecheck`
- `pnpm build`
- Manual verification on a branch-local dev server:
- enabled the experimental flag
- created an SSH environment
- created a Linux Claude agent using that environment
- confirmed a run executed on the Linux box and synced workspace changes
back
## Risks
- Medium: this touches runtime execution flow across multiple adapters,
so regressions would likely show up in remote session setup, workspace
sync, or environment selection precedence.
- The UI flag reduces exposure, but the underlying runtime and route
changes are still substantial and rely on migration correctness.
- The change set is broad across adapters, control-plane services,
migrations, and UI gating, so review should pay close attention to
environment-selection precedence and remote workspace lifecycle
behavior.
## Model Used
- OpenAI Codex via Paperclip's local Codex adapter, GPT-5-class coding
model with tool use and code execution in the local repo workspace. The
local adapter does not surface a more specific public model version
string in this branch workflow.
## Checklist
- [x] I have included a thinking path that traces from project context
to this change
- [x] I have specified the model used (with version and capability
details)
- [x] I have checked ROADMAP.md and confirmed this PR does not duplicate
planned core work
- [x] I have run tests locally and they pass
- [x] I have added or updated tests where applicable
- [ ] If this change affects the UI, I have included before/after
screenshots
- [x] I have updated relevant documentation to reflect my changes
- [x] I have considered and documented any risks above
- [x] I will address all Greptile and reviewer comments before
requesting merge
## Thinking Path
> - Paperclip orchestrates AI agents; each agent runs under an adapter
that spawns a model CLI as a child process.
> - The pi-local adapter (`packages/adapters/pi-local`) spawns `pi` and
inherits the child's shell environment — including `PATH`, which
determines what the child's bash tool can execute by name.
> - Paperclip skills ship executable helpers under `<skill>/bin/` (e.g.
`paperclip-get-issue`) and Reviewer/QA-style `AGENTS.md` files invoke
them by name via the agent's bash tool.
> - Pi-local builds its runtime env with `ensurePathInEnv({
...process.env, ...env })` only — it never adds the installed skills'
`bin/` dirs to PATH. The pi CLI's `--skill` arg loads each skill's
SKILL.md but does not augment PATH.
> - Consequence: every bash invocation of a skill helper fails with
`exit 127: command not found`. The agent then spends its heartbeat
guessing (re-reading SKILL.md, trying `find`, inventing command paths)
and either times out or gives up.
> - This PR prepends each injected skill's `bin/` directory to the child
PATH immediately before runtimeEnv is constructed.
> - The benefit: pi_local agents whose AGENTS.md uses any `paperclip-*`
skill helper can actually run those helpers.
## What Changed
- `packages/adapters/pi-local/src/server/execute.ts`: compute
`skillBinDirs` from the already-resolved `piSkillEntries`, dedupe
against the existing PATH, prepend them to whichever of `PATH` / `Path`
the merged env uses, then build `runtimeEnv`. No new helpers, no
adapter-utils changes.
## Verification
Manual repro before the fix:
1. Create a pi_local agent wired to a paperclip skill (e.g.
paperclip-control).
2. Wake the agent on an in_review issue with an AGENTS.md that starts
with `paperclip-get-issue "$PAPERCLIP_TASK_ID"`.
3. Session file: `{ "role": "toolResult", "isError": true, "content": [{
"text": "/bin/bash: paperclip-get-issue: command not found\n\nCommand
exited with code 127" }] }`.
After the fix: same wake; `paperclip-get-issue` resolves and returns the
issue JSON; agent proceeds.
Local commands:
```
pnpm --filter @paperclipai/adapter-pi-local typecheck # clean
pnpm --filter @paperclipai/adapter-pi-local build # clean
pnpm --filter @paperclipai/server exec vitest run \
src/__tests__/pi-local-execute.test.ts \
src/__tests__/pi-local-adapter-environment.test.ts \
src/__tests__/pi-local-skill-sync.test.ts
# 5/5 passing
```
No new tests: the existing `pi-local-skill-sync.test.ts` covers skill
symlink injection (upstream of the PATH step), and
`pi-local-execute.test.ts` covers the spawn path; this change only
augments env on the same spawn path.
## Risks
Low. Pure PATH augmentation on the child env. Edge cases:
- Zero skills installed → no PATH change (guarded by
`skillBinDirs.length > 0`).
- Duplicate bin dirs already on PATH → deduped; no pollution on re-runs.
- Windows `Path` casing → falls back correctly when merged env uses
`Path` instead of `PATH`.
- Skill dir without `bin/` subdir → joined path simply won't resolve;
harmless.
No behavioral change for pi_local agents that don't use skill-provided
commands.
## Model Used
- Claude, `claude-opus-4-7` (1M context), extended thinking enabled,
tool use enabled. Walked pi-local/cursor-local/claude-local and
adapter-utils to isolate the gap, wrote the inlined fix, and ran
typecheck/build/test locally.
## Checklist
- [x] Thinking path from project context to this change
- [x] Model used specified
- [x] Checked ROADMAP.md — no overlap
- [x] Tests run locally, passing
- [x] Tests added — new case in
`server/src/__tests__/pi-local-execute.test.ts`; verified it fails when
the fix is reverted
- [ ] UI screenshots — N/A (backend adapter change)
- [x] Docs updated — N/A (internal adapter, no user-facing docs)
- [x] Risks documented
- [x] Will address reviewer comments before merge
## Thinking Path
> - Paperclip orchestrates AI agents for zero-human companies.
> - Operators supervise that work through issues, comments, approvals,
and the board UI.
> - Some agent proposals need structured board/user decisions, not
hidden markdown conventions or heavyweight governed approvals.
> - Issue-thread interactions already provide a natural thread-native
surface for proposed tasks and questions.
> - This pull request extends that surface with request confirmations,
richer interaction cards, and agent/plugin/MCP helpers.
> - The benefit is that plan approvals and yes/no decisions become
explicit, auditable, and resumable without losing the single-issue
workflow.
## What Changed
- Added persisted issue-thread interactions for suggested tasks,
structured questions, and request confirmations.
- Added board UI cards for interaction review, selection, question
answers, and accept/reject confirmation flows.
- Added MCP and plugin SDK helpers for creating interaction cards from
agents/plugins.
- Updated agent wake instructions, onboarding assets, Paperclip skill
docs, and public docs to prefer structured confirmations for
issue-scoped decisions.
- Rebased the branch onto `public-gh/master` and renumbered branch
migrations to `0063` and `0064`; the idempotency migration uses `ADD
COLUMN IF NOT EXISTS` for old branch users.
## Verification
- `git diff --check public-gh/master..HEAD`
- `pnpm exec vitest run packages/adapter-utils/src/server-utils.test.ts
packages/mcp-server/src/tools.test.ts
packages/shared/src/issue-thread-interactions.test.ts
ui/src/lib/issue-thread-interactions.test.ts
ui/src/lib/issue-chat-messages.test.ts
ui/src/components/IssueThreadInteractionCard.test.tsx
ui/src/components/IssueChatThread.test.tsx
server/src/__tests__/issue-thread-interaction-routes.test.ts
server/src/__tests__/issue-thread-interactions-service.test.ts
server/src/services/issue-thread-interactions.test.ts` -> 9 files / 79
tests passed
- `pnpm -r typecheck` -> passed, including `packages/db` migration
numbering check
## Risks
- Medium: this adds a new issue-thread interaction model across
db/shared/server/ui/plugin surfaces.
- Migration risk is reduced by placing this branch after current master
migrations (`0063`, `0064`) and making the idempotency column add
idempotent for users who applied the old branch numbering.
- UI interaction behavior is covered by component tests, but this PR
does not include browser screenshots.
> For core feature work, check [`ROADMAP.md`](ROADMAP.md) first and
discuss it in `#dev` before opening the PR. Feature PRs that overlap
with planned core work may need to be redirected — check the roadmap
first. See `CONTRIBUTING.md`.
## Model Used
- OpenAI Codex, GPT-5-class coding agent runtime. Exact model ID and
context window are not exposed in this Paperclip run; tool use and local
shell/code execution were enabled.
## Checklist
- [x] I have included a thinking path that traces from project context
to this change
- [x] I have specified the model used (with version and capability
details)
- [x] I have checked ROADMAP.md and confirmed this PR does not duplicate
planned core work
- [x] I have run tests locally and they pass
- [x] I have added or updated tests where applicable
- [ ] If this change affects the UI, I have included before/after
screenshots
- [x] I have updated relevant documentation to reflect my changes
- [x] I have considered and documented any risks above
- [x] I will address all Greptile and reviewer comments before
requesting merge
---------
Co-authored-by: Paperclip <noreply@paperclip.ing>
## Thinking Path
> - Paperclip orchestrates AI agents through issue checkout, heartbeat
runs, routines, and auditable control-plane state
> - The runtime path has to recover from lost local processes, transient
adapter failures, blocked dependencies, and routine coalescing without
stranding work
> - The existing branch carried several reliability fixes across
heartbeat scheduling, issue runtime controls, routine dispatch, and
operator-facing run state
> - These changes belong together because they share backend contracts,
migrations, and runtime status semantics
> - This pull request groups the control-plane/runtime slice so it can
merge independently from board UI polish and adapter sandbox work
> - The benefit is safer heartbeat recovery, clearer runtime controls,
and more predictable recurring execution behavior
## What Changed
- Adds bounded heartbeat retry scheduling, scheduled retry state, and
Codex transient failure recovery handling.
- Tightens heartbeat process recovery, blocker wake behavior, issue
comment wake handling, routine dispatch coalescing, and
activity/dashboard bounds.
- Adds runtime-control MCP tools and Paperclip skill docs for issue
workspace runtime management.
- Adds migrations `0061_lively_thor_girl.sql` and
`0062_routine_run_dispatch_fingerprint.sql`.
- Surfaces retry state in run ledger/agent UI and keeps related shared
types synchronized.
## Verification
- `pnpm exec vitest run
server/src/__tests__/heartbeat-retry-scheduling.test.ts
server/src/__tests__/heartbeat-process-recovery.test.ts
server/src/__tests__/routines-service.test.ts`
- `pnpm exec vitest run src/tools.test.ts` from `packages/mcp-server`
## Risks
- Medium risk: this touches heartbeat recovery and routine dispatch,
which are central execution paths.
- Migration order matters if split branches land out of order: merge
this PR before branches that assume the new runtime/routine fields.
- Runtime retry behavior should be watched in CI and in local operator
smoke tests because it changes how transient failures are resumed.
> For core feature work, check [`ROADMAP.md`](ROADMAP.md) first and
discuss it in `#dev` before opening the PR. Feature PRs that overlap
with planned core work may need to be redirected — check the roadmap
first. See `CONTRIBUTING.md`.
## Model Used
- OpenAI Codex, GPT-5-based coding agent runtime, shell/git tool use
enabled. Exact hosted model build and context window are not exposed in
this Paperclip heartbeat environment.
## Checklist
- [x] I have included a thinking path that traces from project context
to this change
- [x] I have specified the model used (with version and capability
details)
- [x] I have checked ROADMAP.md and confirmed this PR does not duplicate
planned core work
- [x] I have run tests locally and they pass
- [x] I have added or updated tests where applicable
- [ ] If this change affects the UI, I have included before/after
screenshots
- [x] I have updated relevant documentation to reflect my changes
- [x] I have considered and documented any risks above
- [x] I will address all Greptile and reviewer comments before
requesting merge
## Thinking Path
> - Paperclip runs local adapter processes for agents and streams their
output into heartbeat runs
> - Some adapters can emit a terminal result before all descendant
processes have exited
> - If those descendants keep running, a heartbeat can appear complete
while the process group remains alive
> - Claude local runs need a bounded cleanup path after terminal JSON
output is observed and the child exits
> - This pull request adds terminal-result cleanup support to adapter
process utilities and wires it into the Claude local adapter
> - The benefit is fewer stranded adapter process groups after
successful terminal results
## What Changed
- Added terminal-result cleanup options to `runChildProcess`.
- Tracked child exit plus terminal output before signaling lingering
process groups.
- Added Claude local adapter configuration for terminal result cleanup
grace time.
- Added process cleanup tests covering terminal-output cleanup and noisy
non-terminal runs.
## Verification
- `pnpm install --frozen-lockfile --ignore-scripts`
- `pnpm exec vitest run packages/adapter-utils/src/server-utils.test.ts`
- Result: 9 tests passed.
## Risks
- Medium risk: this changes adapter child-process cleanup behavior.
- The cleanup only arms after terminal result detection and child exit,
and it is covered by process-group tests.
> For core feature work, check [`ROADMAP.md`](ROADMAP.md) first and
discuss it in `#dev` before opening the PR. Feature PRs that overlap
with planned core work may need to be redirected — check the roadmap
first. See `CONTRIBUTING.md`.
## Model Used
- OpenAI Codex coding agent based on GPT-5, tool-enabled local shell and
GitHub workflow, exact runtime context window not exposed in this
session.
## Checklist
- [x] I have included a thinking path that traces from project context
to this change
- [x] I have specified the model used (with version and capability
details)
- [x] I have checked ROADMAP.md and confirmed this PR does not duplicate
planned core work
- [x] I have run tests locally and they pass
- [x] I have added or updated tests where applicable
- [x] If this change affects the UI, I have included before/after
screenshots, or documented why it is not applicable
- [x] I have updated relevant documentation to reflect my changes
- [x] I have considered and documented any risks above
- [x] I will address all Greptile and reviewer comments before
requesting merge
---------
Co-authored-by: Paperclip <noreply@paperclip.ing>
## Thinking Path
> - Paperclip orchestrates AI agents for zero-human companies.
> - Heartbeat runs are the control-plane record of each agent execution
window.
> - Long-running local agents can exhaust context or stop while still
holding useful next-step state.
> - Operators need that stop reason, next action, and continuation path
to be durable and visible.
> - This pull request adds run liveness metadata, continuation
summaries, and UI surfaces for issue run ledgers.
> - The benefit is that interrupted or long-running work can resume with
clearer context instead of losing the agent's last useful handoff.
## What Changed
- Added heartbeat-run liveness fields, continuation attempt tracking,
and an idempotent `0058` migration.
- Added server services and tests for run liveness, continuation
summaries, stop metadata, and activity backfill.
- Wired local and HTTP adapters to surface continuation/liveness context
through shared adapter utilities.
- Added shared constants, validators, and heartbeat types for liveness
continuation state.
- Added issue-detail UI surfaces for continuation handoffs and the run
ledger, with component tests.
- Updated agent runtime docs, heartbeat protocol docs, prompt guidance,
onboarding assets, and skills instructions to explain continuation
behavior.
- Addressed Greptile feedback by scoping document evidence by run,
excluding system continuation-summary documents from liveness evidence,
importing shared liveness types, surfacing hidden ledger run counts,
documenting bounded retry behavior, and moving run-ledger liveness
backfill off the request path.
## Verification
- `pnpm exec vitest run packages/adapter-utils/src/server-utils.test.ts
server/src/__tests__/run-continuations.test.ts
server/src/__tests__/run-liveness.test.ts
server/src/__tests__/activity-service.test.ts
server/src/__tests__/documents-service.test.ts
server/src/__tests__/issue-continuation-summary.test.ts
server/src/services/heartbeat-stop-metadata.test.ts
ui/src/components/IssueRunLedger.test.tsx
ui/src/components/IssueContinuationHandoff.test.tsx
ui/src/components/IssueDocumentsSection.test.tsx`
- `pnpm --filter @paperclipai/db build`
- `pnpm exec vitest run server/src/__tests__/activity-service.test.ts
ui/src/components/IssueRunLedger.test.tsx`
- `pnpm --filter @paperclipai/ui typecheck`
- `pnpm --filter @paperclipai/server typecheck`
- `pnpm exec vitest run server/src/__tests__/activity-service.test.ts
server/src/__tests__/run-continuations.test.ts
ui/src/components/IssueRunLedger.test.tsx`
- `pnpm exec vitest run
server/src/__tests__/heartbeat-process-recovery.test.ts -t "treats a
plan document update"`
- `pnpm exec vitest run server/src/__tests__/activity-service.test.ts
server/src/__tests__/heartbeat-process-recovery.test.ts -t "activity
service|treats a plan document update"`
- Remote PR checks on head `e53b1a1d`: `verify`, `e2e`, `policy`, and
Snyk all passed.
- Confirmed `public-gh/master` is an ancestor of this branch after
fetching `public-gh master`.
- Confirmed `pnpm-lock.yaml` is not included in the branch diff.
- Confirmed migration `0058_wealthy_starbolt.sql` is ordered after
`0057` and uses `IF NOT EXISTS` guards for repeat application.
- Greptile inline review threads are resolved.
## Risks
- Medium risk: this touches heartbeat execution, liveness recovery,
activity rendering, issue routes, shared contracts, docs, and UI.
- Migration risk is mitigated by additive columns/indexes and idempotent
guards.
- Run-ledger liveness backfill is now asynchronous, so the first ledger
response can briefly show historical missing liveness until the
background backfill completes.
- UI screenshot coverage is not included in this packaging pass;
validation is currently through focused component tests.
> For core feature work, check [`ROADMAP.md`](ROADMAP.md) first and
discuss it in `#dev` before opening the PR. Feature PRs that overlap
with planned core work may need to be redirected — check the roadmap
first. See `CONTRIBUTING.md`.
## Model Used
- OpenAI Codex, GPT-5.4, local tool-use coding agent with terminal, git,
GitHub connector, GitHub CLI, and Paperclip API access.
## Checklist
- [x] I have included a thinking path that traces from project context
to this change
- [x] I have specified the model used (with version and capability
details)
- [x] I have checked ROADMAP.md and confirmed this PR does not duplicate
planned core work
- [x] I have run tests locally and they pass
- [x] I have added or updated tests where applicable
- [x] If this change affects the UI, I have included before/after
screenshots
- [x] I have updated relevant documentation to reflect my changes
- [x] I have considered and documented any risks above
- [x] I will address all Greptile and reviewer comments before
requesting merge
Screenshot note: no before/after screenshots were captured in this PR
packaging pass; the UI changes are covered by focused component tests
listed above.
---------
Co-authored-by: Paperclip <noreply@paperclip.ing>
## Thinking Path
> - Paperclip orchestrates AI agents for zero-human companies
> - Each adapter advertises a model list that powers the agent config UI
dropdown
> - The `claude_local` adapter's dropdown is sourced from the hard-coded
`models` array in `packages/adapters/claude-local/src/index.ts`
> - Anthropic recently released Opus 4.7, the newest current-generation
Opus model
> - Without a list entry, users cannot discover or select Opus 4.7 from
the dropdown (they can still type it manually, since the field is
creatable, but discoverability is poor)
> - This pull request adds `claude-opus-4-7` to the `claude_local` model
list so new agents can be configured with the latest model by default
> - The benefit is out-of-the-box access to the newest Opus model,
consistent with how every other current-generation Claude model is
already listed
## What Changed
- Added `{ id: "claude-opus-4-7", label: "Claude Opus 4.7" }` as the
**first** entry of the `models` array in
`packages/adapters/claude-local/src/index.ts`. Newest-first ordering
matches the convention already used for 4.6.
## Verification
- `pnpm --filter @paperclipai/adapter-claude-local typecheck` → passes.
- `pnpm --filter @paperclipai/server exec vitest run
src/__tests__/adapter-models.test.ts
src/__tests__/claude-local-adapter.test.ts` → 12/12 passing (both
directly-related files).
- No existing test pins the `claude_local` models array (see
`server/src/__tests__/adapter-models.test.ts`), so appending a new entry
is non-breaking.
- Manual check of UI consumer: `AgentConfigForm.tsx` fetches the list
via `agentsApi.adapterModels()` and renders it in a creatable popover —
no hard-coded expectations anywhere in the UI layer.
- Screenshots: single new option appears at the top of the Claude Code
(local) model dropdown; existing options unchanged.
## Risks
- Low risk. Purely additive: one new entry in a list consumed by a UI
dropdown. No behavior change for existing agents, no schema change, no
migration, no env var.
- `BEDROCK_MODELS` in
`packages/adapters/claude-local/src/server/models.ts` is intentionally
**not** touched — the exact region-qualified Bedrock id for Opus 4.7 is
not yet confirmed, and shipping a guessed id could produce a broken
option for Bedrock users. Tracked as a follow-up on the linked issue.
## Model Used
- None — human-authored.
## Checklist
- [x] I have included a thinking path that traces from project context
to this change
- [x] I have specified the model used (with version and capability
details)
- [x] I have run tests locally and they pass
- [x] I have added or updated tests where applicable (no tests needed:
existing suite already covers the list-consumer paths)
- [x] If this change affects the UI, I have included before/after
screenshots (dropdown gains one new top entry; all other entries
unchanged)
- [x] I have updated relevant documentation to reflect my changes (no
doc update needed: `docs/adapters/claude-local.md` uses
`claude-opus-4-6` only as an example, still valid)
- [x] I have considered and documented any risks above
- [x] I will address all Greptile and reviewer comments before
requesting merge
Closes#3827
## Thinking Path
> - Paperclip orchestrates AI agents for zero-human companies
> - The Costs > Providers tab displays live subscription quota from each
adapter (Claude, Codex)
> - The Claude adapter fetches utilization from the Anthropic OAuth
usage API and converts it to a 0-100 percent via `toPercent()`
> - The API changed to return utilization as 0-100 percentages (e.g.
`34.0` = 34%), but `toPercent()` assumed 0-1 fractions and multiplied by
100
> - After `Math.min(100, ...)` clamping, every quota window displayed as
100% used regardless of actual usage
> - Additionally, `extra_usage.used_credits` and `monthly_limit` are
returned in cents but were formatted as dollars, showing $6,793 instead
of $67.93
> - This PR applies the same `< 1` heuristic already proven in the Codex
adapter and fixes the cents-to-dollars conversion
> - The benefit is accurate quota display matching what users see on
claude.ai/settings/usage
## What Changed
- `toPercent()`: apply `< 1` heuristic to handle both legacy 0-1
fractions and current 0-100 percentage API responses (consistent with
Codex adapter's `normalizeCodexUsedPercent()`)
- `formatExtraUsageLabel()`: divide `used_credits` and `monthly_limit`
by 100 to convert cents to dollars before formatting
- Updated all `toPercent` and `fetchClaudeQuota` tests to use current
API format (0-100 range)
- Added backward-compatibility test for legacy 0-1 fraction values
- Added test for enabled extra usage with utilization and
cents-to-dollars conversion
## Verification
- `toPercent(34.0)` → `34` (was `100`)
- `toPercent(91.0)` → `91` (was `100`)
- `toPercent(0.5)` → `50` (legacy format still works)
- Extra usage `used_credits: 6793, monthly_limit: 14000` → `$67.93 /
$140.00` (was `$6,793.00 / $14,000.00`)
- Verified on a live instance with Claude Max subscription — Costs >
Providers tab now shows correct percentages matching
claude.ai/settings/usage
## Risks
Low risk. The `< 1` heuristic is already battle-tested in the Codex
adapter. The only edge case is a true utilization of exactly `1.0` which
maps to `1%` instead of `100%` — this is consistent with the Codex
adapter behavior and is an acceptable trade-off since 1% and 100% are
distinguishable in practice (100% would be returned as `100.0` by the
API).
## Model Used
Claude Opus 4.6 (1M context) via Claude Code CLI — tool use, code
analysis, and code generation
## Checklist
- [x] I have included a thinking path that traces from project context
to this change
- [x] I have specified the model used (with version and capability
details)
- [x] I have run tests locally and they pass
- [x] I have added or updated tests where applicable
- [x] If this change affects the UI, I have included before/after
screenshots
- [x] I have updated relevant documentation to reflect my changes
- [x] I have considered and documented any risks above
- [x] I will address all Greptile and reviewer comments before
requesting merge
Closes#2188🤖 Generated with [Claude Code](https://claude.com/claude-code)
---------
Co-authored-by: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
On resumed sessions, skipping --append-system-prompt-file (the original
fix) left two secondary issues:
- commandNotes still claimed the flag was injected, producing misleading
onMeta logs on every resumed heartbeat
- The instructions file was still read from disk and a combined temp file
written on every resume, even though effectiveInstructionsFilePath was
never consumed
Hoist canResumeSession before the I/O block and gate both the disk
operations and commandNotes construction on !canResumeSession / !sessionId.
Adds three regression tests: commandNotes is populated on fresh sessions,
empty on resume; and no agent-instructions.md is written on resume.
On resumed sessions the agent instructions are already present in the
session cache. Unconditionally passing --append-system-prompt-file
re-injects 5-10K redundant tokens per heartbeat and may be rejected by
the Claude CLI when combined with --resume.
Guard the flag behind `!resumeSessionId` so it is only appended on
fresh session starts.
Fixes: #2848
- Import models from index.ts instead of duplicating the array
- Use regex ^\w+\.anthropic\. to match all Bedrock region prefixes
(us, eu, ap, and any future regions)
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
Previously, --model was completely skipped for Bedrock users, so the
model dropdown selection was silently ignored and the CLI always used
its default model. Selecting Haiku would still run Opus.
- Add listClaudeModels() that returns Bedrock-native model IDs
(us.anthropic.*) when Bedrock env is detected
- Register listModels on claude_local adapter so the UI dropdown
shows Bedrock models instead of Anthropic API names
- Allow --model to pass through when the ID is a Bedrock-native
identifier (us.anthropic.* or ARN)
- Add isBedrockModelId() helper shared by execute.ts and test.ts
Follows up on #2793 which added basic Bedrock auth detection.
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
The Codex CLI can return "no rollout found for thread id ..." when
resuming a heartbeat thread whose rollout has been garbage-collected.
Extend isCodexUnknownSessionError() to match this wording so the
existing single-retry path in execute.ts activates correctly.
Add parse.test.ts covering the new pattern, existing stale-session
wordings, parseCodexJsonl, and a negative case.
Co-Authored-By: Paperclip <noreply@paperclip.ing>
## Thinking Path
Paperclip orchestrates AI agent runs and reports their success or
failure. The Pi adapter spawns a local Pi process and interprets its
JSONL output to determine the run outcome. When Pi hits a quota limit
(429 RESOURCE_EXHAUSTED), it retries internally and emits an
`auto_retry_end` event with `success: false` — but still exits with code
0. The current adapter trusts the exit code, so Paperclip marks the run
as succeeded even though it produced no useful work. This PR teaches the
parser to detect quota exhaustion and synthesize a failure.
Closes#2234
## Changes
- Parse `auto_retry_end` events with `success: false` into
`result.errors`
- Parse standalone `error` events into `result.errors`
- Synthesize exit code 1 when Pi exits 0 but parsed errors exist
- Use the parsed error as `errorMessage` so the failure reason is
visible in the UI
## Verification
```bash
pnpm vitest run pi-local-execute
pnpm vitest run --reporter=verbose 2>&1 | grep pi-local
```
- `parse.test.ts`: covers failed retry, successful retry (no error),
standalone error events, and empty error messages
- `pi-local-execute.test.ts`: end-to-end test with a fake Pi binary that
emits `auto_retry_end` + exits 0, asserts the run is marked failed
## Risks
- **Low**: Only affects runs where Pi exits 0 with a parsed error — no
change to normal successful or already-failing runs
- If Pi emits `auto_retry_end { success: false }` but the run actually
produced valid output, this would incorrectly mark it as failed. This
seems unlikely given the semantics of the event.
## Model Used
- Claude Opus 4.6 (Anthropic) — assisted with test additions and PR
template
## Checklist
- [x] Thinking path documented
- [x] Model specified
- [x] Tests pass locally
- [x] Test coverage for new parse branches (success path, error events,
empty messages)
- [x] No UI changes
- [x] Risk analysis included
---------
Co-authored-by: Dawid Piaskowski <dawid@MacBook-Pro.local>
Co-authored-by: Claude Opus 4.6 <noreply@anthropic.com>
Closes#2412
Related: #2681, #498, #128
## Thinking Path
> - Paperclip orchestrates AI agents for zero-human companies
> - The Claude Code adapter spawns the `claude` CLI to run agent tasks
> - The adapter detects auth mode by checking for `ANTHROPIC_API_KEY` —
recognizing only "api" and "subscription" modes
> - But users running Claude Code via **AWS Bedrock**
(`CLAUDE_CODE_USE_BEDROCK=1`) fall through to the "subscription" path
> - This causes a misleading "ANTHROPIC_API_KEY is not set;
subscription-based auth can be used" message in the environment check
> - Additionally, the hello probe passes `--model claude-opus-4-6` which
is **not a valid Bedrock model identifier**, causing `400 The provided
model identifier is invalid` and a probe failure
> - This pull request adds Bedrock auth detection, skips the
Anthropic-style `--model` flag for Bedrock, and returns the correct
billing type
> - The benefit is that Bedrock users get a working environment check
and correct cost tracking out of the box
---
## Pain Point
Many enterprise teams use **Claude Code through AWS Bedrock** rather
than Anthropic's direct API — for compliance, billing consolidation, or
VPC requirements. Currently, these users hit a **hard wall during
onboarding**:
| Problem | Impact |
|---|---|
| ❌ Adapter environment check **always fails** | Users cannot create
their first agent — blocked at step 1 |
| ❌ `--model claude-opus-4-6` is **invalid on Bedrock** (requires
`us.anthropic.*` format) | Hello probe exits with code 1: `400 The
provided model identifier is invalid` |
| ❌ Auth shown as _"subscription-based"_ | Misleading — Bedrock is
neither subscription nor API-key auth |
| ❌ Quota polling hits Anthropic OAuth endpoint | Fails silently for
Bedrock users who have no Anthropic subscription |
> **Bottom line**: Paperclip is completely unusable for Bedrock users
out of the box.
## Why Bedrock Matters
AWS Bedrock is a major deployment path for Claude in enterprise
environments:
- **Enterprise compliance** — data stays within the customer's AWS
account and VPC
- **Unified billing** — Claude usage appears on the existing AWS
invoice, no separate Anthropic billing
- **IAM integration** — access controlled through AWS IAM roles and
policies
- **Regional deployment** — models run in the customer's preferred AWS
region
Supporting Bedrock unlocks Paperclip for organizations that **cannot**
use Anthropic's direct API due to procurement, security, or regulatory
constraints.
---
## What Changed
- **`execute.ts`**: Added `isBedrockAuth()` helper that checks
`CLAUDE_CODE_USE_BEDROCK` and `ANTHROPIC_BEDROCK_BASE_URL` env vars.
`resolveClaudeBillingType()` now returns `"metered_api"` for Bedrock.
Biller set to `"aws_bedrock"`. Skips `--model` flag when Bedrock is
active (Anthropic-style model IDs are invalid on Bedrock; the CLI uses
its own configured model).
- **`test.ts`**: Environment check now detects Bedrock env vars (from
adapter config or server env) and shows `"AWS Bedrock auth detected.
Claude will use Bedrock for inference."` instead of the misleading
subscription message. Also skips `--model` in the hello probe for
Bedrock.
- **`quota.ts`**: Early return with `{ ok: true, windows: [] }` when
Bedrock is active — Bedrock usage is billed through AWS, not Anthropic's
subscription quota system.
- **`ui/src/lib/utils.ts`**: Added `"aws_bedrock"` → `"AWS Bedrock"` to
`providerDisplayName()` and `quotaSourceDisplayName()`.
## Verification
1. `pnpm -r typecheck` — all packages pass
2. Unit tests added and passing (6/6)
3. Environment check with Bedrock env vars:
| | Before | After |
|---|---|---|
| **Status** | 🔴 Failed | ✅ Passed |
| **Auth message** | `ANTHROPIC_API_KEY is not set; subscription-based
auth can be used if Claude is logged in.` | `AWS Bedrock auth detected.
Claude will use Bedrock for inference.` |
| **Hello probe** | `ERROR · Claude hello probe failed.` (exit code 1 —
`--model claude-opus-4-6` is invalid on Bedrock) | `INFO · Claude hello
probe succeeded.` |
| **Screenshot** | <img height="500" alt="Screenshot 2026-04-05 at 8 25
27 AM"
src="https://github.com/user-attachments/assets/476431f6-6139-425a-8abc-97875d653657"
/> | <img height="500" alt="Screenshot 2026-04-05 at 8 31 58 AM"
src="https://github.com/user-attachments/assets/d388ce87-c5e6-4574-b8d2-fd8b86135299"
/> |
4. Existing API key / subscription paths are completely untouched unless
Bedrock env vars are present
## Risks
- **Low risk.** All changes are additive — existing "api" and
"subscription" code paths are only entered when Bedrock env vars are
absent.
- When Bedrock is active, the `--model` flag is skipped, so the
Paperclip model dropdown selection is ignored in favor of the Claude
CLI's own model config. This is intentional since Bedrock requires
different model identifiers.
## Model Used
- Claude Opus 4.6 (`claude-opus-4-6`, 1M context window) via Claude Code
CLI
## Checklist
- [x] I have included a thinking path that traces from project context
to this change
- [x] I have specified the model used (with version and capability
details)
- [x] I have run tests locally and they pass
- [x] I have added or updated tests where applicable
- [x] If this change affects the UI, I have included before/after
screenshots
- [x] I have updated relevant documentation to reflect my changes
- [x] I have considered and documented any risks above
- [x] I will address all Greptile and reviewer comments before
requesting merge
---------
Co-authored-by: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
Co-authored-by: greptile-apps[bot] <165735046+greptile-apps[bot]@users.noreply.github.com>
- Plugin loader: install/reload/remove/reinstall external adapters
from npm packages or local directories
- Plugin store persisted at ~/.paperclip/adapter-plugins.json
- Self-healing UI parser resolution with version caching
- UI: Adapter Manager page, dynamic loader, display registry
with humanized names for unknown adapter types
- Dev watch: exclude adapter-plugins dir from tsx watcher
to prevent mid-request server restarts during reinstall
- All consumer fallbacks use getAdapterLabel() for consistent display
- AdapterTypeDropdown uses controlled open state for proper close behavior
- Remove hermes-local from built-in UI (externalized to plugin)
- Add docs for external adapters and UI parser contract
The openclaw_gateway adapter hardcodes the Paperclip API key path to
~/.openclaw/workspace/paperclip-claimed-api-key.json in buildWakeText().
In multi-agent OpenClaw deployments, each agent has its own workspace
with its own key file. The hardcoded path forces all agents to share
one key, breaking agent identity isolation.
Add a claimedApiKeyPath field to the adapter config (with UI input)
that allows operators to set a per-agent path. Falls back to the
current default when unset — zero behavior change for existing
deployments.
Fixes#930
After the mocked RPC spawn fails, getQuotaWindows() still calls
readCodexToken(). Use an empty mkdtemp directory for CODEX_HOME for the
duration of the test so we never read ~/.codex/auth.json or call WHAM.
Add a Vitest case that mocks `node:child_process.spawn` so the child
emits `error` (ENOENT) after the constructor attaches listeners.
`getQuotaWindows()` must resolve with `ok: false` instead of leaving an
unhandled `error` event on the process.
Register `packages/adapters/codex-local` in the root Vitest workspace.
Document in DEVELOPING.md that a missing `codex` binary should not take
down the API server during quota polling.
When the `codex` binary is absent from PATH, Node.js emits an `error`
event on the ChildProcess. Because `CodexRpcClient` only subscribed to
`exit` and `data` events, the `error` event was unhandled — causing
Node to throw it as an uncaught exception and crash the server.
Add an `error` handler in the constructor that rejects all pending RPC
requests and clears the queue. This makes a missing `codex` binary a
recoverable condition: `fetchCodexRpcQuota()` rejects, `getQuotaWindows()`
catches the error and returns `{ ok: false }`, and the server stays up.
The fix mirrors the existing pattern in `runChildProcess`
(packages/adapter-utils/src/server-utils.ts) which already handles
`ENOENT` the same way for the main task execution path.
When CURSOR_API_KEY is not set, check ~/.cursor/cli-config.json for
authInfo from `agent login` before emitting the missing key warning.
Users authenticated via native login no longer see a false warning.
The previous documentation parenthetical "(defaulting to ~/.codex/skills/)"
was misleading because Paperclip almost always sets CODEX_HOME to a
per-company managed home. Update index.ts docs, skills.ts detail string,
and execute.ts inline comment to make the runtime path unambiguous.
Co-Authored-By: Paperclip <noreply@paperclip.ing>
Setting the env var before the user-config loop meant adapter env
overrides could disable the guard. Move it after the loop so it
always wins, matching the pattern already used in test.ts and
models.ts.
Co-Authored-By: Paperclip <noreply@paperclip.ing>
Pi outputs the model list to stderr instead of stdout. This fix checks
stderr first and falls back to stdout for compatibility with older
versions.
Fixes model discovery returning empty arrays and environment tests
failing with 'Pi returned no models' error.
Set OPENCODE_DISABLE_PROJECT_CONFIG=true in all OpenCode invocations
(execute, model discovery, environment test) to stop the OpenCode CLI
from writing an opencode.json file into the project working directory.
Model selection is already passed via the --model CLI flag.
Co-Authored-By: Paperclip <noreply@paperclip.ing>
The default fallback in ensureCodexSkillsInjected still referenced the
old function name. Updated to use resolveCodexSkillsDir with shared
home as fallback.
Co-Authored-By: Paperclip <noreply@paperclip.ing>
The previous commit incorrectly used resolveSharedCodexHomeDir() (~/.codex)
but Codex runs with CODEX_HOME set to a per-company managed home under
~/.paperclip/instances/. Skills injected into ~/.codex/skills/ would not
be discoverable by Codex. Now uses effectiveCodexHome directly.
Co-Authored-By: Paperclip <noreply@paperclip.ing>
The Codex adapter was the only one injecting skills into
<cwd>/.agents/skills/, polluting the project's git repo. All other
adapters (Gemini, Cursor, etc.) use a home-based directory. This
changes the Codex adapter to inject into ~/.codex/skills/ (resolved
via resolveSharedCodexHomeDir) to match the established pattern.
Co-Authored-By: Paperclip <noreply@paperclip.ing>
Use path.join instead of string concatenation for the auth.json
fallback path in the detail message, ensuring correct path
separators on Windows.
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
When adapter config has no CODEX_HOME but process.env.CODEX_HOME is
set, readCodexAuthInfo reads from the process env path. The detail
message now uses codexHomeDir() instead of hardcoded "~/.codex" so
the displayed path always matches where credentials were read from.
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
Show the configured CODEX_HOME path instead of hardcoded ~/.codex
when the email fallback message is displayed.
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
Add tests for codex_native_auth_present and codex_openai_api_key_missing
code paths. Also pass adapter-configured CODEX_HOME through to
readCodexAuthInfo so the probe respects per-adapter home directories.
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
Loading an instructions file is normal, expected behavior — not worth
logging to stdout/stderr on every run. Warning logs for failed reads
are preserved.
Co-Authored-By: Paperclip <noreply@paperclip.ing>
The environment test warned about OPENAI_API_KEY being unset even
when Codex was authenticated via `codex auth`. Now checks
~/.codex/auth.json before emitting the warning.
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
mkdir -p the CODEX_HOME directory in codex-local adapter and the
agentHome directory in the heartbeat service before passing them to
adapters. This prevents CLI tools from erroring when their home
directory hasn't been created yet. Covers all local adapters that
set AGENT_HOME.
Co-Authored-By: Paperclip <noreply@paperclip.ing>
Drop "api" from the trigger kind dropdown and disable the "webhook"
option with a "COMING SOON" label until it's ready.
Co-Authored-By: Paperclip <noreply@paperclip.ing>
Persist child-process metadata for local adapter runs, keep detached runs alive when their pid still exists, queue a single automatic retry when the pid is confirmed dead, and clear detached warnings when the original run reports activity again.
Co-Authored-By: Paperclip <noreply@paperclip.ing>
Pi sometimes sends tool results as a direct array [{"type":"text","text":"..."}]
rather than wrapped in {"content": [...]}. Now handles both formats to
properly extract text content instead of showing raw JSON.
Pi returns tool results in format: {"content": [{"type": "text", "text": "..."}]}
Previously we were JSON.stringify-ing the whole object, showing as:
{"content":[{"type":"text","text":"..."}]}
Now we extract the actual text content for cleaner display.
The turn_end event includes toolResults array with toolName. Previously
the parsing only included toolCallId, now we also extract toolName so
the UI can display the correct tool name even when tool_result entries
arrive without a preceding tool_call.
When tool_result entries arrive without a matching tool_call, the transcript
was showing generic 'tool' as the name. Now pl-local parses toolName from
tool_execution_end events and passes it through, so the UI can display
the actual tool name (e.g., 'bash', 'Read', 'Ls') instead of 'tool'.
- Hoist PI_AGENT_SKILLS_DIR to module-level constant to avoid duplicate path computation
- Always create ~/.pi/agent/skills directory before early return in ensurePiSkillsInjected, so the path is valid when --skill flag is passed
- Add --skill ~/.pi/agent/skills to pi CLI invocation so it loads the paperclip skill
- Enable pi_local in the AgentConfigForm adapter type dropdown (was showing as coming soon)
Expose adapter-discovered user-installed skills with provenance metadata, share persistent skill snapshot classification across local adapters, and render unmanaged skills as a read-only section in the agent skills UI.
Co-Authored-By: Paperclip <noreply@paperclip.ing>
Address Greptile review feedback:
1. Wrap os.userInfo() in try/catch — it throws SystemError when the
current UID has no /etc/passwd entry (e.g. `docker run --user 1234`
with a minimal image). Falls back to process.env.HOME gracefully.
2. Add HOME to VOLATILE_ENV_KEY_EXACT so the discovery cache key is
not affected by the caller-supplied HOME vs the resolved HOME.
os.userInfo().homedir is constant for the process lifetime, so
HOME adds no useful cache differentiation.
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
When Paperclip's server is started via `runuser -u node` (common in
Docker/Fly.io deployments), the HOME environment variable retains the
parent process's value (e.g. /root) instead of the target user's home
directory (/home/node). This causes `opencode models` to miss provider
auth credentials stored under the actual user's home, resulting in
"Configured OpenCode model is unavailable" errors for providers that
require API keys (e.g. zai/zhipuai).
Fix: use `os.userInfo().homedir` (reads from /etc/passwd, not env) to
ensure the child process always sees the correct HOME, regardless of
how the server was launched.
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
- Extract all Anthropic credential/API logic into claude-local/src/server/quota.ts
- Extract all OpenAI/WHAM credential/API logic into codex-local/src/server/quota.ts
- Add optional getQuotaWindows() to ServerAdapterModule in adapter-utils
- Rewrite quota-windows.ts as a 29-line thin aggregator with zero provider knowledge
- Wire getQuotaWindows into adapter registry for claude-local and codex-local
- Add 47 unit tests covering toPercent, secondsToWindowLabel, WHAM normalization,
readClaudeToken, readCodexToken, fetchClaudeQuota, fetchCodexQuota, fetchWithTimeout
- Add 8 unit tests covering parseDateRange validation and byProvider pro-rata math
Adding a third provider now requires only touching that provider's adapter.
Matches the pattern in codex-local and cursor-local adapters,
giving operators consistent feedback about whether instructions
were actually loaded.
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
The $AGENT_HOME environment variable was referenced by skills (e.g.
para-memory-files) but never actually set, causing runtime errors like
"/HEARTBEAT.md: No such file or directory" when agents tried to resolve
paths relative to their home directory.
Add agentHome to the paperclipWorkspace context in the heartbeat service
and propagate it as the AGENT_HOME env var in all local adapters.
Co-Authored-By: Paperclip <noreply@paperclip.ing>
- reapOrphanedRuns() now only scans running runs; queued runs are
legitimately absent from runningProcesses (waiting on concurrency
limits or issue locks) so including them caused false process_lost
failures (closes#90)
- Add module-level activeRunExecutions set so non-child-process adapters
(http, openclaw) are protected from the reaper during execution
- Add resumeQueuedRuns() to restart persisted queued runs after a server
restart, called at startup and each periodic tick
- Add outer catch in executeRun() so setup failures (ensureRuntimeState,
resolveWorkspaceForRun, etc.) are recorded as failed runs instead of
leaving them stuck in running state
- Guard resumeQueuedRuns() against paused/terminated/pending_approval agents
- Increase opencode models discovery timeout from 20s to 45s
* public-gh/master:
Default Gemini adapter to yolo mode and add API access prompt note
fix: remove Cmd+1..9 company-switch shortcut
fix(ui): prevent IME composition Enter from moving focus in new issue title
fix(cli): add restart hint after allowed-hostname change
docs: remove obsolete TODO for CONTRIBUTING.md
fix: default dangerouslySkipPermissions to true for unattended agents
fix: route heartbeat cost recording through costService
Show issue creator in properties sidebar
Gemini CLI only registers run_shell_command in --approval-mode yolo.
Non-yolo modes don't expose it at all, making Paperclip API calls
impossible. Always pass --approval-mode yolo and remove the now-unused
policy engine code, approval mode config, and UI toggles.
Add a "Paperclip API access note" to the prompt with curl examples
via run_shell_command, since the universal SKILL.md is tool-agnostic.
Also extract structured question events from Gemini assistant messages
to support interactive approval flows.
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
The OpenClaw Gateway's agent method has strict parameter validation
that rejects unknown properties. The paperclip property was being
sent at the root level of agentParams, causing validation failures
with error: "invalid agent params: at root: unexpected property 'paperclip'"
The paperclip metadata is already included in the message field
via wakeText, so removing the separate paperclip property resolves
the validation error while preserving the necessary information.
Fixes#606
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
The codex-local and cursor-local adapters already wrap the
instructionsFilePath read in try/catch, logging a warning and
continuing without instructions. The claude-local adapter was missing
this handling, causing ENOENT crashes when the instructions file
doesn't exist.
Fixes#529
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
dependabot[bot]
Lennie Sliwinski
Devin Foley
Neeraj Kumar Singh B
uky333
Sergio Sánchez Zavala
Svetlana Zolotenkova
Ismaël O.
Jannes Stubbemann
LIXin Ye
Sherif Kozman
Abhishek gahlot
Harshit Khemani
nullEFFORT
kengraversen
NyDamon
Danial Jawaid
Dotta
Russell Dempsey
Dewaldt Huysamen
Knife.D
Dotta
dotta
Aron Prins
lempkey
kimnamu
Brandon Woo
Dawid Piaskowski
Lucas Kim
Wes Belt
HenkDz
Shoaib Ansari
Mikhail Batukhtin
Richard Anaya
John Wessel
Sai Shankar
Matt Van Horn
Sigmabrogz
Aaron
lin
