## Thinking Path
> - Paperclip is the open source app people use to manage AI agents for
work
> - The Claude local adapter can run agent turns through an ACP (Agent
Client Protocol) server, `claude-agent-acp`, instead of the plain CLI
> - Two separate packages each pin their own copy of that dependency:
`packages/adapters/claude-local` (the server-side adapter) and
`packages/paperclip-runner` (which builds the provider pack baked into
every managed sandbox image)
> - `claude-local` moved to `^0.73.0` in #12730, but `paperclip-runner`
was never bumped past `0.70.0` — nothing keeps the two in sync when only
one changes
> - That split means a sandbox image built from `paperclip-runner`'s
provider pack ships a `claude-agent-acp` the server-side adapter was
never actually compatible with
> - This pull request bumps `paperclip-runner`'s pin to `0.73.0`, the
only version that satisfies both packages' declared ranges at once, and
fixes the matching hardcoded version assertion in
`docker/daytona-runner/Dockerfile`
> - The benefit is one consistent, compatible `claude-agent-acp` version
across both the server host and every sandbox image built from this
source, instead of a silent split that only surfaces as a runtime
failure
## Linked Issues or Issue Description
No public issue exists for this specific split; opening directly per
CONTRIBUTING.md path B, following the bug report template fields.
**What happened?**
`packages/paperclip-runner/package.json` pins
`@agentclientprotocol/claude-agent-acp` at an exact `0.70.0`.
`packages/adapters/claude-local/package.json` requires `^0.73.0` (added
in #12730, 2026-09-02). Nobody re-synced `paperclip-runner`'s pin after
that change — the two packages' dependency graphs are independent, so a
bump in one doesn't propagate to the other. `paperclip-runner`'s copy is
what the fleet sandbox image's provider pack actually ships, so every
managed sandbox built from current source carries a `claude-agent-acp`
version the server-side adapter's own declared compatibility range
excludes.
**Expected behavior**
The two packages' `claude-agent-acp` pins should stay within a mutually
compatible range, so a sandbox image built from this source always ships
a version the server-side adapter actually supports.
**Steps to reproduce**
1. Check `packages/adapters/claude-local/package.json`'s
`@agentclientprotocol/claude-agent-acp` range (`^0.73.0`).
2. Check `packages/paperclip-runner/package.json`'s pin for the same
package (`0.70.0` before this PR).
3. Note that `^0.73.0` on a `0.x` version only admits patch releases
(`>=0.73.0 <0.74.0` per semver caret rules), so `0.70.0` falls outside
it.
**Paperclip version or commit**
`master` as of this PR (paperclip-runner still at `0.70.0` prior to this
change; claude-local's `^0.73.0` requirement landed in #12730).
**Deployment mode**
Any deployment that runs `claude_local` agents through the ACP engine
against a sandbox image built from `packages/paperclip-runner`'s
provider pack (managed cloud sandboxes in particular).
Related PRs for context (not duplicates — none of these touch
`paperclip-runner`'s pin):
- #12730 — introduced the `^0.73.0` requirement in `claude-local`
- #11873 — the last time `paperclip-runner`'s pin moved (`0.69.0` →
`0.70.0`)
- #13105 — separately made an unavailable ACP engine a hard failure
instead of a silent CLI fallback, which is what turned this version
split into a visible, run-blocking error rather than a quiet downgrade
## What Changed
- Bump `@agentclientprotocol/claude-agent-acp` from `0.70.0` to `0.73.0`
(exact pin, matching this package's existing pin style for its other
agent-CLI dependencies) in `packages/paperclip-runner/package.json`.
- Update the corresponding hardcoded version assertion (`test
"$(claude-agent-acp --version)" = "0.70.0"`) in
`docker/daytona-runner/Dockerfile` to `0.73.0`, so its own build-time
check stays accurate instead of failing on the next build for an
unrelated reason.
- `pnpm-lock.yaml` is intentionally **not** included —
`pr-trusted.yml`'s `Validate dependency resolution and regenerate stale
lockfile` step already regenerates it for the merge tree and hands it to
downstream `--frozen-lockfile` jobs as an artifact, so a manual lockfile
commit here would just be stale the moment CI runs.
## Verification
- `0.73.0` is a real published version on npm (confirmed via `npm view
@agentclientprotocol/claude-agent-acp versions`), and it's the *only*
version satisfying claude-local's `^0.73.0` range, so this isn't a guess
at compatibility — it's the unique intersection of both packages'
declared ranges.
- `grep -rn "0\.70\.0" docker/ packages/paperclip-runner/package.json`
after this change shows no remaining stale references to the old pin.
- I did not run a full local install/test pass against a hand-updated
lockfile, since regenerating one locally would conflict with leaving
`pnpm-lock.yaml` untouched per the note above; CI's own
lockfile-regeneration step is the intended verification path for a
manifest-only dependency bump like this one.
- Downstream/full verification (does a sandbox image actually built with
this pin work end-to-end) is tracked separately in `paperclip-cloud` —
an unrelated internal-only repo, so not linked here — where a sibling
fix restores the ACP servers to the runtime `PATH` in the fleet sandbox
image itself; both fixes are needed together for a working sandbox, but
this PR is scoped to the version pin alone.
## Risks
- Low risk: single-line dependency version bump plus a matching
test-assertion update, no code changes. `0.73.0` is a patch release
within claude-local's own already-declared-safe range, so there's no
reason to expect it changes behavior tenants depend on.
- The main risk is unknown breaking changes between `claude-agent-acp`
0.70.0 and 0.73.0 that aren't caught by the version-string assertion
alone (that check only confirms the binary reports the right version,
not that its behavior is unchanged). I have not audited that package's
own changelog between those versions.
- `docker/daytona-runner/Dockerfile` is a parallel/reference image (per
its own header comment, meant to stay aligned with the private
`paperclip-cloud/fleet-sandbox-image/Dockerfile`, which is out of scope
here) — this PR does not touch that other Dockerfile.
## Model Used
Claude Sonnet 5 (`claude-sonnet-5`), via Claude Code, with tool use
(file edits, shell/git, `gh` CLI, `npm view` for version verification).
No extended-thinking mode. Standard Claude Code context window.
## 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 — see Verification: a
manifest-only bump with the lockfile intentionally left to CI's own
regeneration step; no local test run applicable
- [x] I have added or updated tests where applicable — version-pin bump
only, no new behavior to test
- [x] I have updated relevant documentation to reflect my changes — none
applicable
- [x] I have considered and documented any risks above
- [x] All Paperclip CI gates are green — pending CI run on this PR
- [x] 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
🤖 Generated with [Claude Code](https://claude.com/claude-code)
---------
Co-authored-by: Claude Sonnet 5 <noreply@anthropic.com>
## Thinking Path
> - Paperclip manages agents that must write work and report task
outcomes through its API.
> - Local adapters select an execution engine and its permission
settings.
> - A higher ACP Node requirement can make an unchanged installation
lose access to its default engine.
> - The adapter then silently selects CLI, which can change permissions
and block API access.
> - This pull request keeps the engine choice fixed and reports missing
prerequisites before work starts.
> - It also gives explicit Codex CLI runs usable defaults and keeps
managed services on a supported Node runtime.
## Linked Issues or Issue Description
Refs #12215. Related changes: #11792 raised the Node requirement; #13094
addressed separate runner networking behavior. This change fixes the
engine-selection and managed-launcher paths.
**What happened?**
An unchanged agent could switch from ACP to CLI after an upgrade. Codex
CLI then used read-only permissions with networking disabled. The run
could finish without updating its task. Repeated recovery attempts used
the same unavailable setup. Managed updates also skipped the Node check
and did not refresh old launchers.
**Expected behavior**
An unavailable engine must fail with a clear setup error. It must not
silently select another engine. Explicit CLI runs must be able to write
workspace files and call the API unless the operator configures stricter
settings. Managed updates must validate Node and keep child tools on
that runtime.
**Steps to reproduce**
1. Run an ACP-default agent under Node 22 after the ACP minimum rises to
24.11.
2. Leave the engine unset and disable the approval/sandbox bypass.
3. Observe the old adapter select CLI and fail to write task disposition
through the API.
4. Start a managed service with an old launcher and a supervisor PATH
that selects a different Node for child tools.
## What Changed
- Remove automatic engine fallback for Codex, Claude, Gemini, and Kimi.
Check prerequisites for default and explicit ACP selections.
- Return a configuration error with proof that provider work did not
start. Stop automatic continuation retries for this error.
- Enable Codex ACP workspace networking at the actual turn boundary.
Upstream mode presets otherwise force it off even when config.toml
enables it. Preserve explicit network denial and read-only mode.
- Set workspace-write and network access defaults for explicit Codex CLI
runs. Preserve explicit sandbox modes, profiles, and network
restrictions.
- Pin the validated Node directory in managed launcher PATH. Refresh
legacy launchers during installs and npm/Git updates.
- Reject updates on unsupported Node. Keep update checks, dry runs, and
rollback available.
- Synchronize the qualified Codex ACP executable identity across server,
TypeScript runner, Rust runner, and provider-pack launch paths.
- Add regression tests and update engine and installation documentation.
## Verification
- [Full CI passed on the final
head](https://github.com/paperclipai/paperclip/actions/runs/34387099695):
typecheck, build/native runner verification, all general and serialized
test shards, all browser shards, release registry, canary dry run, and
policy checks.
- Greptile: 5/5 on `2c1d6e2815830a5cd39e36c8a082cc0c4441b6c0`, with no
unresolved review findings. Security gates are green.
- Full workspace typecheck and build also passed locally. The final
deployed Linux build passed.
- Full Codex, Claude, Gemini, and Kimi source test suites: 804 passed, 2
skipped. Installer, updater, and launcher tests: 47 passed. Installed
ACP turn-boundary tests: 3 passed. ACP packaging tests: 14 passed.
Focused recovery classification tests also passed.
- Real Linux Codex CLI runs, both fresh and resumed, wrote a workspace
file and reached the control-plane health API with the new defaults.
- Explicit read-only and network-disabled control probes retained those
restrictions.
- A real ACP run on the final deployed Linux build wrote a file and
reached the control-plane API with HTTP 200, without engine fallback.
The same probe failed DNS before the turn-policy patch.
- Executable-identity and installed-policy contracts: 12 passed.
Affected native server tests: 197 passed. Runner factory tests: 21
passed. Rust qualification and native provider integration tests: 11
passed.
- Deployed the production changes to a Linux service on Node 24.20 after
a verified database backup. Health, bootstrap readiness, static UI,
executable/cwd identity, and guarded restart checks passed. The restart
lost no runs.
- Corrected stale Kimi skill-default and Gemini remote-archive fixtures;
both suites pass.
## Risks
- Default or legacy auto engine settings now fail when ACP is
unavailable. Operators who intend to use CLI must select it explicitly.
- Codex CLI now permits workspace writes and networking by default, and
ACP workspace-write turns permit networking by default. Explicit
operator sandbox settings remain authoritative.
- Old managed launchers keep their pinned Node until they are
reinstalled under a supported runtime. An old updater cannot repair
itself; the documentation gives the current installer command.
- Custom service wrappers and global/source installations must configure
their runtime PATH. No database migration is required.
## Model Used
OpenAI Codex, based on GPT-6, with reasoning, repository inspection,
shell execution, and test tools. The exact serving model identifier and
context-window size are 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 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
- [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.
> - Agent adapters select a provider, a model, and a runtime.
> - Runner conversion rejected existing Claude agents. The model list
mixed providers.
> - The native Claude runner rejected custom models and could not launch
on macOS.
> - This pull request fixes conversion, model selection, and verified
macOS execution.
> - It also groups configuration fields consistently across adapters and
opens artifact images in the task gallery.
> - Operators can change an agent configuration and run the selected
model on their Mac.
## Linked Issues or Issue Description
**What happened?**
Converting an existing Claude agent to Paperclip Runner failed with a
Codex-only restriction. ACPX Claude showed unrelated models and required
`claude-sonnet-5`. Its native runtime rejected macOS. Configuration
mixed common model settings with process controls. Artifact cards
labeled “Open gallery” navigated to attachment URLs instead of opening
the task gallery.
**Expected behavior**
Conversion keeps agent identity and compatible settings. ACPX Claude
uses the normal Claude catalog and accepts typed model IDs. Codex uses
the native runner. The verified Claude runtime can launch on macOS ARM64
and x64. Common configuration sections place the same fields together
across adapters. Artifact images open in the shared task gallery with
navigation and downloads.
**Steps to reproduce**
1. Open the configuration of an existing Claude agent.
2. Convert it to Paperclip Runner.
3. Select ACPX Claude and a different catalog model or a typed model ID.
4. Save the agent and run a disposable task on macOS.
5. Inspect configuration and advanced run-policy controls across
adapters.
**Paperclip version or commit**
The bugs were reproduced on `165ca56a22adb60e5fda56045442d9c8498116a8`.
This branch was rebased onto `7ed122911`.
**Deployment mode**
Built from source. Local test-drive instance on macOS ARM64 with an
isolated database.
Related work: #11798 addresses unsupported ACP session options in the
existing adapter path. #13048 addresses working-folder preservation.
This change fixes native runner configuration and launch behavior.
## What Changed
- Remove the Codex-only conversion restriction. Preserve agent identity,
instructions, directories, credentials, and compatible model settings.
Reset incompatible sessions while retaining history.
- Show ACPX Claude and native Codex as distinct provider choices. Remove
ACPX Codex from advertised configuration. Normalize legacy
configurations before fresh runs without rewriting historical run
descriptors.
- Select model catalogs and cache entries by provider. Support refresh
and typed model IDs. Pass exact Claude IDs through session creation,
model changes, and recovery.
- Add verified macOS ARM64 and x64 Claude SDK snapshots. Bound
executable allocation and total snapshot size. Preserve package checks,
dependency isolation, process ownership, cancellation, and Linux
descriptor loading.
- Probe local runtime readiness. Report remote platform checks as
incomplete until the remote runner verifies its runtime.
- Surface actual model rejection and allow correction and retry.
- Repair missing ACPX goal-capability helpers exposed by the post-rebase
live test. Persist and restore the optional capability without breaking
session startup.
- Put Agent identity first and intentionally remove the Capabilities
editor, as requested. This is removal of UI editing, not relocation:
preserve existing capability metadata and API compatibility without
adding another editor. Use the themed select for configurable permission
modes, with normal text instead of monospace.
- Put model and provider under Adapter. Give environment variables their
own section. Fold command and arguments under Configuration. Fold
lifecycle, timeout, and interrupt grace under Advanced Run Policy. Hide
single-option permission controls.
- Open image and video artifact cards in the existing task gallery,
including cards in the artifacts panel. Chat attachment images use the
same gallery. Preserve standalone media previews and download links.
## Verification
- Rebased focused UI/API/database suites: 293 tests passed.
- Rebased native runtime and ACPX suites: 242 passed, 7 skipped.
- Repository typecheck, build, and token gates passed for the runner
changes. Gallery follow-up UI typecheck, build, and token gates also
passed.
- Follow-up UI suites passed (86 tests), packaging checks passed (14
tests), and the final focused runtime suites passed (126 passed, 7
skipped).
- Linux container isolation and lifecycle fixtures passed before rebase
(57 passed, 2 skipped). Rust ACPX provider-session tests passed after
rebase (8 tests).
- Browser tests completed actual Claude and native Codex tasks on macOS
ARM64. They covered conversion, catalog refresh, a non-default catalog
model, a typed `haiku` ID, save/reload, cancel, follow-up session
continuity, invalid-model errors, and recovery.
- Final-revision live tests completed a typed Claude task, a follow-up
with the same provider session, and a native Codex task on macOS ARM64.
- Browser tests confirmed the moved interrupt-grace field saves and
survives reload. Cross-adapter tests cover Claude, Codex, Gemini,
process, gateway, and schema forms.
- Full local run: 7,080 passed, 30 skipped, and two timeouts. Both
timeout suites passed on isolated rerun (84 tests); the failures were
the plugin login-worker exit diagnostic and the runner real-server
vertical slice.
- Final follow-up checks: 50 registry tests and 45 snapshot/installation
tests passed (6 platform-specific skips). Oversized executable rejection
is covered before allocation or reading; unsupported-platform tests
invoke the real installation probe.
- Runner head `ddb5101c483a297f74875ab96b3c66035b002d50`: all CI gates
green, including full runner verification, repository build, typecheck,
general/serialized server suites, browser tests, and canary dry run. [CI
run](https://github.com/paperclipai/paperclip/actions/runs/34286178670).
- Greptile: 5/5 on that runner head. All four review threads resolved.
Superagent, Socket, and Snyk checks green.
- After snapshot hardening, another real Claude task completed on this
Mac using the rebuilt runtime.
- Gallery follow-up: 148 focused tests passed, covering artifact
selection, shared attachment collections, deduplication, image/video
cards, standalone previews, downloads, and closing. Live browser
verification completed on the settings follow-up: artifact selection,
6-image pagination with wrapping, download action, and closing all
stayed on the same task URL. All checks passed on gallery head
`96136da58ff195bf6ca00b281eb3022ad12d7bd8`: [CI
run](https://github.com/paperclipai/paperclip/actions/runs/34287987536).
Greptile returned 5/5 on that exact head with no unresolved threads.
- Final settings polish: 96 focused tests, UI typecheck/build, and token
gates passed. A real browser walkthrough verified readable permission
options, identity placement, Capabilities removal, and permission
save/reload. Original test-agent permission mode restored. All 31 checks
passed on final head `e46540d6bf32bfb0566dca16b2f4a75ba437618c`: [CI
run](https://github.com/paperclipai/paperclip/actions/runs/34292797886).
Greptile returned 5/5 with no unresolved threads.
## Risks
- Capabilities intentionally has no editable UI field after this change.
Existing values remain readable and API-compatible; removing the field
does not erase stored metadata.
- macOS launch now copies verified package files into private snapshots.
The implementation must retain isolation and clean up snapshots on exit.
- Runtime provider or model changes reset the current session.
Historical runs remain available.
- The macOS x64 SDK executable digest was verified, but a live Intel Mac
run was not available. Linux verification used container fixtures, not a
real Claude task.
- Remote environment tests report a warning when only the platform has
been checked. They do not claim package readiness from the server host.
## Model Used
OpenAI Codex, based on GPT-6. The exact served model identifier and
context-window limit are not exposed in this session. Used reasoning,
repository inspection, code execution, Rust and TypeScript tests, and
browser automation.
## 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 (focused suites and both
timeout suites on rerun; full-run counts above)
- [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
- [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 Runner needs its runtime preinstalled for fast sandbox
startup.
- Native and local adapters should launch one current CLI installation
per provider.
- An older global copy can shadow that installation, and exact native
compatibility pins must match it.
- Update the qualified releases and binary digests, expose shared CLI
entrypoints from the provider pack, and prefer the image-owned bin
directory.
- Keep dependency installation in the image build; task startup only
discovers, links, and verifies artifacts.
## Linked Issues or Issue Description
**What happened?**
Remote native startup rejected a stale global Codex, while CLI-only
images lacked runnerd entirely.
**Expected behavior**
An image-baked runtime starts without uploading binaries or installing
packages. All adapters share the same current provider CLI.
**Steps to reproduce**
Start a native remote task with the old global Codex and the updated
runtime available only under `/opt/paperclip-runner/bin`.
**Paperclip version or commit**
Discovery behavior at `54a99d884`.
**Deployment mode**
Docker with a remote sandbox.
## What Changed
- Prefer `/opt/paperclip-runner/bin`, then the user's local bin
directory, then PATH. Existing metadata and version validation remains
in force.
- Qualify Codex 0.153.4, OpenCode 1.18.29, and Claude SDK 0.3.263 / CLI
2.1.263. Update binary digests, TypeScript/Rust checks, registry
defaults, and the displayed OpenCode version together.
- Share Codex and Claude's native executable with the ACP bridges
through exact dependency overrides. Preserve the separately qualified
ACP bridge implementations and their security patches.
- Expose shared provider-pack CLI launchers; fail the pack build if
Codex ACP resolves a separate Codex installation. Update the eval
image's other agent CLIs to current stable releases and remove duplicate
global provider installs.
- Document the single-current-CLI policy in source comments and
development guidance. Latest stable releases are resolved at
review/build preparation and pinned; task startup never auto-updates.
## Verification
- Native-session and adapter-registry suites: 158 tests passed.
- Provider suites: 88 tests passed, 7 Linux-only checks skipped on
macOS. One existing macOS temporary-path alias assertion passed when
rerun with canonical `TMPDIR=/private/tmp`.
- Package-contract and OpenCode materialization tests: 11 passed.
- Full typecheck, build, and token gates passed. Rust
native-provider/recovery tests: 19 passed.
- Broad local suite: 5,974 passed, 23 failed, 41 skipped. Failures are
in unchanged macOS workspace/path/port and connection suites; focused
runtime tests pass. All latest-head Linux PR checks passed, including
the full test shards, typecheck, build, runner verification, browser
suites, and canary dry run.
- The standalone fleet image built with one current provider CLI each
and passed native Codex/Claude binary-integrity checks. A disposable
Daytona sandbox reported ready in 798 ms; its baked runner completed an
API-key `gpt-5.6-luna` turn in 2,430 ms and returned the expected marker
with a usage receipt. No runtime artifacts were uploaded or installed.
- The normal shared `codex exec` entrypoint also completed an API-key
`gpt-5.6-luna` turn in 2,321 ms.
- Both image builds verify the complete generated lockfile against a
reviewed SHA-256 before package installation or lifecycle execution.
Root lockfile changes remain CI-owned. Merge and rollout remain on hold
for operator review.
## Risks
- Updating provider CLIs changes their behavior for all adapters;
version probes and live native smoke testing are required before image
promotion.
- The image-owned directory takes precedence. Its entries must launch
the same shared CLI as the global PATH, not a private older/newer copy.
- Application qualification pins and the deployed image must move
together. No startup fallback installation is added.
- No schema or authentication-policy changes.
## Model Used
OpenAI GPT-6 (Codex). The session does not expose a more specific model
ID or context-window size. Used reasoning, repository inspection, code
execution, and browser 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 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
- [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 manages AI agents that perform work.
> - Paperclip Runner connects durable task runs to local provider
processes.
> - The full-stack paid matrix exposed failures after the runner
integrity repair.
> - Verified JavaScript entrypoints lost their relative module graph
when Linux executed them through descriptor paths.
> - Returned provider startup errors also remained pending and became
indeterminate after recovery.
> - Sparse Codex tool lifecycle events lost the `write_document`
identity before task transcript projection.
> - This pull request repairs those three boundaries and makes the
structured-question fixture deterministic.
> - The benefit is repeatable provider startup, exact failure replay,
and correct inline Plan placement.
## Linked Issues or Issue Description
Refs #12721 and #12700.
**What happened?**
The paid runner matrix failed ACPX and OpenCode startup before provider
session creation. The runner journal then replaced the original startup
error with an indeterminate recovery result. Native Codex saved a Plan
but rendered it only as a fallback card. A legacy Claude waiting reply
could also echo the reserved terminal marker before the answer arrived.
**Expected behavior**
Verified JavaScript providers must start from immutable
descriptor-backed artifacts. Returned startup failures must persist as
terminal failed command results. Native tool lifecycle updates must
preserve the `write_document` boundary. Pre-answer fixture output must
not contain the reserved terminal marker.
**Steps to reproduce**
1. Run the local provider cells in the Runner Full-Stack E2E workflow.
2. Observe ACPX and OpenCode fail during `session.open` before provider
execution.
3. Observe recovery report `execution_indeterminate` instead of the
original startup error.
4. Run the native Codex Plan cell and observe the fallback Plan card
after the tool activity row.
5. Run the legacy Claude structured-question resume cell and observe an
early marker echo in waiting prose.
**Paperclip version or commit**
`0f9452101740835ce0b1488a204bf48acd5bafc3`
**Deployment mode**
Local development with the paid GitHub Actions acceptance workflow.
## What Changed
- Bundle the ACPX sidecar and OpenCode proxy as self-contained Node ESM
entrypoints before hashing and verified descriptor launch.
- Anchor ACPX dynamic provider package resolution at a
controller-derived provider-pack root and keep that root out of the
provider child environment.
- Persist executor-returned startup errors as redacted durable failed
command results while retaining indeterminate recovery for true process
death.
- Coalesce sparse native tool items by stable ID so a late
`write_document` name, input, and result reach the transcript boundary
once.
- Forbid the structured-question fixture from spelling or announcing its
reserved terminal marker before the user answers.
## Verification
- Rust and TypeScript regression tests cover durable failed replay, true
crash ambiguity, bundle closure, package-root derivation, environment
filtering, exact Codex tool lifecycle coalescing, and prompt
determinism.
- Local execution is intentionally limited to formatters and static diff
checks. GitHub Actions will run tests, type checks, builds, and security
checks.
- After ordinary CI is green, scoped paid cells will validate one ACPX
launch, one OpenCode launch, native Codex Plan projection, and legacy
Claude structured resume before a complete matrix rerun.
- Prior failing matrix:
https://github.com/paperclipai/paperclip/actions/runs/33682434315
## Risks
- Bundling changes the bytes covered by provider launch hashes.
Provider-pack generation already hashes the final built files.
- ACPX still loads qualified provider packages dynamically. The
controller supplies a normalized package root, while existing version,
digest, path, and descriptor checks remain active.
- Durable `failed` is terminal. Replays return the same redacted result
and do not execute the provider effect twice.
> 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 based on GPT-5 with agentic reasoning, repository
inspection, code editing, Git, parallel subagents, and GitHub Actions
coordination. The exact deployed snapshot and context-window size are
not exposed to this task.
## 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 linked related public work or described the bug in
this PR
- [x] I have not referenced internal or instance-local Paperclip issues
or links
- [x] My branch name describes the change and contains no internal
ticket id
- [ ] I have run tests locally and they pass (intentionally deferred to
GitHub Actions)
- [x] I have added or updated tests where applicable
- [x] No documentation change is required for this runtime repair
- [x] I have considered and documented the 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
## Thinking Path
> - Paperclip is the open source app people use to manage AI agents for
work.
> - The Claude local adapter lets operators select a Claude model for an
agent.
> - Claude Fable 5.1 was absent from the adapter model lists.
> - The adapter runtime also used a Claude Code build that rejected
Fable 5.1.
> - This pull request adds the direct Anthropic ID and the AWS Bedrock
inference profile ID.
> - It also updates the Claude ACP runtime and keeps the Paperclip usage
and isolation patches.
> - The benefit is that operators can select and run Claude Fable 5.1
through the Claude adapter.
## Linked Issues or Issue Description
Refs #8810. That issue covers related model ID handling. This change
does not change provider-prefixed model IDs.
**Agent or provider**
Claude Code through the built-in `claude_local` adapter. The requested
model is Claude Fable 5.1.
**Why this adapter is useful**
Operators can use Fable 5.1 without entering an undocumented model ID.
The configured model also reaches both supported Claude execution lanes.
**How the agent is invoked**
The CLI lane sends `--model claude-fable-5-1`. The ACP lane sends
`ANTHROPIC_MODEL=claude-fable-5-1` to
`@agentclientprotocol/claude-agent-acp`.
**Are you willing to implement it?**
Yes. This pull request includes the implementation and tests.
**Additional context**
Claude Code 2.1.232 rejected Fable 5.1 and required version 2.1.251 or
newer. ACP package 0.73.0 includes Claude Code 2.1.257. The update keeps
Paperclip's usage metadata and isolated-context behavior.
## What Changed
- Added `claude-fable-5-1` to the direct Claude fallback list.
- Added `us.anthropic.claude-fable-5-1` to the AWS Bedrock list.
- Kept the existing default model at the first position in each list.
- Updated the Claude ACP dependency from 0.70 to 0.73.
- Carried the Paperclip usage and isolated-context changes into the 0.73
patch.
- Added a Claude Code 2.1.251 minimum-version preflight for Fable 5.1
when using the standard `claude` executable, surfaced in both adapter
Test and execution. Explicit custom wrappers retain their existing
compatibility contract.
- Kept local adapter Tests from executing caller-selected binaries: when
runtime `PATH` selects a different Claude executable than the trusted
probe, the Test warns and defers the authoritative version check to
execution instead of approving or rejecting the alternate installation.
- Added tests for model listing, discovery deduplication, Bedrock
filtering, model pass-through in both execution lanes, old-CLI rejection
before launch, custom-wrapper compatibility, and local runtime-PATH
mismatch handling.
## Verification
- `pnpm --filter @paperclipai/adapter-claude-local typecheck`
- `pnpm exec vitest run
packages/adapters/claude-local/src/server/execute.remote.test.ts
packages/adapters/claude-local/src/server/test.remote.test.ts
packages/adapters/claude-local/src/server/test.probe.test.ts
packages/adapters/claude-local/src/server/acp.test.ts
server/src/__tests__/adapter-models.test.ts` (72 tests passed)
- `node --test scripts/acpx-patch-packaging.test.mjs` (13 tests passed)
- `pnpm -r typecheck`
- `pnpm build`
- A local Paperclip agent run completed with `usageJson.model` set to
`claude-fable-5-1` through ACP 0.73.0 and its bundled Claude Code
2.1.257.
- `pnpm test:run` completed 5,638 passing tests and 24 skipped tests. It
also found 24 failures in unrelated workspace-runtime,
path-canonicalization, and runtime-exposure tests on macOS with Node 26.
These failures do not touch this diff. Clean pull request CI is the
final full-suite gate.
## Risks
- The ACP dependency update can change Claude runtime behavior outside
model selection. Focused ACP tests, the full typecheck, the production
build, and a real local Fable run reduce this risk.
- The 0.73 patch must stay aligned with the installed ACP version.
Dependency-resolution CI verifies the manifest and patch pair.
- Fable 5.1 adds a short `claude --version` preflight to standard
CLI-lane Tests and runs. The result is intentionally not cached so an
in-place Claude Code upgrade takes effect without restarting Paperclip.
Explicit custom wrappers are not version-probed because their output and
compatibility contract can differ from the standard executable.
- Local Tests preserve the existing deny-by-default probe boundary and
do not execute a binary selected by caller-controlled `PATH`. A
mismatched runtime binary produces an explicit warning without blocking
an otherwise valid setup; execution independently validates the actual
runtime-selected CLI before launch.
- The AWS Bedrock identifier differs from earlier IDs because Fable 5.1
has no `-v1` suffix. The model-list test locks this exact value.
- There is no schema change or migration.
> 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
Provider: OpenAI. Model: GPT-5 Codex. The host did not expose a more
specific model ID or context-window size. Capabilities used: agentic
reasoning, repository editing, shell execution, web research, and local
runtime 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 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
- [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 runner package already provides the production protocol and
execution spine.
> - Contributors still need stable SDK surfaces, deterministic test
tools, and local inspection tools.
> - Those surfaces share generated contracts and must change as one
package boundary.
> - This pull request adds the package-local SDK, labs, examples, and
drift checks.
> - The benefit is a reviewable developer platform that does not change
application execution selection.
## Linked Issues or Issue Description
**Subsystem affected**
`packages/paperclip-runner` — runner SDK, conformance tools, and
developer tooling.
**Problem or motivation**
The production runner spine is present, but package consumers cannot
build deterministic integrations, inspect sessions, or verify
provider-neutral behavior through supported surfaces.
**Proposed solution**
Add browser, React, standalone, live-session, scenario, conformance, and
evaluation surfaces. Add generated contract inventories and
package-local verification scripts. Keep production application routing
unchanged.
**Alternatives considered**
We considered splitting each generated catalog, SDK surface, and demo
into separate pull requests. Those changes share exports, fixtures, and
drift gates. Splitting them would create intermediate package states
that do not build.
**Roadmap alignment**
No overlapping item appears in `ROADMAP.md`. This work extends the
runner package that is already on `master`.
## What Changed
- Add browser, React, standalone, live-session, and issue-thread SDK
surfaces.
- Add deterministic mock control-plane, scenario, conformance, replay,
and evaluation tools.
- Add bounded Codex, OpenCode, and ACPX development transports and
fixtures.
- Keep deferred managed-provider execution fail-closed. Persisted
compatibility data remains readable.
- Add generated capability inventories with their source files and drift
checks.
- Add examples, package documentation, browser checks, and
clean-consumer checks.
- Preserve the reviewed protocol bounds, replay compatibility aliases,
process environment isolation, and semantic redaction limits.
- Update the ACPX package patch that the existing workspace patch
registry already tracks.
- Do not change `pnpm-lock.yaml`, repository workflows, server runtime
selection, or the application UI.
## Verification
GitHub Actions is the verification authority for this pull request. The
repository CI, package TypeScript and Rust checks, package tests,
generated-output drift checks, browser checks, security scans, and
Greptile review must pass on the exact head.
Local test suites were not run because this series uses parallel GitHub
Actions for verification.
## Risks
This is a large greenfield package change. The main risks are public
export drift, generated-output drift, and optional React consumer
compatibility. Package boundary checks, clean-consumer checks, and
browser tests cover those risks. Production adapter selection and server
execution are outside this pull request.
## Stack
1. **This PR:** runner SDK and developer tooling.
2. [Codex production server
integration](https://github.com/paperclipai/paperclip/pull/12616).
3. [Provider-neutral task-thread
UI](https://github.com/paperclipai/paperclip/pull/12617).
## Model Used
OpenAI Codex, GPT-5, high-reasoning mode, with tool use and code
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 searched GitHub for duplicate or related PRs and linked
them above
- [x] I have described the issue in-PR following the feature request
template
- [x] I have not referenced internal/instance-local Paperclip issues or
links
- [x] My branch name describes the change and contains no internal
Paperclip ticket id
- [ ] 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 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
## Thinking Path
> - Paperclip Runner already has a hardened ACPX path for Codex.
> - Claude can reuse that protocol only with an exact package/model
profile and provider-lifetime fencing.
> - Pi needs a separately spawned runtime whose executable does not yet
have the descriptor-confined verified launch used by the ACP server.
> - This pull request therefore activates Claude only and keeps Pi
unavailable before installation or process launch.
## Linked Issues or Issue Description
**Subsystem affected**
Paperclip Runner ACPX driver, runtime host, sidecar, backend factory,
package dependency, and provider conformance tests.
**Problem or motivation**
The production ACPX backend was Codex-only. Claude needs the same
fail-closed model, authorization, cancellation, cleanup, and recovery
boundaries without exposing an unsafe secondary runtime path.
**Proposed solution**
Generalize the hardened ACPX runtime for the exact qualified `claude`
profile, add the pinned Claude ACP package and reviewed isolation patch,
and reject Pi before installation, backend construction, sidecar
initialization, Rust session admission, or process creation.
**Alternatives considered**
Activating Pi in this PR was rejected after security review because its
secondary runtime executable was pathname-based and lacked the verified
descriptor/snapshot boundary. Pi is deferred to a dedicated follow-up.
Replaying the older generic ACPX implementation was rejected because it
predates current hardening.
**Roadmap alignment**
ROADMAP.md does not list a conflicting ACPX-provider project. This
extends the existing Runner provider architecture.
## What Changed
- Generalized the ACPX backend, driver, runtime adapter, host, and
sidecar for the qualified Claude profile.
- Added Claude ACPX activation through its exact pinned package/model
pair and isolated-settings patch.
- Added provider-lifetime fencing for non-Codex qualified ACPX sessions.
- Kept Pi dependencies and its patch out of the package and build
configuration.
- Added fail-closed Pi rejection at driver validation, backend
construction, runtime-host admission, sidecar initialization, and Rust
session validation.
- Added focused tests for Claude selection, model enforcement, lifecycle
fencing, cancellation, recovery, and Pi rejection.
- Did not change or commit `pnpm-lock.yaml`; CI regenerates the PR
lockfile under the existing repository policy.
## Verification
- GitHub Actions is the authoritative verification environment for this
PR.
- CI runs dependency policy, runner package checks, protocol parity,
typecheck, build, security, and stack policy.
- Local tests were not run because this checkout is resource
constrained, per the requested workflow.
## Risks
- Claude package behavior can drift from the qualified protocol; the
package and patch are pinned and admission verifies the exact profile.
- Unsupported providers and models fail closed.
- Pi remains unavailable until descriptor-confined verified launch
exists for its separate runtime.
- Existing Codex ACPX behavior remains covered by shared conformance
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.
## Model Used
OpenAI Codex, GPT-5.6, with repository tools, code execution, and
parallel agent review.
## Checklist
- [x] I have included a thinking path that traces from project context
to this change
- [x] I have specified the model used
- [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
- [x] I have described the issue in the PR
- [x] I have not referenced internal or instance-local Paperclip issues
or links
- [x] My branch name contains no internal task identifier
- [ ] I have run tests locally and they pass — GitHub Actions is
authoritative for this resource-constrained checkout
- [x] I have added or updated tests where applicable
- [x] I have updated relevant documentation
- [x] I have considered and documented risks
- [ ] All applicable Paperclip CI gates are green
- [ ] Greptile is 5/5 with no actionable findings
## Stack
- Position: lowest unmerged PR
- Base: `master`
- Previous:
[#12588](https://github.com/paperclipai/paperclip/pull/12588), merged
qualified OpenCode runtime
- Next: [#12591](https://github.com/paperclipai/paperclip/pull/12591),
native application integration
## Thinking Path
> - Paperclip is the open source app people use to manage AI agents for
work.
> - The package-local host boundary is ready for a concrete ACP
implementation, but the first production profile is Codex only.
> - ACPX must not inherit the server process environment or choose an
executable by pathname after admission.
> - Codex must not re-enable ambient apps, memory, skills, MCP
configuration, or instructions inside its isolated home.
> - This pull request pins only the two required production packages and
applies narrowly tested host patches.
> - The benefit is a minimal dependency boundary that follows the
repository's CI-owned lockfile process.
## Linked Issues or Issue Description
**Agent or provider**
Codex through `acpx@0.13.1` and `@agentclientprotocol/codex-acp@1.6.2`.
**Why this adapter is useful**
The injected runtime host needs a concrete ACP session manager and the
exact reviewed Codex ACP server. Upstream ACPX does not yet expose a
host-owned spawn callback, and upstream Codex ACP does not yet apply
Paperclip's isolated instruction, MCP, app, memory, and skill boundary.
Both behaviors are required before the dependency can execute inside the
runner.
**How the agent is invoked**
The next pull request will adapt these pinned packages to the private
runtime host. ACPX receives a host-owned callback that consumes the
already verified executable lease. Codex receives only the isolated
environment, explicit base instructions, explicit MCP servers, and the
skills rooted in its private `CODEX_HOME`. This pull request alone does
not spawn either package or register an adapter.
**Additional context**
This pull request is stacked on #12399. It adds no Pi, Claude, AWS, SDK,
lab, browser, or UI dependency. It intentionally does not commit
`pnpm-lock.yaml`: the repository policy job regenerates a manifest-only
PR lockfile artifact for downstream frozen installs, and the lockfile
bot updates master separately.
## What Changed
- Pin `acpx` to `0.13.1` and the Codex ACP server to `1.6.2` in the
runner package.
- Register both patches in the pnpm 9 root configuration and newer-pnpm
workspace configuration.
- Preserve the existing embedded-Postgres and ACPX 0.12 patch entries
used by other packages.
- Patch ACPX to evaluate an allowlisted environment at child-spawn time
and keep spawn cwd out of provider-visible session identity.
- Patch ACPX to accept a host-owned spawn callback with the resolved
arguments and options, allowing the verified command lease to own
execution.
- Patch Codex ACP to retain runner-owned MCP server identity in
permission requests.
- Patch Codex ACP to pass explicit Paperclip base instructions on both
start and resume.
- In isolated mode, disable ambient apps, memory, and existing MCP
configuration; load skills only from `CODEX_HOME`; and configure only
requested servers.
- Add a package contract test that enforces exact versions, Codex-only
dependency scope, both pnpm patch registries, and every required patch
hook.
## Verification
- Both patch files dry-apply successfully to fresh published tarballs
for `acpx@0.13.1` and `@agentclientprotocol/codex-acp@1.6.2`.
- A local no-lockfile install applied both patches; their runtime
markers and exact installed versions were inspected.
- Runner TypeScript typecheck — passed against the patched packages.
- Runner package tests — passed: 16 Node protocol/package tests and 426
Vitest tests.
- `pnpm -r typecheck` — passed for all applicable workspaces.
- `pnpm build` — passed, including runner binary, server, UI, and
workspace packages.
- `git diff --check` — passed.
- The diff contains 6 files and does not change `pnpm-lock.yaml`, a
GitHub workflow, server selection, or UI behavior.
## Risks
The primary risk is drift between published package contents and
checked-in compiled patches. Exact versions are pinned, both patches are
exercised by package-contract gates, and CI performs the authoritative
regenerated-lockfile frozen install. The spawn callback does not grant a
new executable path: the following adapter must consume the opaque
verified command lease. Codex isolation changes activate only when
`PAPERCLIP_ACPX_ISOLATED_CONTEXT=1`, so existing direct Codex adapters
are unaffected.
## Model Used
OpenAI Codex with GPT-5 and 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 linked an existing public item or described the
issue in this PR
- [x] I have not referenced internal or instance-local Paperclip issues
or links
- [x] My branch name describes the change and contains no internal task
identifier
- [x] I have run the affected tests locally and they pass
- [x] I have added or updated tests where applicable
- [x] I have documented the dependency, patch, isolation, and lockfile
boundaries
- [ ] All applicable GitHub Actions are green
- [ ] Greptile is 5/5 with every actionable comment resolved
- [x] I will address all review findings before requesting merge
## Thinking Path
> - Paperclip is the open source app people use to manage AI agents for
work.
> - The Paperclip runner needs a safe boundary before it can launch
ACP-compatible agents.
> - A caller-controlled command, model, environment, or frame could
bypass that boundary.
> - The ACPX transport contract in #12386 defines the allowed messages
but does not bind a runtime profile.
> - This pull request defines closed, versioned profiles and validates
the launch inputs around that contract.
> - The benefit is a small and reviewable trust boundary before any ACPX
process can become available.
## Linked Issues or Issue Description
**Agent or provider**
ACPX sidecar support for the qualified Pi, Claude, and Codex ACP
servers.
**Why this adapter is useful**
The runner needs one bounded process boundary for ACP-compatible
providers. A closed profile prevents an untrusted run from selecting an
arbitrary executable, package version, or model.
**How the agent is invoked**
A later pull request will launch an internal sidecar from an exact
profile. This pull request only validates profiles, environment values,
and protocol frames. It does not add an executable dependency or enable
an adapter.
**Additional context**
This pull request is stacked on #12386. It keeps the existing direct
adapters and the Codex runner path unchanged.
## What Changed
- Add a closed profile table for the qualified Pi, Claude, and Codex ACP
servers.
- Require the exact qualified model and return an isolated profile value
to callers.
- Add an agent-specific environment allowlist with entry and aggregate
size limits.
- Add strict parsing for bounded sidecar requests and structured plan
values.
- Reject unknown fields, unsupported protocol versions, invalid
identifiers, null bytes, cyclic values, and oversized input.
## Verification
- Runner TypeScript typecheck — passed.
- Runner TypeScript tests — 40 files and 362 Vitest tests passed; 11
Node contract tests passed.
- `pnpm -r typecheck` — passed for all applicable workspaces.
- `pnpm build` — passed, including runner binary, server, UI, and
workspace packages.
- Prettier and `git diff --check` — passed.
- The diff contains 6 files and does not change `pnpm-lock.yaml`, a
workflow, a package dependency, or a public export.
## Risks
The main risk is accepting more launch state than the sidecar needs. The
implementation uses an agent-specific allowlist, rejects null bytes, and
enforces per-entry and aggregate bounds. This pull request does not
launch a process or expose a new adapter, so production and
direct-adapter behavior remain unchanged.
## Model Used
OpenAI Codex with GPT-5 and 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 linked an existing public item or described the
issue in this PR
- [x] I have not referenced internal or instance-local Paperclip issues
or links
- [x] My branch name describes the change and contains no internal task
identifier
- [x] I have run the affected tests locally and they pass
- [x] I have added or updated tests where applicable
- [x] I have documented the compatibility and security boundary
- [ ] All applicable GitHub Actions are green
- [ ] Greptile is 5/5 with every actionable comment resolved
- [x] I will address all review findings before requesting merge
## Thinking Path
> - Paperclip is the open source app people use to manage AI agents for
work
> - #12227 vendored a pnpm patch of the `postgres` driver to stop a
teardown race (`nextWrite` firing after the socket is nulled) from
crashing the process and failing green CI shards
> - Maintainer call: carrying a vendored driver patch is not worth it
for a CI flake — the patch adds a maintenance obligation on every future
driver upgrade
> - The race is an upstream bug in `postgres@3.4.9`; the plan is to wait
for an upstream release that fixes it and bump the dependency instead
> - This pull request reverts #12227 in full: the patch file, its
`package.json` registration, and the regression test that exercised the
patched behavior
> - The benefit is an unmodified dependency graph; the known flake
signature returns and is retried when it bites
## Linked Issues or Issue Description
Reverts #12227.
**What existing behavior does this improve?**
Dependency hygiene: `postgres@3.4.9` is consumed unmodified again, with
no `pnpm.patchedDependencies` entry to re-evaluate on every driver
upgrade.
**Current behavior**
The repo carries `patches/postgres@3.4.9.patch` (null-socket guard in
the driver's deferred write flush, plus an `execute()` refusal on
socketless connections) and a regression test for it.
**Proposed behavior**
Plain upstream `postgres@3.4.9`. The teardown race stays an upstream
bug: a green test shard can occasionally fail with `Vitest caught 1
unhandled error` and `TypeError: Cannot read properties of null (reading
'write')` at `Immediate.nextWrite`; the remedy is retrying the shard
until an upstream driver release fixes the race and we bump.
**Reason and benefit**
A vendored driver patch is a standing maintenance cost that outweighs
the flake it suppressed.
## What Changed
- Reverts #12227 (`6c7c0fd1f`) in full: removes
`patches/postgres@3.4.9.patch`, its `pnpm.patchedDependencies`
registration in root `package.json`, and
`packages/db/src/postgres-driver-teardown.test.ts`. No lockfile
involvement — the merged commit never touched `pnpm-lock.yaml` and the
refresh bot had not yet recorded the patch.
## Verification
- `pnpm install` on the reverted tree is coherent; the full
`packages/db` suite passes (26 files / 100 tests).
- `git revert` applied cleanly with no conflicts.
## Risks
- Low. This restores the exact pre-#12227 state. The known flake
signature returns; it fails jobs whose tests all passed and is cleared
by retrying the shard.
## Model Used
Claude Fable 5 (`claude-fable-5`, Anthropic) via Claude Code — agentic
coding session 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
- [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
- [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
## Thinking Path
> - Paperclip is the open source app people use to manage AI agents for
work
> - The server and its test suites talk to PostgreSQL through the
`postgres` (postgres.js) driver, and tests routinely tear their
databases down while connections still carry traffic
> - The driver flushes small buffered frames from a `setImmediate`, and
that deferred flush calls `socket.write()` without checking that the
socket still exists; a reserved connection whose backend died keeps
accepting queries, so the flush can fire with a null socket
> - The resulting `TypeError` escapes from a timer callback with no
try/catch above it, crashing the process — in CI this fails suites whose
tests all passed ("Vitest caught 1 unhandled error"), and the same crash
is reported against the driver in the wild after ECONNRESET
> - The latest driver release (3.4.9) still has the bug, so this pull
request adds a pnpm patch guarding the flush and normalizing timer state
on close, plus a deterministic regression test
> - The benefit is CI that no longer fails randomly on a teardown race,
and production processes that survive a database connection dying at the
wrong moment
## Linked Issues or Issue Description
No public issue exists; the underlying problem follows the bug-report
template.
**What happened?**
CI jobs fail with all tests passing: vitest reports `Vitest caught 1
unhandled error during the test run` with `TypeError: Cannot read
properties of null (reading 'write')` at `postgres/src/connection.js`
`Immediate.nextWrite`. The attribution points at whichever test file
happened to be running (e.g. `native-codex-runner.integration.test.ts`),
because the throw comes from a process-level timer callback, not from a
test. The identical crash is reported against the upstream driver by
other projects after `ECONNRESET` (e.g. immich-app/immich#25098).
**Expected behavior**
A connection dying between a write being scheduled and its deferred
flush must settle the affected queries through the driver's normal
connection-error path, never throw from a bare timer callback.
**Steps to reproduce**
Run the new `packages/db/src/postgres-driver-teardown.test.ts` with the
patch removed: reserve a connection (`sql.reserve()` — the same surface
`sql.begin()` uses), destroy the backend socket, wait for the client to
process the close, then issue one query on the reserved connection. The
deferred flush fires one tick later with `socket === null` and crashes
the process with exactly the CI signature.
**Paperclip version or commit**
master `198fc8b28`, `postgres@3.4.9` (latest release; bug still present
on the driver's master branch).
## What Changed
- `patches/postgres@3.4.9.patch` (new, wired via
`pnpm.patchedDependencies`): `nextWrite` returns without writing when
`socket === null`, dropping the buffered bytes — the close path has
already settled every in-flight query, so those bytes have nowhere to
go. The `closed()` and `terminate()` handlers additionally reset
`nextWriteTimer`/`chunk` after `clearImmediate`, so a stale cleared
handle cannot silently block a future reconnect's first flush. All three
shipped builds (`src`, `cjs`, `cf`) get the identical change.
- `packages/db/src/postgres-driver-teardown.test.ts` (new):
deterministic reproduction against a minimal in-process fake wire server
(startup auth + an empty result for the `fetch_types` bootstrap).
Asserts the late query settles with `CONNECTION_DESTROYED` through
`sql.end()` instead of crashing the process.
## Verification
- The regression test fails against unpatched `postgres@3.4.9` with the
exact CI signature (verified by running the same scenario against an
unpatched checkout) and passes with the patch.
- Full `packages/db` suite: 27 files / 101 tests pass.
- Spot-checked server suites that exercise the database through the
patched driver.
## Risks
- Low. The behavioral change activates only in a state that previously
crashed the process (write flush with no socket). Dropping the buffered
bytes matches what the connection's close path already promised callers:
every in-flight query has been settled with a connection error.
- The timer/chunk reset in `closed()`/`terminate()` prevents a
theoretical stale-handle hang after reconnect; on the normal path both
were already reset by `nextWrite`.
- The patch pins to `postgres@3.4.9`; a future driver upgrade will
surface the patch for re-evaluation (pnpm fails loudly on version
mismatch), and the guard can be dropped if the fix lands upstream.
## Model Used
Claude Fable 5 (`claude-fable-5`, Anthropic) via Claude Code — agentic
coding session with tool use (driver source analysis, wire-protocol fake
server, local test 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 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
- [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
## Thinking Path
> - Paperclip is the open source control plane people use to manage AI
agents for work.
> - Local agent heartbeats need durable process identity so the server
can supervise them.
> - The ACPX runtime owns the child process used by `codex_local`
sessions.
> - ACPX did not expose the child PID and start time to the Paperclip
adapter.
> - Warm ACPX runtimes can also serve a later heartbeat without a new
spawn event.
> - A hot restart could therefore classify a live Codex run as lost
because its heartbeat row had no process identity.
> - This pull request forwards ACPX spawn identity, reuses it for
compatible warm heartbeats, and fails closed when identity cannot be
persisted.
> - The benefit is reliable hot-restart adoption for eligible local
Codex runs.
## Linked Issues or Issue Description
No matching public GitHub issue was found.
**What happened?**
A `codex_local` heartbeat could run through ACPX without a persisted
`processPid` or `processStartedAt`. A Paperclip hot restart then had no
durable identity for the live ACP child. Recovery could classify the run
as `process_lost` even while the child was still alive.
**Expected behavior**
ACPX reports the real child PID and start time before the first prompt.
A compatible warm runtime reports the same known identity to each later
heartbeat that reuses the child. ACPX stops the child if the identity is
invalid or persistence fails. Hot-restart recovery can then adopt the
live run.
**Steps to reproduce**
1. Start a `codex_local` heartbeat through the ACPX execution lane.
2. Keep the run active during a Paperclip hot restart.
3. Inspect the heartbeat row before this change.
4. Observe that the process identity can be null and recovery cannot
adopt the live child.
**Reproduced on**
- Paperclip `master` before this change.
- Linux source deployment.
- `codex_local` with ACPX `0.12.0`.
## What Changed
- Add an awaited `onAgentSpawn` lifecycle hook to the patched ACPX
runtime.
- Forward the ACP child PID and start time through the adapter `onSpawn`
callback.
- Keep a mutable callback sink for cached runtimes so a later respawn
updates the current heartbeat.
- Reuse the last known process identity when a compatible warm heartbeat
reuses the existing child.
- Kill the ACP child and fail session startup when the PID is invalid or
identity persistence rejects.
- Add ACPX and heartbeat recovery tests for callback ordering, warm
reuse, failure cleanup, durable row identity, and hot-restart adoption.
- Document the one-time drain required when an installed pre-fix run
already lacks process metadata.
## Verification
-
`PAPERCLIP_HOME="$PAPERCLIP_RUN_SCRATCH_DIR/test-home-execute-escalated"
pnpm exec vitest run
packages/adapter-utils/src/acpx-engine/execute.test.ts` — 89 passed.
-
`PAPERCLIP_HOME="$PAPERCLIP_RUN_SCRATCH_DIR/test-home-recovery-escalated"
pnpm exec vitest run
server/src/__tests__/heartbeat-process-recovery.test.ts` — 92 passed.
-
`PAPERCLIP_HOME="$PAPERCLIP_RUN_SCRATCH_DIR/test-home-remote-smoke-escalated"
pnpm exec vitest run
packages/adapter-utils/src/acpx-engine/remote-spawn-smoke.test.ts` — 3
passed.
-
`PAPERCLIP_HOME="$PAPERCLIP_RUN_SCRATCH_DIR/test-home-ci-repro-escalated"
pnpm exec vitest run
server/src/__tests__/heartbeat-dependency-scheduling.test.ts` — 6
passed.
- `pnpm --filter @paperclipai/adapter-utils typecheck` — passed.
- Reverse and forward dry-run application of `patches/acpx@0.12.0.patch`
— passed.
- `git diff --check` — passed.
- `git diff --exit-code origin/master...HEAD -- pnpm-lock.yaml` —
passed.
- `git diff --exit-code origin/master...HEAD -- .github/workflows` —
passed.
## Risks
- Runtime risk is low to moderate. ACPX now awaits process-identity
persistence during child startup.
- ACPX kills the child when persistence fails. This prevents an
unsupervised process, but it makes that heartbeat fail visibly.
- A compatible warm heartbeat reuses the identity of the existing ACP
child. Regression tests verify that identity is persisted before the
next prompt.
- The change updates the vendored ACPX patch. Package installation must
apply that patch.
- There are no schema, migration, public API, UI, workflow, or lockfile
changes.
> 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 used GPT-5.3-Codex for the earlier implementation.
- OpenAI Codex used GPT-5 for the lifecycle-hook revision and the
current fail-closed review fix. The runtime did not expose a more
specific snapshot ID or context-window size. Both runs used reasoning,
repository tools, and code 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 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
- [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
> - Local Claude, Codex, Gemini, and custom ACP adapters run through the
shared embedded ACPX engine
> - That engine wrapped every local agent command in a generated Bash
script to inject environment variables and filter child stderr
> - Windows cannot directly spawn that Bash wrapper, and npm/pnpm ACP
binaries are exposed through `.cmd` shims there
> - ACPX 0.12 already supports per-session child environment variables,
so the wrapper is unnecessary
> - This pull request registers agent commands directly, injects env
through ACPX session options, captures child stderr in-process, and adds
a real Node ACP spawn smoke on Ubuntu and Windows
> - The benefit is one cross-platform spawn path with a reusable smoke
test instead of parallel shell-wrapper implementations
## Linked Issues or Issue Description
Fixes#9941.
Refs #9428 and #9771.
**What happened**
ACPX-backed local agents failed to start on Windows because Paperclip
registered a generated POSIX `.sh` wrapper as the agent command. Windows
also needs the `.cmd` npm/pnpm shim when resolving built-in ACP
binaries, and symlink creation can fail with `EPERM` for seeded
auth/skill files.
**Expected behavior**
The same ACPX engine path should spawn a real ACP agent on Windows and
Linux, forward Paperclip/runtime env without mutating `process.env`,
preserve filtered/unfiltered child stderr behavior, and fall back to
copies where Windows symlinks are unavailable.
**Steps to reproduce**
Run a local ACPX adapter on Windows with the prior wrapper path. ACPX
attempts to spawn the generated `.sh` file and the agent never
initializes.
**Deployment mode**
Local Paperclip adapters using
`packages/adapter-utils/src/acpx-engine/`.
## What Changed
- Removed generated Bash agent/env wrappers and registered local
commands directly with ACPX.
- Passed the resolved child environment through ACPX
`sessionOptions.env`, including resume retry paths.
- Added a minimal `acpx@0.12.0` package patch exposing child stderr
callbacks and allowing documented uppercase env-map keys in persisted
session options.
- Moved stderr tee/filter behavior in-process: raw stderr remains in the
per-run file while benign `nes/close` noise is omitted from live stderr.
- Preferred `.cmd` ancestor binaries on Windows and added `EPERM` copy
fallbacks for Codex auth seeding and Gemini skill materialization.
- Added a real Node ACP echo-agent spawn smoke that can run directly on
any supported platform.
## Verification
- `pnpm exec vitest run
packages/adapter-utils/src/acpx-engine/execute.test.ts
packages/adapter-utils/src/acpx-engine/spawn-smoke.test.ts` — 57 passed.
- `pnpm --filter @paperclipai/adapter-utils typecheck` — passed.
- `node --test scripts/acpx-patch-packaging.test.mjs
scripts/release-lib.test.mjs` — 10 passed.
- Full canary release dry run under Node 24.18.0 / npm 11.16.0 — passed
in an isolated scratch clone.
- `git diff --check` — passed during implementation verification.
- One-time GitHub Actions proof: [Ubuntu ACPX spawn
smoke](https://github.com/paperclipai/paperclip/actions/runs/29924348927/job/88937774579),
[Windows ACPX spawn
smoke](https://github.com/paperclipai/paperclip/actions/runs/29924348927/job/88937774558),
and [Canary Dry
Run](https://github.com/paperclipai/paperclip/actions/runs/29924348927/job/88937774497)
passed on head `f345ac69f2`; the dedicated smoke jobs are intentionally
not retained in the recurring PR workflow.
## Risks
- The ACPX stderr callback and env persistence exemption are carried as
a pnpm dependency patch until ACPX exposes/fixes those behaviors
upstream.
- Child stderr is synchronously appended to preserve ordering and
failure diagnostics; unusually high-volume agent stderr could briefly
block the Node event loop.
- The Windows-specific `.cmd` resolution and symlink `EPERM` branches
are proven by the standalone smoke test and the linked one-time
`windows-latest` run rather than a permanent CI gate.
> 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-5.4 via Codex CLI, medium reasoning, repository/tool
execution enabled; context-window size is 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 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
- [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>
Rebuild the patch file with valid unified-diff hunks so pnpm can
apply the locale and environment fixes during install.
Co-Authored-By: Paperclip <noreply@paperclip.ing>
The vendor package still hardcoded LC_MESSAGES to en_US.UTF-8.
That locale is missing in slim containers, and initdb fails during
bootstrap even when --lc-messages=C is passed later.
Co-Authored-By: Paperclip <noreply@paperclip.ing>
Use globalThis.process.env in the vendor patch so the spawned child
process config does not trip over the local process binding inside
embedded-postgres.
Co-Authored-By: Paperclip <noreply@paperclip.ing>
The embedded-postgres library hardcodes --lc-messages=en_US.UTF-8 and
strips the parent process environment when spawning initdb/postgres.
In slim Docker images (e.g. node:20-bookworm-slim), the en_US.UTF-8
locale isn't installed, causing initdb to exit with code 1.
Two fixes applied:
1. Add --lc-messages=C to all initdbFlags arrays (overrides the
library's hardcoded locale since our flags come after in the spread)
2. pnpm patch on embedded-postgres to preserve process.env in spawn
calls, preventing loss of PATH, LD_LIBRARY_PATH, and other vars
Co-Authored-By: Paperclip <noreply@paperclip.ing>
Nicky Leach
Dotta
Michael Nguyen
Devin Foley
dotta