## Thinking Path
> - Paperclip is the open source app people use to manage AI agents for
work.
> - The native runner connects task messages, provider execution, and
task outcomes.
> - First-time user tests exposed gaps in recovery, completion
permissions, message delivery, and Stop behavior.
> - These gaps left usable output hidden, completed work waiting for
bookkeeping, or safe work unable to continue.
> - This pull request fixes the shared lifecycle and receipt paths while
preserving process ownership and action checks.
> - Users can continue work with accurate task state and durable
messages.
## Linked Issues or Issue Description
**What happened?**
A stopped local Codex execution could remain blocked even after its
processes had stopped and its complete transcript proved that no
external action needed replay. Claude under Conservative permissions
could fail to call task completion tools. Recovery could reuse an
assistant item ID and overwrite prior output. A delivered comment could
remain marked uncertain after navigation. Stop could look like Pause or
a new recovery incident. Workspace contention could look like
cancellation. A direct reply reopening Done could enter a clarification
loop.
**Expected behavior**
Recover automatically only with verified termination and complete action
receipts. Preserve answers and messages. Keep task completion available
under Conservative permissions without broad tool access. Show crashes
as Blocked, actual human decisions as In Review, and ordinary workspace
contention as waiting. Stop the current response and allow a new
direction.
**Steps to reproduce**
1. Create ordinary response tasks with local Codex and Claude Code, then
send follow-up messages through the task composer.
2. Interrupt a disposable local Codex runner during text-only work.
Verify automatic continuation and retained output.
3. Stop a response, send a new request, answer a clarification, and
reopen completed work with another message.
4. Navigate or reload while a comment submission is pending. Confirm the
exact persisted request receipt settles it without removing newer draft
text.
5. Run two tasks in a shared Daytona workspace. Confirm waiting does not
appear as failure.
**Paperclip version or commit**
Initial acceptance baseline: `c9021c6721f91e2c74bd9fee9d3fd41c999d17b7`.
Current integration base: `6cef9743c`. Both operator-interruption and
workspace-waiting guards are preserved; native restart and legacy
permission rules remain documented.
**Deployment mode**
An isolated source-built test-drive instance, with real local Codex and
Claude Code providers and disposable Daytona environments.
Related work: #13314, #13316, #13327, #13344, #13239, #13254, #13163.
This PR addresses additional failures from ordinary task journeys,
including controller restart handoff and repeated warm sandbox setup.
Historical task status reconciliation is excluded.
## What Changed
- Persist runner ownership immediately at spawn and resume an explicitly
adopted runner even when the controller crashed before the first driver
checkpoint. Detach the controller safely across graceful restarts,
including session startup. Prevent an old finalizer from suspending or
signaling an adopted runner. Checkpoint idle warm sessions before
shutdown. Preserve the same run and queued follow-up messages.
- Scope saved legacy queue successor checks to the queue owner while
preserving ordinary task locks, operator identity, assignment gates, and
exactly-once delivery.
- Preserve managed Codex credential files when an old session is
detached for restart; normal owned cleanup still copies refreshed auth
back and removes the scoped copy.
- Reuse the bound warm shared sandbox and fully verify an existing
staged provider pack before using it. This avoids repeated uploads when
the pack is already valid.
- Add a narrow local Codex replacement path with stopped-process proof,
a closed transcript inventory, exact completion receipts, and
fresh-session lineage. Preserve no-replay holds when evidence is
incomplete. Recovery may clear only the same run's recorded Blocked
status version; manual re-blocking and dependency changes invalidate
that receipt, while queued comments do not. Later blocks stop scheduled,
queued, and final dispatch; queued/final checks re-read dependencies
even when the task status stays In Progress.
- Permit only task delivery and human-input tools through the isolated
Claude runner's exact task bridge.
- Scope assistant item identity to the provider turn and ignore only
authority-free Codex skill-change notifications during startup.
- Reconcile composer submissions by client request ID across response
loss, navigation, and reload. Retain text typed during delivery.
- Keep acknowledged run-only Stop neutral and show workspace contention
as waiting. Project exhausted native failures as Blocked.
- Restore the guarded task-page retry action for failed legacy runs,
including the server-supported explicit new-attempt path for stopped
conversation adapters. Preserve native/process recovery holds and avoid
promising Retry while a decision or execution gate hides it.
- Refresh delivered artifacts and handle direct user replies that reopen
completed work without a clarification loop.
- Check the embedded PostgreSQL PID, data directory, and actual port
before connecting or migrating.
- Document accepted behavior and add focused regressions at lifecycle,
route, transcript, and UI boundaries.
## Verification
- Final head `fece606ac2` passes the complete GitHub CI matrix: **34
green checks, two expected Storybook skips, no failures or pending
checks**, including `ci / verify`, `ci / e2e`, full runner verification,
typecheck, build, every server/workspace shard, and all browser shards.
[CI
run](https://github.com/paperclipai/paperclip/actions/runs/34727183287).
Greptile is **5/5 with no open findings**. The final two commits only
refine test fixtures; both affected suites pass 24/24 locally and in CI,
with server typecheck green.
- Complete local Vitest coverage uses the canonical groups/shards: all
635 general server suites, all 145 serialized suites, and all workspace
packages. The aggregate began on `0a8001c18` while the final queue fix
arrived: 23,903 passed, five failed, 87 skipped. The five
port/socket/timing failures passed unchanged in follow-ups (60 tests in
the exposure/file suites and 412 tests covering the serialized failures
and unrun tails). The final queue/operator-identity suites separately
passed 52/52. This is aggregate coverage plus explicit reruns, not a
pristine single-command final-head run.
- After integration with current master,
queue/operator-identity/continuation suites passed 162/162 and affected
UI suites passed 140/140. ACP Stop/continuation and legacy
task/Inbox/message browser suites passed 9/9, including both task
recovery Retry and thread Try again, automatic saved-message delivery,
exactly one new run, Done, and retained output after reload. The default
process Stop/Pause/Resume browser case passed (the native-provider case
is opt-in and skipped by default). The complete Board attachment/receipt
browser suite passed 11/11 on a disposable instance, covering both
composers, exact receipts after lost responses, no replay, bound
attachments, and newer drafts after reload.
- Blocking-intent regressions cover pre-existing Blocked, a mismatched
run/cause, an explicit manual re-block, changed dependencies, a queued
comment after failure, and a block arriving between scheduling and
provider dispatch. The negative cases reproduced before the fix. All 478
affected executor/recovery/dispatch tests passed; both database suites
ran separately after availability-probe skips in the first combined
command. The final late-dependency check passed all 143 affected
recovery/dispatch tests (zero skips) after two new negative cases
reproduced the bug.
- Focused runtime regressions cover awaited runner ownership
publication, authenticated adoption before the first checkpoint,
old-finalizer detachment, idle and busy warm-session shutdown, rejected
checkpoint propagation, provider-pack verification, and managed-Codex
credential preservation. Four managed credential detachment cases
reproduced the bug before the fix; normal owned cleanup still succeeds
exactly once.
- Live local Claude: SIGKILL 2.6 seconds into startup recovered the same
run automatically in 53 seconds, then a normal follow-up completed in 24
seconds. SIGTERM 2.5 seconds into startup preserved the same run (54
seconds) and its queued follow-up (21 seconds). Answers remained visible
and the task reached Done.
- Live Claude Daytona: a warm follow-up retained its sandbox and fell
from 121 seconds to 44 seconds. A separate cold turn took 127 seconds;
after controller shutdown and checkpointing, its follow-up completed in
33 seconds with the same sandbox, workspace, native session, and runner.
Both answers remained visible and the task was Done.
- Other live journeys covered task completion and follow-up with local
and Daytona Codex, local Codex crash recovery, Stop then new direction,
clarification response, live artifact refresh, and shared-workspace
waiting.
- Validation limits: the opt-in native composer Stop/Pause→subtree
Resume fixture exposes terminal/result ordering and subtree-cancellation
attribution bugs that can leave a child task blocked; that new finding
is assigned to a separate follow-up and is not claimed fixed here.
Default CI skips this optional native-provider fixture. Managed-Codex
credential handoff and the queue-agent integration use automated
regression evidence. Cold custom provider-pack uploads still add startup
latency.
## Risks
- Automatic replacement remains deliberately narrow: local Codex,
verified stopped identities, unchanged retained state, and a complete
text/completion-only turn. Unknown actions, partial history, or changed
ownership remain blocked.
- Claude completion permission handling changes an upstream package
patch. The exact isolated task bridge must remain pinned; unrelated
tools keep their existing permissions.
- New task failure projection changes user-visible status. No historical
status backfill or database migration is included.
- This is a broad lifecycle fix across server and UI. Live proof covers
graceful local Claude restart during startup and idle Claude Daytona
session recovery across controller shutdown. Live abrupt SIGKILL during
local Claude startup also recovered the same run. Unknown ownership or
missing action evidence still blocks reuse. Cold custom provider-pack
uploads still add startup latency; this change avoids unnecessary repeat
uploads.
## Model Used
OpenAI GPT-6 (Codex), with reasoning, code execution, browser
automation, and tool use. The exact hosted model ID and context window
are not exposed in 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 (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.
> - Task conversations save messages that arrive during an active turn.
> - Legacy adapters deliver these messages in a later turn.
> - A run can stop before the saved queue is delivered.
> - The Interrupt button previously required an active run, so it could
not release this queue.
> - This pull request lets a board operator send the saved queue after
the run stops and retries queues missed during finalization.
> - Manual dispatch must use the clicking operator and must not require
permission to create agents.
> - The task can continue without a duplicate message or a second
execution owner.
## Linked Issues or Issue Description
**What happened?**
A legacy task retained a queued message after its run stopped. Interrupt
was disabled because the queue had no active target. Finalization and
deferred message admission can also leave a queue without a successor.
**Expected behavior**
Interrupt sends the saved messages when no runner is active. Messages
that arrive during normal completion are delivered automatically. An
uncertain previous execution still requires proof that its process or
sandbox stopped.
**Steps to reproduce**
1. Queue a user message during a legacy conversation turn.
2. Let the turn stop or simulate a server restart before queue
promotion.
3. Open the task with a deferred queue and no active run.
4. Try Interrupt. Before this change, the button is disabled.
**Paperclip version or commit**
Reproduced against `8f40b4ad4`.
**Deployment mode**
Legacy conversation adapter. The same persisted queue state is covered
with an isolated PostgreSQL fixture.
Related public work: #13275 adds active legacy interruption. #13291
addresses automatic sandbox conversation recovery. This change handles
explicit saved-queue delivery and late queue promotion.
## What Changed
- Accept a null Interrupt target while retaining queue identity,
revision, company, and assignee checks.
- Save the operator's request on the existing queue. Reuse normal
admission after verified stop, including older messages, different
authors, and queues whose original wake came from the system.
- Strip interruption authority from caller-supplied wake payloads. Only
the board queue route can persist that authority.
- Retry durable interruption requests after restart and deferred queues
after legacy cleanup.
- Let an explicit Interrupt retry cleanup for its stopped run, including
old ephemeral leases that recorded success without a provider stop
receipt. Preserve retained resources, other lease owners, and the
automatic retry limit.
- Preserve the server's waiting explanation when normalizing and
combining queue entries.
- Revalidate the consumed board queue receipt at dispatch so a different
message author does not cause setup failure.
- Use the Interrupt user's execution identity for the new run. Preserve
original message authors. Validate the receipt independently at startup
and inherit the resulting identity on retry.
- Persist authenticated board authority for ordinary manual wakes too.
Adopting someone else's queued messages cannot switch a manual run to
that author's permissions. Strip caller-supplied authority markers and
retain private conversation ownership checks.
- Keep the clicking user when a manual wake is merged into an older
deferred receipt. Update its requester and payload in the same
transaction.
- Use the same current-queue/revision API on task details and pipeline
conversations; show Interrupt after a legacy target stops.
- Keep manual wakes out of active runs, including unscoped agent wakes.
They receive their own execution identity; a matching receipt requester
is not sufficient because an exact retry can retain a different
originating identity.
- Authorize both existing-agent wake endpoints with `agent:wake`,
available to active non-viewer company members. Keep `agents:create` for
hiring. Validate the stored task and current assignee before an exact
task retry.
- Reject viewer Interrupt requests before saving intent or stopping
execution. Keep external chat retry authorization and per-action
agent/user permission checks.
- Preserve edits and discards until dispatch. Prevent another queue
promotion when the same agent already has a successor. Keep independent
reviewer recovery available.
- Suppress cancelled/failed run toasts for intentional operator
interruption. Keep ordinary runtime error notices.
- Add UI, route, admission, restart, successor ownership, and toast
regression tests. Document the behavior.
- Reuse the existing socket reservation helper for both
credential-quorum test cases after CI exposed an ambient-port collision.
This changes test preparation only; production credential staging is
still called exactly once.
## Verification
- Failing regression tests reproduced the message-author identity bug
and an operator's `agents:create` rejection before the fixes.
- All 316 focused tests pass across eight route, queue, identity,
authorization, continuation, and responsible-user suites, including the
44-test rerun of queue admission and actual startup after the final
manual-wake restriction. Regressions reproduce cross-user merging both
with and without a task, and same-requester receipt ambiguity. The
cross-company existence guard also passes both tests.
- Startup integration tests reach adapter execution under the clicking
operator and retain that identity through follow-up. Coverage includes
mixed authors, adopted queues, system-origin queues, restarts, forged or
stale receipts, viewers, suspended memberships, changed assignees,
private conversations, and caller-supplied authority markers.
- The earlier queue/cleanup/UI regression suite passed 402 tests. The
final review corrections pass another 180 tests across queue
admission/persistence, real heartbeat startup, UI API, conversation
rendering, and pipeline suites. Regression tests reproduced both review
findings before correction. The final head has a 5/5 review with no
unresolved threads. Full CI passes on `c2002979c`, including every
general and serialized server shard, all browser shards, Paperclip
Runner verification, typecheck, build, canary dry run, and the aggregate
gates.
- Full `pnpm -r typecheck`, `pnpm build`, and UI token gates pass after
the final application changes. CI identified an outdated task-page API
mock after the shared helper extraction; the fixture now exercises the
real helper, and all 131 task-page/API tests pass. The final application
build passes with the additional manual-wake restriction.
- CI exposed a pre-existing port collision in the Codex
credential-quorum fixture. It reproduced locally; both listener cases
now use the existing bounded reservation helper. All 41 credential tests
pass on rerun. One intervening local run hit a separate ambient bind
collision in the two-occupied-port case.
- The full local `pnpm test:run` attempt was stopped after host
contention caused focused-suite timeouts. The affected focused tests
passed on rerun. An expiring trace fixture and a missing
private-conversation state were corrected. The successful full CI run is
the complete-suite verification.
- Hosted Interrupt previously cleared the original queue and produced
exactly one successor with neutral interruption feedback. It exposed the
dispatch authorization defect. Retry on that earlier build was rejected
for missing `agents:create` before creating another run.
- Deployed the final application build (`38257f391`) to the scoped
hosted instance and verified readiness. The latest PR commit changes
only the credential test fixture; application code matches that
deployment. A live Retry by the same operator without `agents:create`
created one successor attributed to that operator, passing the former
dispatch permission gate. Startup then stopped at
`configuration_incomplete` because that operator has not configured
their required personal Claude Code OAuth secret; the post-deployment
run page confirms the operator identity and no provider work started,
and the My secrets UI still shows the token as not set. Provider
execution remains unverified pending that credential. No permission
grants or credentials were changed.
## Risks
Queue admission and finalization can race. The task lock, durable queue
receipt, current comment IDs, and successor guard prevent duplicate
dispatch. Process and lease stop checks, task pauses, approvals,
ownership, and budgets remain in force. The API change only allows null
on legacy Interrupt; native steering still requires an active run. No
schema migration is required. Active non-viewer board members can now
invoke existing agents without agent-creation permission. Agent
self-invocation rules, raw provider-trace admin access, task retry
scope, external chat authorization, and action-specific user/agent
permissions remain enforced.
## Model Used
OpenAI GPT-6 through Codex. The session does not expose an exact backend
model ID or context-window size. Used reasoning, repository search, code
execution, tests, and browser tools.
## 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
#` 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.
> - Native runs report their outcome through paperclip_finish.
> - The server previously turned incomplete reports into human approval
requests.
> - Those requests could block a later successful run, even when no
person had requested review.
> - This pull request creates review cards only for explicit attention
requests and withdraws proven old fallback cards.
> - Agents receive useful completion feedback, while explicit approval
gates and task state protections remain in force.
## Linked Issues or Issue Description
Related: #13266 removed reviews caused by policy upgrades. This change
removes the separate completion fallback.
**What happened?**
An agent reported needs_review while waiting for checks without
requesting a human decision. Paperclip created a generic Native
completion review. A later successful report could not complete the task
because that old card remained pending.
**Expected behavior**
Ordinary low-risk work completes after a valid done report, a successful
run, and workspace finalization. Incomplete work stays with the agent.
Explicit approval requests remain visible and must be resolved.
**Steps to reproduce**
1. Complete a native run with needs_review and no attention requests.
2. Continue the task and submit a successful done report.
3. Observe that the old implementation leaves the task in review behind
a generic confirmation card.
## What Changed
- Require explicit attention requests to create native review cards.
Route each request independently and preserve pending or declined
decisions.
- Withdraw only pending system cards with matching old decision,
assessment, effect, contract, and prompt provenance. Preserve history
and explicit or answered requests.
- Reassess an affected current result without overwriting later task
edits, runs, contracts, or workspace failures.
- Return pending approval links and required actions through the
completion tool. Reject contradictory done reports and empty review
requests before accepting a result.
- Allow one corrective continuation for incomplete results, then expose
a recovery action.
- Update status fixtures, database regressions, runner tests, and the
completion contract documentation.
## Verification
- `pnpm -r typecheck` passed after merging current master.
- `pnpm build` passed after merging current master.
- The combined branch passed 89 completion and Agent Chat tests. Other
targeted tests passed: 170 external-chat and reconciliation tests; 50
runner-resume and control-plane tests; 13 arbiter tests; 7 chat delivery
tests; 21 runner completion and runtime-context tests.
- The full local test attempt exposed old review fixtures and a missing
fake-provider binary. The fixtures are fixed and the helper is built.
All affected suites pass in fresh reruns. The timing-sensitive Discord
test also passed on rerun.
- All latest-head CI checks passed, including build, typecheck, general
and serialized tests, runner verification, browser tests, and canary dry
run. Greptile is 5/5 with zero unresolved comments.
## Risks
- Cleanup changes existing pending cards. It requires exact system
provenance and only applies to low-risk agent-claim contracts. It does
not delete history or dismiss explicit requests.
- Status still commits after the turn and workspace finalization.
Completion feedback reports current constraints and does not claim an
early status commit.
- Incomplete reports now request a bounded corrective run instead of an
automatic approval. Repeated failures expose recovery.
## Model Used
OpenAI Codex, GPT-6, with repository inspection, code execution, and
test tools. The exact deployment 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.
> - The first agent helps the board define work and hire other agents.
> - That agent can have the general role while its instructions require
hiring skills.
> - Missing skills and blocked schema discovery make valid requests
fail.
> - Repeated confirmation and invalid waiting guidance can turn these
failures into extra runs.
> - This PR supplies the required skills, opens read-only schema
discovery, and corrects the guidance.
> - The agent can complete an authorized hire while company approval and
duplicate checks still apply.
## Linked Issues or Issue Description
Refs #13068 — the first-task onboarding flow that this change repairs.
Refs #12029 — related drift between the sandbox allowlist and bundled
hiring guidance. This PR adds schema access; it does not replace the
earlier hiring-route fix.
**What happened?**
A general-role onboarding chief received hiring instructions without the
core hiring skills. Sandbox requests to the documented OpenAPI endpoint
failed. The agent then guessed question and hire payloads. The persona
required new confirmation after validation errors and described waiting
states that agents cannot set.
**Expected behavior**
A direct request authorizes the requested hire. The chief asks only for
material missing details, uses valid API payloads, and completes the
task. Formal company approval gates still apply. A saved human-input
card gives the task a valid waiting state.
**Steps to reproduce**
1. Create an onboarding chief with role `general` through the board.
2. Ask it to hire a friendly robot with a supplied name and
responsibilities.
3. Check its assigned skills, schema requests, question cards, hire
requests, and final task state.
**Paperclip version or commit**
Reproduced on the first-task onboarding implementation after #13068. The
live local verification used this branch at `112f44610`.
**Deployment mode**
The original failure used a hosted sandbox with legacy Codex ACP. Live
verification used an isolated local instance and real `codex_local`
execution. Queue and HTTP/2 transport access is covered by automated
tests.
## What Changed
- Give board-created onboarding chiefs the existing core skills
regardless of role. Preserve explicit skill version pins, including
aliases. Keep ordinary general-agent defaults and authorization checks.
- Allow exactly `GET /api/openapi.json` through both sandbox bridge
transports.
- Publish validator-tested question, free-text, hire, and waiting
examples. Regenerate the runner API reference and capability inventory.
- Clarify direct authorization, material ambiguity, and correction of
confirmed pre-creation validation failures. Preserve uncertain-outcome
reconciliation, duplicate protection, and company approval gates.
- Align disposition instructions with agent permissions and the saved
human-input waiting path.
## Verification
- After rebasing onto current `master`: 69 targeted server tests, 110
queue/HTTP2 bridge tests, and 4 capability inventory tests passed. These
cover core skill defaults, version pins, actor restrictions, schema
access, published examples, hire validation, idempotency, and approval
gates. Waiting recovery tests and live question flows also passed before
the rebase.
- `pnpm -r typecheck` and `pnpm build` passed again after the rebase.
Frozen dependency installation and both generated capability checks
passed.
- Ran the full `pnpm test:run` suite. The initial run had 14 failed
server files due to local database resource limits, a missing built test
fixture, and socket failures. All 14 files passed after fixture repair
and isolated retries. UI, CLI, workspace packages, database tests, and
all 145 serialized server files passed.
- Real one-request hiring replay: one hire, one successful run, task
done in 2m16s. No repeated approval or recovery escalation.
- Real two-turn browser conversation: start with an unspecified hire,
then supply a name and friendly robot responsibilities. One
clarification card, one hire, two successful runs, task done in 3m27s of
execution. No failed writes, confirmation cards, or recovery actions.
- Assigned the hired robot a welcome-message task through the browser.
It produced a warm message under 100 words and finished in one
successful 66-second run, with no questions or recovery actions.
- The two-turn flow still asked an optional preferences question and
gave a technical final reply. These are remaining presentation limits.
- Greptile: 5/5 on `b71f83ba2`, with zero unresolved review threads.
Fixed its generator finding and passed 1,655 published-example/runtime
API tests plus server typecheck. All latest-head CI checks are green (32
passed; 2 unrelated Storybook checks skipped). The signoff-policy
browser test initially timed out while waiting for an approver run. Its
shard passed on one rerun without code changes. [CI
run](https://github.com/paperclipai/paperclip/actions/runs/34698211049).
## Risks
- Onboarding chiefs receive more default skills. Ordinary general agents
retain existing defaults, and explicit versions take precedence.
- Prompt guidance can affect model behavior. The live replays are
examples, not a guarantee that every model follows the guidance.
- Retry guidance applies only when validation confirms that nothing was
created. Uncertain outcomes still require checking existing agents.
- No database migration or new public endpoint. Existing company
boundaries, approval gates, and bounded recovery remain in force.
## Model Used
OpenAI Codex, model `gpt-6-astra`, with reasoning, tool use, code
editing, and live browser verification. The exact 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
- [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.
> - Conversations must use the same tasks, controls, and execution
history.
> - Users need an ongoing chat with an agent without managing task
properties.
> - Agents should clarify and plan work, then hand execution to assigned
project tasks.
> - This pull request combines the reviewed Agent Chat stack for one
squash merge.
> - The benefit is persistent conversation with normal task governance
and shared UI.
## Linked Issues or Issue Description
**Subsystem affected**
Task lifecycle, agent runtime tools, shared task UI, and browser/paid
runner tests.
**Problem or motivation**
Users need one persistent conversation with each agent. A separate chat
store or renderer would duplicate task behavior and bypass existing
controls.
**Proposed solution**
Use a task-backed chat per company, user, and agent. Reuse the task
composer and transcript. Clarify and plan in chat, then create assigned
project tasks with the relevant plan. Keep Agent Chat behind its own
disabled-by-default experimental setting.
**Roadmap alignment**
This implements the task-backed direction in [CEO
Chat](https://github.com/paperclipai/paperclip/blob/master/ROADMAP.md#-ceo-chat).
Related proposals: #2504 and #9693. Related request: #7981. The
maintainer requested one squash merge of the complete stack.
Consolidates the reviewed runtime
[#13281](https://github.com/paperclipai/paperclip/pull/13281), backend
[#13282](https://github.com/paperclipai/paperclip/pull/13282), and UI
[#13283](https://github.com/paperclipai/paperclip/pull/13283) layers
with this PR's E2E coverage. All four layers passed CI and received
Greptile 5/5 before consolidation. This PR targets master and includes
the complete feature.
## What Changed
- Add personal canonical chat tasks with ordinary company visibility,
immutable identity, idempotent first sends, and an idle waiting state.
- Process `/new` in queue order. Preserve history, release a chat pause,
and fence old provider context and delayed writes.
- Keep chat lifecycle rules across recovery, finalization, assignment,
task lists, and rollups.
- Support research and plan revision in chat. Hand plans to ordinary
assigned project tasks before execution starts. Reject new chat
subtasks.
- Add repository-aware project creation and discovery tools, including
multiple repository IDs and GitHub URLs, authorization, idempotency, and
durable project-created cards.
- Reuse task UI components for chat, with starred/recent agent
navigation and a separate `enableAgentChat` experimental flag.
- Add deterministic browser tests and 24 paid chat cells across four
Codex/Claude profiles, with validated reports and screenshots.
- Integrate current master recovery, controller lease, queued-message,
and task UI changes. Gate chat interruption and deferred promotion on
ownership/feature policy. Guarantee lease renewal and active controls
are stopped even if teardown fails.
- Preserve master's migration 0273 and generate chat migration 0274 with
idempotent replay for development databases.
## Verification
- Prior exact heads of all four PRs passed Linux CI, including build,
typecheck, general/serialized tests, and browser E2E. Each had Greptile
5/5 and no unresolved findings.
- Integrated local verification passed: full repository typecheck and
production build, Storybook build, token gates, 340 focused UI tests,
all 20 deterministic chat browser tests, two migration replay tests, 88
focused chat/queue/native/controller tests, and provider/session
regressions including real lease expiry. These include the three
lifecycle regressions for the final admission/teardown fixes; server
typecheck also passes. Current head
`1268eda16cc2af892055917e7292f068820be135` has Greptile 5/5 with no
unresolved findings and passing security scans. All final-head CI gates
passed: build, full Runner verification, typecheck/release registry,
canary, all general/serialized test shards, and all browser E2E shards
([CI
run](https://github.com/paperclipai/paperclip/actions/runs/34696739927)).
Local PostgreSQL startup contention required serialized retries; skipped
fixtures do not count as passing coverage.
- The earlier paid campaign passed all 24 chat cells and retained 32
screenshots:
[report](https://d1p6rlowie26tp.cloudfront.net/runner-e2e/campaigns/gha-34648511170-1/index.html?report=agent-chat#suite-agent-chat).
It tested `abacbdfd2f660709ec37312cdb758284c8399d04`; it is prior
evidence, not a paid run of this integrated head.
- Manual check: enable Agent Chat in Experimental settings, open an
agent, clarify and revise a plan, then hand off to an assigned project
task. Stop a reply, send `/new`, and verify fresh context with retained
history. Disable the setting and verify agent shortcuts/new chat turns
are blocked.
## Risks
- Queue/session integration can affect retries and delayed writes. Tests
cover ownership, cancellation, reset boundaries, idle recovery, and
ordinary task behavior.
- Migration 0274 adds conversation fields and constraints. Replay is
idempotent and preserves existing development chat history.
- This combines the previously reviewed stack at the maintainer's
request. Agent Chat remains off by default and is separate from
Conference Room.
## Model Used
OpenAI Codex, GPT-6 Astra (`gpt-6-astra`), with reasoning, code
execution, browser tools, and parallel review. The exact context-window
size is not exposed in this session. Codex and Claude also ran as test
subjects in the linked paid campaign.
## 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.
> - Native runner protocol tests verify that invalid authenticated input
fails closed.
> - Cloud readiness requires those tests to pass before a source commit
can deploy.
> - Two test fixtures hash the host Node executable twice even though
its process launcher is synthetic.
> - Cold reads of a large Linux executable consume time unrelated to
protocol failure handling.
> - This PR gives both synthetic runner fixtures tiny real artifacts and
keeps their existing deadlines.
> - Real authentication, encrypted frames, and request and notification
failures remain covered.
## Linked Issues or Issue Description
**What happened?**
Cloud readiness repeatedly failed in `DurablePrpControlPlane > promptly
fails the real transport request and notification paths on authenticated
bad semantic input (throwing observer: false)` with `Test timed out in
5000ms`. The image built, but the failed source check prevented
deployment. See runs
[34656885170](https://github.com/paperclipai/paperclip/actions/runs/34656885170)
and
[34657111560](https://github.com/paperclipai/paperclip/actions/runs/34657111560).
The first case took 5.6–11.4 seconds; the following case took about 0.4
seconds. The original test reads and hashes `process.execPath` once
itself and once through the real transport. The Node executable is
122,678,944 bytes in the local Linux Node 24 container, versus 68,672
bytes on the development Mac. Instrumented Mac runs pass; cold
executable reads are a likely cause of the CI-only timeout.
**Expected behavior**
The deadline should measure real protocol rejection and consumer
failure, without reading a large unrelated executable as test fixture
data.
**Steps to reproduce**
1. Run the named test with the real transport and synthetic process
launcher.
2. For a deterministic probe, inject a six-second delay into the first
`readFileSync(process.execPath)` call.
3. The original test exceeds its existing five-second deadline. Both
cases pass after this change because neither reads the host executable.
The probe is temporary instrumentation, not part of this commit.
**Paperclip version or commit**
f12b647ae; the same failure also occurred on a38ccf997 and 9031516a7.
**Deployment mode**
Cloud readiness in GitHub Actions on Linux AWS runners. No duplicate
open PR found after searching runner timeout and durable PRP changes.
## What Changed
- Write a 26-byte synthetic runner artifact inside each existing
temporary test directory.
- Bind both the real authority and test client to that artifact's actual
SHA-256 digest.
- Retain synthetic process launchers, real wire exchange, all observer
and composed-runtime cases, integrity assertions, cleanup, and existing
five- and fifteen-second timeouts.
## Verification
- Passed all 107 tests across
`src/control-plane/durable-prp-control-plane.test.ts` and
`src/drivers/codex/codex-protocol-integrity.test.ts` using the package
Vitest configuration.
- With a temporary six-second host-executable read delay, the original
case times out. Both fixed cases pass in 235 ms combined under the same
fault injection.
- Passed `git diff --check`.
- Full workspace typecheck and build passed. The full local test command
stopped after the general server group: 10,694 tests passed and 15
failed. Thirteen failures involve macOS directory rename permissions in
the skill-cache suites. Two failures involve a missing local AgentMail
skill path. Later groups did not run in that local invocation. The
targeted runner tests passed; all Linux CI test groups passed.
- [Current-head CI run
34661107658](https://github.com/paperclipai/paperclip/actions/runs/34661107658)
passed. All 32 checks are green or intentionally skipped at
`e7053594e647fd8e9b71fc569cab6b7a3e9cd210`. Native runner verification
passed in 15m25s. Greptile is 5/5 on that head, with no unresolved
review findings.
## Risks
Low risk: this changes test fixture data only. The launcher never
executes the file. Artifact hashing and authentication still use real
bytes. Protocol behavior and existing test deadlines do not change. The
current Linux CI run passed both fixtures and the full native runner
verification. Further post-merge runs will test repeatability.
## Model Used
OpenAI GPT-6 through Codex, with reasoning, repository tools, and code
execution. The exact serving model ID and context window are not exposed
by 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 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 (targeted runner tests;
full local suite limitations are listed 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 is the open source app people use to manage AI agents for
work.
> - Native sessions keep a provider process and its work alive across
control-plane operations.
> - A hidden 15-minute turn deadline stopped work even when the agent
timeout was zero.
> - A one-hour runner lifetime and fixed connection lease added two more
limits.
> - Recovery also rejected goal commands because it reconstructed their
startup summary with the wrong protocol version.
> - This pull request removes implicit duration limits and renews
authenticated leases in the harness.
> - Healthy sessions can continue without model action or a user
interface change.
## Linked Issues or Issue Description
Refs #13092 and #12845. Related: #13163 covers sandbox recovery after
app restarts; this change covers session duration and lease renewal.
**What happened?**
A native Codex session stopped after 15 minutes while a tool was still
running. The agent had `timeoutSec: 0`. Recovery then rejected a
`session.goal.get` startup command with `invalid provider startup
ownership fence`.
**Expected behavior**
An unlimited session keeps working while its provider and authenticated
controller remain healthy. Lease maintenance is transparent. Explicit
timeouts, cancellation and revoked authority still take effect.
**Steps to reproduce**
Start a native session with `timeoutSec: 0` and run a tool beyond 15
minutes. Before this fix, the runtime cancels the turn. A recovery
startup that uses a goal command also exposes the protocol-version
mismatch.
## What Changed
- Honor the agent turn timeout. Zero disables the timer. Long explicit
durations use timer chunks to avoid Node timer overflow.
- Default native runner lifetime to unlimited. Keep bounded startup,
reconnect and control-operation deadlines.
- Renew leases over the authenticated connection. Persist renewal before
the reply. Validate identity, epoch and expiry. Handle duplicate
requests and a lost reply on reconnect.
- Freeze renewal during warm ownership transitions and terminal
handling.
- Validate persisted goal startup commands with protocol v2.
- Add duration, renewal, ownership, recovery and real-process regression
tests. Update runner protocol and recovery docs.
- Add no UI components or controls. Renewal requires no model output or
user action.
## Verification
Current head: `348e369c35c5da8bb8be378f4b35dcf6f40882e7`. [CI
run](https://github.com/paperclipai/paperclip/actions/runs/34649767113).
- All 32 checks pass on this head. The two Storybook checks are skipped
as expected. CI includes full build, typecheck, runner verification,
browser suites, server tests and the canary package dry run.
- Greptile reports 5/5 on this head. All review threads are resolved,
and the security scan passes.
- Passed `pnpm -r typecheck` and `pnpm build` locally.
- Passed 219 native-runtime and controller tests, including fake-clock
tests for three weeks of renewal and 30-day explicit timeouts. Six
denial tests confirm that renewal cannot extend expired, revoked or
mismatched authority.
- Passed 337 executor, cancellation and restart-recovery tests, plus 278
Rust runner-core library tests.
- Passed a real runner with a silent fake Codex provider across its
original lease expiry. Runner PID, provider PID, thread and active turn
stayed unchanged. Warm-attach recovery tests also pass.
- Passed all 83 plugin-worker tests and 159 of 161 workspace-runtime
tests locally. The two remaining assertions passed with a canonical
macOS temporary directory, as did the changed runtime fixture. The full
affected server shard passes in CI.
- An unchanged GitHub callback-ordering test failed once in CI, passed
locally in isolation, and passed its one test-shard retry. The final CI
summary is successful.
- The full local `pnpm test:run` sweep was interrupted after dependency
setup failures and load-related timeouts. Identified failing suites
passed in isolated reruns after the dependency repair. The complete test
matrix passed remotely in CI.
## Risks
- Deploy the controller and runner together to enable renewal. Older
peers keep their existing bounded lease behavior.
- Unlimited runtime permits long resource use until completion, explicit
cancellation, configured timeout or loss of valid authority.
- Lease renewal changes authenticated protocol handling. Regression
tests cover stale, revoked and mismatched authority, lost replies and
warm handoff behavior.
- Simulated multi-week tests and a real lease-boundary test do not
constitute a weeks-long production soak.
## Model Used
OpenAI GPT-6 through Codex, with repository inspection, code execution
and TypeScript/Rust test tools. The exact backend revision 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 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.
> - Users can stop a run and send another message on the same task.
> - Remote runners need evidence from their sandbox provider that
execution stopped.
> - Local process checks cannot prove that a remote process exited.
> - This pull request records provider stop receipts and uses them for
conversation admission.
> - New user messages can proceed after confirmed cleanup without
repeating interrupted actions.
## Linked Issues or Issue Description
Refs #13237 and #13239. Related: #13163 covers app-restart recovery;
this change covers an explicit stop followed by a new user message.
**What happened?**
A stopped remote Claude ACP task kept its execution hold after Daytona
cleanup succeeded. Native runners also rejected remote process
identities and retained stale session cleanup gates. A message sent
during cleanup could stay deferred after the sandbox stopped.
**Expected behavior**
After the provider confirms that the old execution stopped, a new user
message starts a fresh turn. Pending user messages must not need another
message to trigger admission. Prior action outcomes remain recorded.
**Steps to reproduce**
1. Start a long-running task in Daytona with a legacy Claude ACP or
native ACP runner.
2. Cancel the run while its tool is active.
3. Send a new message immediately, or after cleanup completes.
4. Observe the execution hold despite the old sandbox having stopped.
**Paperclip version or commit**
Reproduced on master at 7b829efdf6. The
branch is rebased on current master.
## What Changed
- Add optional provider stop receipts to sandbox release and destroy
hooks. Old plugins remain compatible.
- Bind receipts to the company, run, lease, and provider resource.
Failed cleanup cannot supply stop authority.
- Acknowledge legacy remote cancellation after confirmed termination.
- Admit native user continuations using remote receipts instead of host
process checks.
- Retire only the settled cleanup owner matching the stopped company,
run, and provider resource. Isolate cleanup gates between remote
sandboxes, including two sandboxes owned by one run.
- Reconsider user messages deferred during remote cleanup through normal
admission, including successful later cleanup retries.
- Preserve receipts through cleanup retries and inline cleanup after
failed startup.
- Permit provider destruction after a terminal remote checkpoint
failure. Busy ownership still blocks destruction.
- Add regression tests and update execution semantics.
- Give the real preview fixture ten seconds for cold startup. Reuse
release-mode Rust artifacts for the filtered parity checks, avoiding a
duplicate debug test build that exhausted CI disk twice; test filters
and assertions are unchanged.
## Verification
- Recursive typecheck and build passed.
- Targeted server, Daytona plugin, and native runtime tests passed. They
cover missing or mismatched receipts, failed cleanup, local process
protection, exact cleanup ownership, and a message sent during cleanup.
- Live Daytona tests passed for legacy Claude ACP, native per-turn,
native warm, and a newly created native runner using disposable
sandboxes. Each original run was cancelled; its explicit follow-up
completed with no execution hold. The disposable case confirmed a new
sandbox after deletion.
- Native tests used the provider's Opus 5 selector, `opus[1m]`, and the
Linux runner bundle from the sandbox image. Existing checkpoint/sync
finalization warnings remained visible before the native follow-ups
reached committed success. This change does not repair those separate
warnings or guarantee recovery of uncheckpointed files.
- A direct live provider test also passed with the final delete-wait
change: the destroy hook returned its receipt only after Daytona
reported the sandbox destroyed. The final Daytona plugin suite passed
all 153 tests; its build passed.
- After rebasing on master, 284 targeted server/plugin tests and 313
native executor tests passed. The native session runtime suite passed
all 128 tests.
- The review regression passed all 145 tests across the continuation,
environment runtime, and pending-cleanup sweep suites. Recursive
typecheck and build passed again after that fix.
- The security review's exact-resource finding is fixed. Cleanup
completion requires the same provider-resource scope used during session
creation. All 128 native runtime tests, 61 server continuation/cleanup
tests, recursive typecheck, and build passed after this fix. The
two-sandboxes-in-one-run regression proves one receipt cannot retire the
other quarantine.
- Greptile reviewed the final commit at 5/5, the security scan passed,
and no review threads remain unresolved. The final native executor suite
also passed all 313 tests.
- The full local suite reached 10,616 passing tests before stopping on
four failures. All four now pass in focused reruns: the final-code
continuation/teardown suites, the attachment suite, and the native
session test after building its required fake-provider binary. The
original full run overlapped source edits and did not reach the
remaining groups; it is not counted as a full-suite pass.
- The preview exposure suite passed all 25 applicable tests (three
Linux-only tests skipped locally) after the startup allowance change.
Both release-mode Rust parity commands passed locally. All final-head PR
checks passed, including full server/workspace/browser test groups, full
native runner verification, build, typecheck, canary dry-run, and the
aggregate gates.
## Risks
- A provider must return a receipt only after confirmed termination.
Incorrect provider claims could permit overlapping execution.
- Older providers without receipts retain the existing hold. Missing
evidence, failed cleanup, active ownership, pauses, approvals, and
budgets still block admission.
- This change preserves unknown action outcomes and old checkpoints. It
does not authorize replay or alter historical runs.
- No database migration or telemetry contract change is required.
## Model Used
OpenAI GPT-6 through Codex, with repository inspection, code execution,
and browser tools. The exact backend revision 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 references)
- [x] My branch name describes the change and contains no internal
Paperclip ticket id or instance-derived details
- [x] I have run targeted 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 and their tasks.
> - A task can outlive a provider process or a server restart.
> - Legacy recovery treated unknown tool outcomes as a permanent
execution hold.
> - That hold could also reject a later user message.
> - A conversation turn can use prior history without replaying prior
tool calls.
> - This pull request lets supported conversation adapters continue
within the existing retry budget.
> - Users can send a new message after automatic attempts stop.
## Linked Issues or Issue Description
**What happened?**
A server restart could interrupt a local ACP run and leave its task
behind a permanent recovery hold. A later user message could be
cancelled before the provider answered. The immediate recovery path
could also create a successor outside the durable failure counter.
**Expected behavior**
Continue with a bounded new conversation turn. Preserve a compatible
provider session or use full task context when it is unavailable. Do not
replay recorded tools. When automatic attempts stop, allow a new user
request through the normal execution gates.
**Steps to reproduce**
1. Start a task with a local conversation adapter.
2. Restart the server while the provider is working.
3. Let the previous run become interrupted.
4. Send a follow-up message and observe the recovery hold on the old
behavior.
Related work: Refs #13075 for durable task recovery. Refs #12946 for
retry-limit and checkout-lock handling. This change routes conversation
recovery through the existing bounded scheduler.
## What Changed
- Mark supported local conversation failures for continuation. Keep
native-runner and non-conversation recovery rules.
- Carry an interruption notice into the next turn. Retain stopped ACP
session history even when a write outcome is unknown.
- Clear unavailable ACP sessions so the next bounded attempt can use
full task context.
- Route immediate failure recovery through the same durable scheduler as
process-loss recovery. Release only the predecessor checkout when its
retry takes ownership.
- Retire obsolete conversation holds using immutable run evidence, in
bounded batches with an activity record. Preserve outcome evidence and
do not wake historical tasks.
- Block actual admission and Resume while a predecessor process or
environment lease is still active. Keep the original interruption notice
after a rejected wake. Preserve the upstream blocked-wake waiting
contract: bounded retry planning can happen during cleanup, while
deferred messages and execution remain gated.
- Add subprocess and database regression tests. Update the execution
contract.
- Add the current thread-status field to the native recovery provider
fixture so its damaged-journal test reaches the intended boundary.
Tolerate an already-exited fixture process during test cleanup while
still asserting both processes terminate.
## Verification
- Workspace typecheck passed: `pnpm -r typecheck`.
- Build passed: `pnpm build`.
- Module boundaries passed: `pnpm check:module-boundaries`.
- Focused tests passed: 293 recovery/session/dispatch tests, 66 retry
and response-gate tests, and 37 native-session tests. Some suites
overlap.
- Tests cover interrupted writes, missing sessions, concurrent retries,
restart persistence, pending questions and approvals, execution gates,
and historical holds.
- Built the Rust test executables with `pnpm --filter
@paperclipai/paperclip-runner build:rust` for native-runner
verification.
- Full Vitest coverage verified locally using the repository’s general
and serialized shards, with focused reruns for failures and files not
reached after a shard stopped. The ownership-gate regression is fixed
and the complete affected server shard passes (1,390 tests). Local
parallel runs also hit temporary-directory, resource, and timing
failures; those suites pass with canonical temporary paths and
sequential reruns. No test timeouts were increased.
- Final merged-branch regression run: 577 tests pass across process
recovery, retry scheduling, liveness, durable chat, wake-queue
application/adapter, dispatch, continuation, native sessions, and task
chat. Earlier focused verification also passed 19 native control tests.
Token gates and whitespace validation pass.
- Browser verification passed all three ACP Stop/continue/pause
scenarios, including a rerun after merging the upstream waiting
behavior: `PAPERCLIP_E2E_PORT=3397 pnpm test:e2e
tests/e2e/acp-stop-continuation.spec.ts`. The interrupted-write case
verifies that follow-up completes without a repeated write.
- Final-head [CI run
34625037394](https://github.com/paperclipai/paperclip/actions/runs/34625037394)
passed on `06ac4bd9d150f8b209a96e5fd609c696958794a0`: all 31 reported
checks are green, including server/workspace suites, all browser shards,
native runner verification, build, typecheck, release dry run, and
aggregate gates. The two conditional Storybook checks were skipped.
Greptile reviewed this exact commit at 5/5; all review threads are
resolved.
## Risks
- A new model turn can choose to repeat an action. Paperclip does not
replay recorded tool calls and does not certify unknown action outcomes.
- Conversation adapters now stop after their retry budget instead of
requiring action reconciliation. Explicit Stop, pause, dependency,
approval, budget, and ownership gates remain in force.
- No schema migration or dependency changes. Historical holds are folded
without changing task status or waking work.
## Model Used
OpenAI GPT-6 through Codex, with reasoning, repository tools, code
execution, and test execution. The session does not expose a more
specific model build ID or context-window size.
## 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 native runner sends events that keep task state and steering
controls current.
> - Debug trace correlation reread and parsed the full trace for each
pending event.
> - Repeated scans blocked event delivery and left task state minutes
behind the provider.
> - This pull request indexes appended trace records once and reuses
those correlations.
> - Operators get current run state while native trace evidence remains
available.
## Linked Issues or Issue Description
**What happened?**
Native runs appeared live after their provider turn ended. Steering was
unavailable while the board still showed an active run. A CPU profile
attributed 93% of sampled time to trace lookup and file reads.
**Expected behavior**
Debug trace correlation must not delay run events or task state by
minutes.
**Steps to reproduce**
1. Enable native provider trace capture for a Codex run.
2. Produce a long event stream with pending correlations.
3. Compare provider event times with persisted run event times.
**Paperclip version**
Reproduced on source build 6dd48cad43.
**Deployment mode**
Self-hosted server with the native Codex runner.
Related event-delivery work: Refs #12208. This fix addresses synchronous
trace lookup rather than disconnect draining. Searches found no
duplicate trace-index fix.
## What Changed
- Add a transport-owned incremental index for native trace frame
correlations.
- Read at most 1 MiB per lookup and retry incomplete prefixes and
partial records.
- Preserve latest-frame and terminal-status ordering. Invalidate
replaced or truncated traces.
- Clear each index on transport close. Keep active indexes independent.
- Skip correlation records over 64 KiB without buffering or parsing
their full contents. Preserve the original trace file.
- Add regression tests for repeated misses, appended records, partial
UTF-8, large traces, replacement, and cleanup.
- Document the trace lookup performance constraint.
## Verification
- Latest revision: 11 index regressions and the existing real-runner
trace correlation test pass (12 tests).
- Includes 24 interleaved active traces, oversized records across
appends, and incomplete later interpretations.
- `pnpm build` and `pnpm -r typecheck` pass on final head `ddc74e42c`.
- The real-process hard-restart regression passes after rebuilding its
fake-provider binary from the rebased sources.
- The full local test command reports a pre-existing assertion mismatch
in `native-session-resume.test.ts`: the damaged-epoch recovery case
expects the old attach error instead of the new startup-history error.
The same exact failure was reproduced in an isolated checkout of
unchanged base `a20ecce40`. The PR does not change that test or startup
behavior.
- A recorded-trace benchmark reduced 120 lookups from 2.20 seconds to 22
milliseconds.
- The deployed hotfix removed the scan hotspot. The final main-thread
profile was 94% idle. A guarded restart reported no lost runs.
- All GitHub checks pass on `ddc74e42c`, including build, typecheck,
general tests, serialized server tests, end-to-end tests, canary dry
run, and security checks. Greptile is 5/5 on that exact head with no
open threads.
- Two timeout cases from the long local run pass in isolation: exhausted
quota-monitor evidence and native-question expiry. The duplicate full
local run was stopped after the complete CI suite passed.
## Risks
Each active transport retains event-to-frame mappings in memory and
clears them on close. Records over 64 KiB do not enter the correlation
index; the original trace file retains them. Large traces can need
multiple pending retries before a correlation is available. These
records are diagnostic; this change does not alter execution authority,
event acknowledgements, or provider commands.
## Model Used
OpenAI Codex, GPT-6. The runtime does not expose a more specific model
ID or context-window size. Used reasoning, code editing, shell
execution, CPU profiling, and tests.
## 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 tests; the
full-suite baseline failure is documented 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 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 is the open source app people use to manage AI agents for
work.
> - People also ask agents for work in their existing chat tools.
> - Each external conversation needs one task and a current authorized
source.
> - Retries, Stop, and provider failures must not duplicate work or
expose private data.
> - The first chat PR establishes the opt-in provider and data
contracts.
> - This PR adds experimental channel integration and its durable
control plane.
> - Users can request work from connected channels and inspect delivery
in Paperclip.
## Linked Issues or Issue Description
Refs #13100 and #13092. This is the second of exactly two chat PRs.
Foundation #13100 is merged and changed 143 files. Runner prerequisite
#13092 is also merged. This PR changes 400 files against master, below
the 500-file review limit. It contains no wireframe images or HTML
galleries.
## What Changed
- Add native Slack, GitHub, Microsoft Teams, Telegram, and Discord chat
connections. Keep chat disabled unless the operator enables experimental
chat connectors. Preserve the production GitHub tool connection and its
normal setup path.
- Bind each provider bot identity to one immutable Paperclip agent. Bind
each admitted external conversation to one task. Paperclip owns tasks,
runs, permissions, and audit records.
- Add durable admission, per-conversation queues, questions, task
controls, progress, final replies, images, files, and delivery receipts.
Board comments remain internal unless explicitly sent to the channel.
- Check current identity, provider reach, resource access, credentials,
runtime generation, and exact source before provider effects. Keep
private responses private. Never send raw reasoning, private logs,
credentials, or tool arguments.
- Hold uncertain sends for explicit audited resolution. Make Board
Send-to-channel atomic and idempotent. Keep reconnect and setup
credentials in Paperclip secret storage.
- Preserve current native-runner authority across retries, lost
acknowledgements, and recovery. Keep immutable input and completion
contracts separate from newer user input. Receipt reconciliation cannot
launch a provider.
- Reconcile chat close/new ordering and provider-effect lock order.
Audit resource access changes in the same transaction. Submit only the
selected resource from each UI toggle so stale pages cannot undo
unrelated access changes.
- Drain Codex stdout before certifying process exit. Bound the drain
with the existing shutdown grace. Preserve observed terminal authority
without treating an undrained process as successful or reusable.
- Incorporate master `018ca5da` with its ACP Stop, mobile task layout,
runner packaging, and official lock changes. Preserve dedicated
chat-answer continuations in both directions when ordinary queued
comments are adopted after Stop.
- Fence late adapter readiness behind an earlier Stop for the same run.
Preserve verified cleanup for registered adapters. Handle single Stop,
agent pause, duplicate Stops, and failure release without creating a
false cancellation receipt.
- Incorporate master's `6dd48cad4` wake-queue extraction. Preserve exact
failed-chat retry authorization and lineage, retired question-source
suppression, and the block on generic recovery that would discard the
admitted source. Fresh deferred input retains its separate promotion
path.
- Incorporate master `2a05b5ed3` and its queue-admission extraction,
simplified transaction ports, and separate runner CI job. Preserve exact
durable receipts, actor separation, and dedicated-answer isolation
through the new module. A failed receipt insert rolls back the
accompanying deferred-wake merge.
## Verification
Current head: `afe19299d06253cb628eb398e91d1200ea9f412a`, incorporating
master `2a05b5ed3457ea33efd6895520447d1d97fe98d8`. The conflicts are
resolved. This successor fixes two test-harness boundaries exposed by
CI: per-case route-module preparation and actual durable-save completion
before intentional runner termination. Production code and all existing
test/turn deadlines are unchanged. [Exact-head Greptile
review](https://github.com/paperclipai/paperclip/pull/13038#issuecomment-5587250594)
is **5/5**, completed September 10 at 13:20:55 UTC, with no actionable
findings or open review threads. [Fresh exact-head
CI](https://github.com/paperclipai/paperclip/actions/runs/34481724341)
passes **all 24 jobs**, including Build and both required aggregates.
Normal exact-head guarded merge was attempted and rejected by the
remaining branch approval policy: CODEOWNER review is required and no
human approval is present. Normal **squash auto-merge is enabled** as of
September 10 at 13:36:26 UTC. Requested CODEOWNERS have been notified;
no approval bypass or self-approval was used. Earlier-head results below
remain historical evidence, not qualification of this successor.
- Final exact-head Linux evidence: 995/995 chat integration cases; 36/36
agent-skills routes; 35/35 runner live-session cases, including real
process kill/resume; 1948 runner Vitest cases with three existing
benchmark/platform guards; 870/870 API-authority cases; and 104 browser
cases with four existing optional skips. Rust, conformance/replay, full
repository build, typecheck, canary, all server/workspace shards, and
both required aggregates pass with normal CI concurrency. Earlier failed
attempts remain recorded below.
- Latest test-only qualification: 141/141
route/permissions/authentication cases pass in separate cold forks, with
plain server types and independent review clear. The real-runner suite
passes 35/35, with plain runner types and independent review clear. A
controlled premature-save acknowledgement fails as expected; matching
ownership/effect/process evidence, rejected saves, real turn outcome,
test abort, and pre-kill liveness are covered. No local reproduction of
the original CI scheduling failure is claimed. The preceding [CI
run](https://github.com/paperclipai/paperclip/actions/runs/34479680858)
passes 21/24 jobs, including all 995 Linux chat cases and browser
aggregate (104 passed, four existing optional skips); only Build, the
skills serialized shard, and the required verification aggregate fail.
Its exact-head Greptile review was 5/5. Both failed job logs are
retained.
- Final fixture qualification: all eight focused Discord cases and all
995 chat integration cases pass. The exact modal statement/PID is
observed before taking the real connection lock; the test then proves
its actual blocking relationship before mutation. Original SQL
execution, provider behavior, negative assertions, and 1s/15s timeouts
remain unchanged. Independent review is clear and test/production hashes
remain frozen. The preceding [CI
attempt](https://github.com/paperclipai/paperclip/actions/runs/34477184777)
passed 22 jobs, including Build/runner, typecheck, canary, all other
test shards, and browser aggregate (104 passed, four existing optional
skips); the two fixture failures and failed verification aggregate
remain recorded, not relabeled as a pass.
- Current queue-module composition: 308/308 recovery/batching/queue/Stop
tests; 995/995 full chat integration; 89/89 module tests, including real
PostgreSQL receipt-insert rollback; 24/24 workflow/module-boundary
tests; plain server and UI types. All four actual local process/ACP
browser paths pass in 1.4 minutes. Fresh databases, no skips or retries,
stable reviewed source hashes. The initial boundary failure is retained;
its no-op service wrapper was removed without changing recovery context
or weakening the check. An exploratory standalone test-directory
typecheck fails because its new upstream transformation config is not a
standalone typechecking project; standard CI/build does not invoke it,
and no configuration was weakened to suppress those diagnostics.
- The preceding head `e02a63d462ce5d47433b0aeb632bb6fd20aab1ba` passed
[all 24 CI
jobs](https://github.com/paperclipai/paperclip/actions/runs/34436462958)
and exact-head Greptile review at 5/5. Required CODEOWNER review
prevented its normal merge before master advanced again.
- Final extracted-module composition: 307/307 recovery, batching, queue
and Stop-control tests; 995/995 full chat integration; 49/49 module
tests including eight PostgreSQL adapter cases; and 19/19 issue-update
tests. Plain server types pass. All four actual local process/ACP
browser paths pass in 1.3 minutes. Fresh databases, no skips or retries
in these cohorts, frozen source hashes, and independent review clear.
- The preceding head `3e4e1c1c` passes [all PR CI
jobs](https://github.com/paperclipai/paperclip/actions/runs/34415826820),
including Build and required `ci / verify` and `ci / e2e`. Both the
original Rust failure and the previously load-sensitive lineage fixture
pass with unchanged Linux concurrency. Master advanced afterward and
required this reconciliation.
- Final master composition: 448/448 focused UI tests, 186/186 adapter
tests, 24/24 queue/control tests, and 11/11 packaging tests. Plain UI,
server, shared, and adapter types pass. Token gates and diff checks
pass. Independent server and UI reviews are clear.
- Stop-registration regression: both real-service cases fail against
exact `a95` source and pass with the fix. The full corrected
recovery/control suite passes 265/265. Duplicate-owner and failed-Stop
controls also pass. Plain server types pass. The readiness barrier
prevents provider startup without adding an acknowledgment to an already
terminal run.
- Final qualification strengthens terminal-field equality and repeats
both affected cases successfully on a fresh database. All four actual
local process/ACP browser paths pass again in 1.3 minutes, without skips
or retries. The final screenshot shows Cancelled, a paused subtree,
retained input, and no error toast.
- Two new actual-service regressions fail before the merge fix. They
prove that queued-comment adoption could consume a dedicated chat answer
or add unrelated input to that answer. The fixed four-case cohort
passes, including ordinary upstream continuation and adapter Stop
controls. Full recovery passes 257/257. All four actual local
process/ACP Stop browser flows pass in 1.4 minutes, without skips or
retries, on a fresh database.
- The unchanged runner artifact was qualified with 171/171 transport
tests, 870/870 API-authority tests, conformance 1/1, and replay 11/11.
Six controlled reader tests prove the exit/drain repair. Its local
serial Rust workspace passed 546 top-level cases plus two invoked
helpers; the later passing Linux CI supplies default-concurrency
evidence.
- Prior exact-source full chat integration passes 995/995. Settings
regressions cover concurrent stale pages, 501 destinations, pending
state, rejected updates, and explicit retry. These deterministic tests
do not prove live provider behavior.
- Retained failed attempts and their causes are in the [qualification
log](afe19299d0/doc/plans/chat-adapters/2026-09-08-chat-queue-and-webhook-repair.md).
The first merge adapter run timed out while macOS slept for 290 seconds.
Its unchanged repeat passed with a temporary sleep guard. No assertion,
deadline, or CI gate was weakened.
Review commands include `pnpm --filter @paperclipai/server exec vitest
run src/__tests__/heartbeat-process-recovery.test.ts
src/__tests__/issue-queued-comments-routes.test.ts` and `pnpm exec
playwright test --config tests/e2e/playwright.config.ts
tests/e2e/acp-stop-continuation.spec.ts`. Database suites require fresh
disposable databases. See the [browser
runbook](afe19299d0/doc/plans/chat-adapters/2026-09-04-chat-adapters-browser-e2e-runbook.md)
for provider setup and separate live acceptance steps.
## Risks
- This remains experimental. Deterministic tests and bounded live
evidence do not establish every provider feature, tenant, permission
layout, or media shape. Teams work-tenant qualification is still open.
- Failed and uncertain provider effects remain visible and can require
operator action. A transport receipt does not prove recipient
visibility.
- Native controller and runner artifacts must remain compatible.
Preserve lease ownership, terminal authority, source binding, and
quarantine during future changes.
- Access and audit rows commit together, but activity notifications
remain best-effort. This is not a new durable event outbox.
- The PR operation does not deploy a live server, replace its runner, or
change provider permissions. Remaining live qualification is documented
in the [temporary
handoff](afe19299d0/doc/plans/chat-adapters/2026-09-08-open-qualification-followups.md).
## Model Used
OpenAI Codex assisted with implementation, tool execution, testing, and
review. The work records `gpt-6-astra` assistance. The environment does
not report a context-window size. No private reasoning traces are
included.
## 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 Paperclip Runner manages live agent sessions and their suspend
path.
> - A turn-timeout promise could reject before `reconcileActiveTurn()`
attached its handler.
> - A short provider-drain budget could reject a valid suspend on a
loaded continuous-integration host.
> - This pull request closes both timing windows and adds deterministic
regression tests.
> - The benefit is a fail-closed suspend path that does not report false
failures under load.
## Linked Issues or Issue Description
**What happened?**
Capability-live tests failed under load. A turn-timeout promise could
raise an unhandled rejection during a slow interrupt round trip. An idle
provider drain could also exceed its one-second proof budget during
suspend.
**Expected behavior**
The suspend path must observe turn-timeout rejections and allow enough
time for one provider command round trip. It must still fail closed when
the runner does not prove durable suspension.
**Steps to reproduce**
1. Run the capability-live tests on a loaded four-vCPU
continuous-integration host.
2. Delay a `turn/interrupt` reply beyond the turn timeout.
3. Close a live session and observe the suspend barrier.
**Paperclip version or commit**
`01b442b2926ac2010a2fcfbda6592802472ef07f`
**Deployment mode**
Built from source with the Paperclip Runner test suite.
## What Changed
- Attach the turn-timeout rejection handler inside `armTurnWaiter()` at
promise creation.
- Remove the redundant per-call-site guard in `sendMessage()`.
- Use a uniform five-second provider-drain proof budget, capped by the
outer preparation deadline.
- Remove the unused boolean return from the provider-turn-stop helper.
- Add deterministic tests for the delayed interrupt and the short close
grace period.
## Verification
- `npx vitest run
packages/paperclip-runner/src/live/live-session.test.ts` passed locally
with zero skipped tests.
- `npx vitest run
packages/paperclip-runner/src/live/runnerd-codex-transport.test.ts`
passed locally with zero skipped tests.
- The two new tests appeared in the local run output and were not
skipped.
- The package type-check passed locally.
- GitHub Actions must confirm the full continuous-integration suite,
including `Verify Paperclip Runner`.
## Risks
- The provider-drain wait now allows up to five seconds before the outer
deadline caps it.
- The fail-closed suspend barrier remains unchanged.
- The test suite still depends on the Rust runner binary for
capability-live tests.
## Model Used
OpenAI Codex, GPT-5. The runtime provided tool use and code execution.
The runtime did not provide a context-window value.
## 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 runs Codex locally and in remote sandboxes.
> - The runner must preserve startup configuration and session identity.
> - Missing project trust can disable repository configuration.
> - Full-history requests use deprecated provider fields.
> - Resume usage describes old work and must not become new run usage.
> - This change corrects startup trust, state reads, and usage
classification.
## Linked Issues or Issue Description
**What happened?**
Normal Codex runs could show repository-trust and history-deprecation
warnings.
Resume could report the preceding turn's token snapshot as a late-turn
warning.
The historical last-usage value could also be attributed to the new run.
**Expected behavior**
Trust the server-selected startup root in isolated configuration. Read
lightweight
provider state and paginated evidence. Use historical cumulative usage
as a
baseline without a new charge or user-facing warning.
**Steps to reproduce**
1. Start a native Codex task in a selected repository.
2. Finish the turn and resume the provider thread.
3. Inspect provider notices, history requests, and per-run usage.
4. Repeat startup and cold resume inside a Daytona sandbox.
**Paperclip version or commit**
Codex CLI 0.153.4 is the pinned runtime and reproduced baseline.
Replayed onto master at 6abeb6733. Related authority work: Refs #13092.
This PR retains its startup cleanup and protocol-integrity checks.
**Deployment mode**
Local source checkout and disposable Daytona sandbox.
## What Changed
- Classify the exact historical resume usage event before the generic
stale-turn warning.
- Persist cumulative usage baselines across recovery of the same run.
- Use excludeTurns on resume and lightweight thread reads.
- Page turn metadata and selected turn items with cursor and identity
validation.
- Reject unsupported or incomplete history instead of guessing that
execution is idle.
- Trust the startup execution root on its host, including Git worktree
trust keys.
- Start Codex in that root and retain the selected sandbox profile on
later turns.
- Keep unrelated isolated configuration and Codex's separate hook trust
policy.
- Add Rust, TypeScript, accounting, native integration, and local
run-log documentation.
## Verification
- Codex and native-transport TypeScript: 333 passed before PR replay.
- Adjacent OpenCode/ACPX driver and accounting tests: 49 passed.
- Rust library, serialized: 226 passed. Native Codex integration: 72
passed, 1 ignored, plus two pagination regressions.
- Repository typecheck and build passed. All repository test groups have
passing coverage after fixture and resource retests; the initial
monolithic command was not clean.
- Fresh real Codex native browser tasks returned correct answers without
the three targeted notices. Answers persisted after refresh and restart.
- Real same-thread TypeScript driver tests passed locally and in
Daytona, including cold resume, configuration, skills, and an approved
harmless hook.
- Local usage summed to 64,607 tokens. Daytona usage summed to 42,737
tokens. Each sum matched its final session total exactly.
- See doc/plans/2026-09-09-codex-integration-acceptance.md for the scope
and limits of the live tests.
- After replay onto current master and review fixes: 334 Codex, backend,
and live-session tests passed, including checkpoint serialization and
real-runner process restart. TypeScript checks passed.
- The native Codex integration run passed 83 tests; the large lineage
test passed separately with the release runner (its debug build exceeded
the test deadline).
- All GitHub checks passed on the final PR head. Greptile is 5/5 with no
unresolved review threads. CI regenerates the lockfile for the added
TOML dependency, per repository policy.
- The first server shard hit a timing-dependent duplicate-key failure in
the unchanged artifact-document concurrency test. Its focused 11-test
suite passed locally. One CI retry on the same head passed all 103 files
and 1,405 tests (2 skipped): [retry
result](https://github.com/paperclipai/paperclip/actions/runs/34398832930/job/102631274667).
## Risks
- Trust applies only to the server-selected startup root and isolated
configuration. Sandbox and tool permissions remain authoritative.
- Codex still requires approval of individual hook hashes. This change
does not bypass that policy.
- Providers without the required history APIs fail explicitly.
- Daytona acceptance used the production TypeScript driver. Remote
Paperclip UI and remote Rust execution were not tested.
- No new public API, database state, recovery policy, or UI control is
included.
## Model Used
OpenAI Codex, GPT-6 (`gpt-6-astra`). Used for reasoning, code edits,
tool use,
and test execution. The exact context-window limit is not exposed in
this
session. Real-provider acceptance used Codex CLI 0.153.4 with
`gpt-5.6-sol`.
## 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
#` 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>
Keep receiving provider events after paperclip_finish, drain pending event persistence, and select the final assistant answer after the provider turn ends. Preserve cancellation, failure, and governed-wait behavior.
Co-Authored-By: Paperclip <noreply@paperclip.ing>
## Thinking Path
> - Paperclip shows agent work in task feeds.
> - Native tools and provider notices appear in that feed.
> - A routine completion call adds no useful action for the user.
> - A provider notice needs readable text to explain its warning.
> - This change hides the completion call and displays the notice
summary.
> - Saved run events remain available for inspection.
## Linked Issues or Issue Description
**What happened?**
The feed showed paperclip_finish as a normal tool call. Provider notices
showed
a generic name or a large key-value table instead of a clear warning
message.
**Expected behavior**
Hide the routine completion call from the task feed. Show a warning
icon,
a short severity heading, and the full notice summary across the row.
**Steps to reproduce**
1. Open a native Codex task that calls paperclip_finish.
2. Inspect its tool activity in the task feed.
3. Inspect a run with a repository-trust provider notice.
**Paperclip version or commit**
Reproduced on the implementation checkout. Replayed onto master at
6abeb6733.
The search found no duplicate PR for this display change.
**Deployment mode**
Local source checkout with native task feeds.
## What Changed
- Hide paperclip_finish calls and results in task-feed adapters.
- Preserve the raw events for run-log inspection.
- Carry provider-notice text into the transcript.
- Render notices with an icon, severity heading, and full-width summary.
- Cover the display and transcript mapping with regression tests.
## Verification
- The affected UI tests and the token gates passed on the implementation
checkout.
- Browser inspection confirmed the warning text is readable and
completion calls are hidden.
- Repository typecheck and build passed. Repository test groups passed
after resource retests.
- On this PR branch, 123 focused UI tests and the token gates passed.
All GitHub checks passed; Greptile is 5/5 with no unresolved threads.
## Risks
- The task feed hides one known internal tool. Raw run events remain
unchanged.
- Long warning text must wrap within the available width.
- No provider warning classification, accounting, or recovery behavior
changes here.
## Model Used
OpenAI Codex, GPT-6 (`gpt-6-astra`). Used for reasoning, code edits,
tool use,
and test execution. The exact context-window limit is not exposed in
this
session. Real-provider acceptance used Codex CLI 0.153.4 with
`gpt-5.6-sol`.
## 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
#` 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 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.
> - The native runner carries tool results and task output to the
control plane.
> - A lost connection must not change which run owns a result.
> - A session must not become reusable while provider output is still
pending.
> - This pull request adds strict recovery evidence and bounded drain
barriers.
> - It preserves current PRP version negotiation and session-goal
support.
> - The benefit is safer reuse of native sessions after a transport
failure.
## Linked Issues or Issue Description
Refs #13038. This is the first of two stacked pull requests. It contains
the native runtime prerequisites. The second pull request contains the
experimental chat-channel integration. It preserves the provider
identity and typed terminal-failure contracts in #13074 and the durable
recovery work in #13075.
**What happened?**
Native session failures could leave retained provider events, incomplete
tool results, or warm handoff state that was not safe to reuse. A later
run could observe output from an earlier authority.
**Expected behavior**
Recovery must preserve exact run, tool, process, artifact, and lease
evidence. Uncertain or corrupt state must fail closed. A successful
close must prove that retained provider output is settled.
**Steps to reproduce**
Run the transport and control-plane regressions. They hold and drop
authenticated frames, fail durable writes, and restart fresh controllers
and runner processes with retained state. Provider executables are local
test fixtures.
## What Changed
- Preserve pending provider cleanup and semantic-result evidence across
session close and restart.
- Add an authenticated warm handoff with exact old and new identities,
durable receipts, and completion acknowledgement.
- Drain retained provider events under the cumulative acknowledgement
fence.
- Reject corrupt tool-result contracts without unsafe provider replay or
reusable checkpoints.
- Keep ordinary PRP v1 sessions and current session-goal behavior.
Require negotiated PRP v2 and acknowledged native session evidence
before warm authority rotation.
- Preserve late semantic inputs and exact durable result receipts until
close can prove settlement.
- Add transport, crash-window, artifact, checkpoint, and final-output
regressions.
- Deduplicate resolved execution delivery under the current issue lock.
Reuse the exact existing successor after concurrent scans or a lost
acknowledgement. Preserve newer operator evidence.
- Persist idle provider integrity/capacity failures before process
retirement, retain permanent model-rejection classification, and keep
external question identifiers out of task instructions.
- Expose only the context source on native status events. Keep thin
dispatch projections compatible without exposing the complete context.
## Verification
- Review-fix revision: 128 runtime-context/native-session tests, five
idle-failure/adjacent Rust cases, 24 warm crash-window cases, three
startup-notification/close cases, and five attach/backlog cases passed.
The security and idle-failure cases were first reproduced failing.
- Prior merged revision: runner production build, TypeScript typecheck,
complete Rust workspace tests and formatting passed; 272 focused runner
tests and two real PostgreSQL regressions passed.
- Earlier full runner runs and CI Build failed on missing
semantic-result fixture receipts, stale local provider fixture bytes,
startup-notification ordering, and a confirmation-loss fixture that
could accidentally send its final ACK. Each cause was reproduced and
corrected without relaxing production authority or close assertions.
These earlier runs are retained as failures, not represented as passing
verification.
- The first local repository-wide run failed before later phases because
the isolated install omitted PostgreSQL's native-library aliases; it
also encountered an unrelated occupied-port fixture. Those results are
retained, not represented as a passing run.
- Exact `335b2ee52709afb3885d4d6ebb2a3ece4b5864d6`: the complete runner
suite passed 1,888 tests, with 10 existing skips. The full Rust release
workspace passed with serial test scheduling. The unchanged parallel
Rust run hit the five-second 300-descendant fixture deadline; that
failure is retained. No deadline or assertion was relaxed.
- The resolved-execution regression suite passed 57 tests, including
concurrent delivery, lost acknowledgement, superseded authority, and
newer operator evidence. Plain server typecheck passed. The
duplicate-delivery cases were first reproduced failing.
- Prior exact `335b2ee52709afb3885d4d6ebb2a3ece4b5864d6` CI passed all
required jobs and Greptile reported 5/5. Its local general-server run
passed 7,208 tests but failed one responsibility fixture; later phases
did not run. The fixture started the next wake while its bounded handoff
was active. It also used nonexistent comment IDs, which hid the current
stored-message-author identity rule. The updated tests use real message
authors, preserve task ownership, and await exact automatic handoffs. No
production identity policy changed.
- Current head `aa39275a1f300f7d1a0b16cd0885eea567cff6b0` includes
current master and the native context-source projection. The focused
identity/status cohort passed 27 tests and plain server typecheck
passed. Fresh full repository tests, types, build, required CI, and
Greptile review are pending. Final results will be updated before merge.
- This is deterministic local-provider evidence. It is not a claim of
complete live-provider qualification.
## Risks
- This changes authenticated recovery and close ordering. The TypeScript
transport and runner binary must be built from the same revision.
- Failed or incomplete evidence intentionally prevents reuse and can
require a fresh run.
- PRP v1 ordinary/cold sessions remain supported. A v1 connection lease
cannot upgrade in place. A current v2-capable runner held on a v1 lease
was qualified through owned-process retirement/join, fresh bootstrap on
the same old authority, v2 observation/ACK, then warm rotation. Legacy
binary replacement and adopted-owner migration are not qualified by that
test; rollout must not present them as automatic same-lease upgrades.
- This pull request has no database migration or chat-channel
activation. The second pull request keeps the channel feature
experimental.
## Model Used
OpenAI Codex assisted with implementation, tool execution, tests, and
reconciliation. The existing implementation records OpenAI `gpt-6-astra`
assistance. The current environment does not report a context-window
size. No private reasoning traces are included.
## 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
- [ ] 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 runs agents for people with different GitHub accounts.
> - Managed operations must use the intended person's eligible
connection.
> - A failed duplicate connection must not hide a healthy grant for the
same account.
> - Legacy hosts also need their existing Git configuration when managed
access is not configured.
> - Runner networking and local Git operations must not depend on GitHub
broker availability.
> - This pull request separates those policies and improves failure
diagnostics.
## Linked Issues or Issue Description
**What happened?** New runs always cleared host Git credentials and
installed managed launchers. Network permission depended on GitHub
environment variables. A launcher failure could stop even local `git
status`. A newer unhealthy duplicate could take precedence over a
healthy connection, and generic health errors were shown as reconnect
requirements.
**Expected behavior:** Use a healthy eligible managed connection for the
intended account. Preserve host authentication only for unconfigured
standard-trust local or SSH execution. Permit local Git during broker
failures and keep network permission independent of GitHub credentials.
**Steps to reproduce:** Configure healthy and unhealthy grants for one
GitHub account, dispatch an agent, and execute Git commands. Separately
run an unconfigured legacy host with existing GitHub CLI authentication.
Stop the broker and run local `git status`.
**Paperclip version or commit:** Master at 668110469. **Deployment
mode:** Self-hosted. **Installation method:** Git checkout. **Agent
adapters involved:** Native Codex runner and shared adapter execution
setup. **Database mode:** Existing instance database; no migration.
**Access context:** Responsible person's managed grant, or explicitly
unconfigured legacy host. **Node.js version:** 26.4.0 locally.
**Operating system:** macOS development and Linux execution hosts.
**Relevant logs or output:** Previously `GitHub credential context
unavailable` hid configuration, transport, and capability errors. New
diagnostics identify these categories without credential values.
**Additional context:** Refs #13005 and #13022. Dependency provisioning
is addressed separately in #13093.
## What Changed
- Prefer healthy eligible grants and retry credential acquisition once
for the same principal and account before starting an operation.
- Preserve host Git configuration only when managed access is
unconfigured on a standard-trust local or SSH target.
- Project authentication mode and validated Git metadata into native
runner boundaries; refresh resumed provider settings when modes change.
- Enable network access through an explicit standard-trust controller
decision, independently of GitHub. Omitted or restricted decisions stay
disabled; replace warm providers when that decision changes.
- Run local Git with cleared credentials when the managed broker fails,
with specific redacted diagnostics.
- Retry access-refresh conflicts once without treating concurrency as
expired authorization.
- Show retry instead of reconnect for transient GitHub health failures.
Add optional authorization and run-diagnostic fields without a database
migration.
## Verification
- All latest-head CI gates are green, including typecheck, general and
serialized suites, browser tests, canary, native runner verification,
and build. Greptile is 5/5 with no remaining findings; the security scan
passed.
- Full recursive typecheck and build passed. UI token gates passed.
- Full general server run: 7,110 passed, one transient socket hangup;
that file passed on retry. All remaining workspace groups passed,
including 5,552 UI and 478 CLI tests. The complete serialized rerun
passed all 144 suites / 2,179 tests after the initial isolated timeout
passed on retry.
- 195 final launcher and native session tests passed, including
host/managed transitions, local/remote warm network-policy changes,
broker rotation, and attempts to override validated controller
filesystem roots.
- GitHub gateway fallback, duplicate connection selection, refresh
conflicts, per-user reauthorization, and native transport/security
suites passed.
- Additional live native fixtures passed SSH public-key authentication
and a Git credential helper in fresh and resumed host-mode sessions. An
unwritable managed configuration directory preserved local Git (exit 0)
while GitHub CLI failed with `configuration_directory_unavailable` (exit
4). The temporary SSH listener and keys were removed.
- Applicable Rust suites passed except two timing failures under load;
each failed case passed in isolation. The final environment contract
test passed.
- Linux native runner acceptance passed in both managed and legacy host
modes: DNS, HTTPS, npm package download, fresh-worktree Git status,
authenticated GitHub user lookup, repository read, and a new run
continuing the same provider conversation. Managed broker outage
preserved local Git and rejected authenticated access without host
fallback.
- Matching Linux server/runner artifacts and the separate provisioning
repair are deployed to the development instance. A fresh UI-dispatched
task and a new run after a server restart both passed all six shell
checks through the live controller and credential broker. Both runs
selected the expected healthy connection/grant and retained the same
provider conversation. Neither connection was repaired or reconnected.
The sandbox roots are assigned from the validated execution-target probe
**after** ordinary bindings are merged. Regression coverage supplies
forged roots and verifies they cannot override the controller paths.
Networking is enabled only for an explicit
`PAPERCLIP_RUNNER_NETWORK_ACCESS=enabled` controller decision; omitted
values remain disabled.
## Risks
Unconfigured standard-trust local and SSH runs regain access to host Git
authentication resources. Managed, sandbox, plugin, and low-trust runs
do not gain this fallback. Revoked managed access never falls back to
another account. Deploy server and runner artifacts together;
already-started operations retain their captured identity. An
unauthenticated command can still fail when it requires GitHub access.
## Model Used
OpenAI GPT-6 through Codex, with code editing, shell execution, tests,
and browser inspection. The exact model variant 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.
> - The runner translates provider events into durable task execution.
> - Provider notifications can refer to another thread or a descendant.
> - Root validation treated these notifications as fatal, and later
layers could lose the original failure.
> - This pull request classifies event identity and preserves structured
terminal failures.
> - The server can then choose safe recovery without weakening tool
authority.
## Linked Issues or Issue Description
Refs #13038. This change incorporates the shared protocol-integrity and
bounded-cleanup prerequisites from that PR. It does not include the chat
feature. Related failure classification work: #13028.
**What happened?**
An informational provider notification for another thread could
terminate the root session. A failed stream could then become a
missing-result error and lose its cause.
**Expected behavior**
Ignore unrelated informational notices with bounded diagnostics. Reject
invalid authoritative events. Preserve the original failure code and
recovery meaning through cleanup.
**Steps to reproduce**
Run a native Codex task. Deliver a notification for an unrelated thread,
or close the stream after a structured failure. Inspect the root outcome
and recorded failure.
**Paperclip version or commit**
Reproduced before e20010472. This branch includes the current
session-goal contract from that commit.
**Deployment mode**
Built from source. Native runner with Codex.
## What Changed
- Classify root, provider-confirmed descendant, stale, unrelated, and
invalid provider events.
- Keep tool requests bound to their original execution authority.
- Preserve typed failures through transport, session, and durable
control-plane cleanup.
- Keep bounded cleanup failures separate from the primary execution
failure.
- Add compatible shared contracts for continuation context, execution
status, and explicit reconciliation. The dependent PR adds their server
and UI consumers.
## Verification
- Runner TypeScript and ACPX suites: 1,733 passed, 7 skipped. Node
contracts: 38 passed.
- Real provider-process fixtures cover 300 descendant identities across
restart, the 4,096-identity capacity boundary, and rejection of
continuation after terminal acknowledgement and restart.
- Repository build, typecheck, and full `pnpm test:run` passed on the
rebased stack (18,448 tests passed, 49 skipped). The full Rust workspace
passed with `--test-threads=1`; parallel execution exposed an existing
fixture port-reservation race. All latest-head CI checks passed. One
unchanged artifact-document concurrency test failed on the first CI run
and passed on its single rerun.
- Added notification, streaming failure, protocol integrity, cleanup
quarantine, and durable failure tests.
## Risks
Provider event classification must retain the new session-goal behavior
on master. Descendant notifications must never gain root tool authority.
Shared contract fields are additive. This PR does not migrate data or
start replacement provider work.
## Model Used
OpenAI GPT-6 through Codex. The exact deployment ID and context window
were not exposed. Used reasoning, tool use, code execution, 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
- [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 is the open source app people use to manage AI agents for
work
> - Agents use connections to reach external services.
> - A fresh native task can have no service tools installed.
> - The agent needs a way to discover services and ask the responsible
person for access.
> - This pull request brings the existing connection-intent flow into
native task execution.
> - The person can connect from the task, and the agent can continue
with updated tools.
## Linked Issues or Issue Description
**Subsystem affected**
Native runner tool authority, connection intents, task interactions, and
shared connection setup.
**Problem or motivation**
A task that needs an unconnected service cannot finish its work. Leaving
the task to configure access also loses context. A resolved request must
survive a restart and resume the correct agent once.
**Proposed solution**
Expose connection discovery and access requests as server-owned native
tools. Render a durable task card and use the shared setup dialog.
Persist outcome delivery and start a fresh provider session after access
is ready.
**Alternatives considered**
Sending the person to the Connections page adds navigation and does not
solve continuation. Polling for authorization consumes runs and can
create duplicate requests.
**Roadmap alignment**
This extends the existing connection-intent runtime and setup
experience. It reuses the shared access model and the native runner.
Related: #12345, #12347. The service-slug fix in #12906 is related but
separate. Companion evaluation PR:
https://github.com/paperclipai/paperclip-evals/pull/21.
## What Changed
- Expose `connections_search` and `connection_request` with server-bound
company, task, agent, and responsible user. Preserve the legacy entry
points.
- Discover catalog services and authorized custom connections. Check
installation, identity, health, and executable permissions before
reporting ready.
- Keep pending cards through ordinary messages. Reuse requests and
retire stale ownership. Put Connect at the right of Not now.
- Reuse the shared setup flow in a task dialog. Keep access additive and
default to the requesting agent. Recover from cancelled or blocked OAuth
windows with a new-tab fallback.
- Persist outcome delivery with an idempotent wake key. Resume in a
fresh session and recheck ownership before dispatch.
- Add native browser fixtures, offline Storybook states, server
contracts, and evaluation fixtures. Update guidance and documentation.
## Verification
- `pnpm build`: passed after replaying the change on current master.
- `pnpm -r typecheck`: passed.
- `pnpm check:token-gates`: passed.
- `pnpm --filter @paperclipai/ui build-storybook`: passed.
- New continuation-policy regression cases: 16 passed.
- Docker-backed PostgreSQL regressions passed for requester-only OAuth
access, assignment-only expiry, terminal expiry, and credential-free
setup metadata.
- Shared setup and task-card UI tests: 121 passed, including configured
MCP reconnect URL recovery and preserving user edits across refetch.
- Storybook browser checks: all 119 passed on the latest reconnect fix.
- `pnpm test:run`: 4,734 tests passed in the first server group, but
embedded PostgreSQL startup failures and resulting cleanup errors
prevented a complete local pass. All Linux CI lanes passed on the latest
reviewed commit. One external-object route test returned an unexplained
500 on the first run; it passed twice locally and the failed shard
passed on retry without code changes.
- Earlier feature-checkout evidence: three deterministic native browser
journeys passed, including restart delivery and an actual fixture tool
result. Legacy scripted coverage also passed. All 59 added stories were
inspected in light and dark themes.
- Live Notion testing recorded successful provider reads. The manual
test used a local-trusted instance. It does not prove
authenticated/cloud deployment or every provider journey.
- Native browser rerun reached the embedded PostgreSQL startup limit
before bootstrap, so the latest checkout’s full native browser journey
remains unverified. Both OAuth page/task regression cases passed against
isolated Docker-backed PostgreSQL 17. They verify no premature task
access, requester-only completion, additive retries, and reconnect
preservation.
- Applied both new migrations twice to isolated PostgreSQL 17. Foreign
keys remained intact, duplicate active delivery keys were rejected, and
failed delivery records did not block retries.
Reviewer path: start a fresh test drive, enable the native runner, use
an agent that can perform work directly, and ask it to summarize a
Notion page. Connect from the card, then verify the resumed provider
call and source-linked answer. The default test-drive CEO is instructed
to delegate, so it can introduce an unrelated hiring step.
## Risks
- Two additive migrations create durable deliveries and a partial unique
wake index. They are idempotent. The wake index can require a
maintenance window on large tables because migrations run in a
transaction.
- OAuth and continuation cross asynchronous boundaries. Tests cover
ownership changes, retries, additive access, and restart delivery; live
provider behavior still varies.
- The latest requester-scope fix has not yet been exercised through live
OAuth. GitHub, API-key, authenticated-user, and all recovery journeys
are not claimed as verified.
## Model Used
OpenAI GPT-6-based Codex assisted with implementation, tests, and review
using tools and code execution. The runtime does not expose the exact
model version, context window, or reasoning setting. Live evaluation
used `gpt-5.6-luna`; manual native testing used `gpt-5.6-sol`.
## Checklist
- [x] I have included a thinking path that traces from project context
to this change
- [x] I have specified the model used and disclosed unavailable runtime
details
- [x] I have checked ROADMAP.md and confirmed this extends existing
connection 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 issue
template
- [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 all required tests locally and they pass
- [x] I have added or updated tests where applicable
- [x] I have updated relevant documentation
- [x] I have considered and documented risks
- [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.
> - Several people can send instructions to the same agent and task.
> - A fixed GitHub token in the provider process can keep the first
person's access after another person's message is accepted.
> - Task ownership cannot select credentials for each accepted
instruction or preserve the identity of an operation already in
progress.
> - This pull request records ordered execution identity contexts and
resolves credentials when managed Git, gh, or GitHub tools start.
> - The benefit is automatic personal GitHub access for shared agents,
with durable continuation rules and no teammate credential fallback.
## Linked Issues or Issue Description
**Subsystem affected**
Cross-cutting: orchestration, connection grants, database, runtime
adapters, native runners, and run details.
**Problem or motivation**
A shared agent must use the person whose instructions it has accepted. A
queued message must retain its author. A retry or approval without new
instructions must retain the originating identity. GitHub must remain
optional for ordinary work.
**Proposed solution**
Persist execution identity separately from task ownership. Give new
processes a run-scoped broker capability and token-free managed
launchers. Capture identity at operation start. Keep an explicit
dedicated-agent grant as an override. Show redacted diagnostics in run
details.
**Alternatives considered**
Per-task ownership, fixed provider tokens, and mutable repository author
configuration do not handle accepted steering or concurrent operations.
A manual account-selection action would add unnecessary setup to each
turn.
**Roadmap alignment**
This completes the existing Multiple Human Users, MCP Tool Gateway &
Apps, Secrets Manager, and Self-healing Runs capabilities. The
implementation follows the maintainer-approved plan.
Related work: Refs #12843, Refs #12907. Existing proposals #4618 and
#8945 cover per-agent or per-worktree author configuration. This change
instead follows the accepted human instruction across runtime types.
Refs #11831 for governed personal connection delegation; this change
preserves connection audience checks and does not use standing
delegation as a personal credential fallback.
## What Changed
- Add durable, ordered identity contexts and active run references.
Preserve message authors through consolidation, steering, retries,
delegation, approvals, routines, and restart.
- Add an authenticated operation-time GitHub credential broker and
local/remote managed git and gh launchers. Keep personal tokens out of
the long-lived provider process.
- Resolve GitHub gateway and server-side Git operations through the same
responsible-person or dedicated-grant selection rules.
- Make absent and unavailable GitHub credentials non-blocking at generic
startup. Clear host and prior-person credentials. Keep anonymous Git
access where supported.
- Add run-detail identity history and the dedicated-account warning.
Keep task ownership and queue-versus-steer decisions unchanged.
- Preserve personal OAuth declarations through connection edits. Retain
exact selected grants in the gateway.
- Fix continuation races found during real acceptance: verify a warm
owner before credential rotation, and wait for bounded durable runner
suspension before the next run starts.
- Make migrations replay-safe. Retain identity through agent/run
deletion, remove it with its company, and clean terminal launcher
directories before releasing execution environments. Document
coordinated release and rollback.
## Verification
- Full workspace typecheck, build, and token gates passed. The complete
local suite passed in its normal test groups: 17,120 passing tests,
including all 143 serialized server suites. After integrating the newly
merged runner API work, full local typecheck and build passed again,
along with 890 focused integration tests. All 31 checks on the
integrated revision passed, including build, browser E2E, release
registry, canary dry run, typecheck, security and all test suites.
Greptile is 5/5 with all review threads resolved.
- Current focused checks passed: 142 native executor tests, 67 runtime
lifecycle tests, 9 durable identity tests, 75 credential/routine tests,
19 low-trust/resumption tests, and the executable migration replay test.
- Authenticated browser acceptance with two Paperclip users and two
GitHub accounts on one shared native agent passed. Real commits and
pushes followed A → B accepted steering → queued A continuation in the
same saved conversation. GitHub commit author and committer identities
matched all three operations. Both runs succeeded and task ownership
stayed unchanged.
- Real GitHub MCP calls switched from A to B after accepted steering. A
delegated subtask retained its originating identity across a server
restart.
- Disabling B's GitHub connection left ordinary work successful. Managed
gh was unauthenticated and the provider had no inherited GH_TOKEN or
GITHUB_TOKEN.
- The browser displayed run-detail diagnostics and the exact
dedicated-account warning. A final controller-restart check followed by
another-person continuation retained the conversation, selected the
correct GitHub login and Git author, and removed each terminal launcher
directory.
- Company-lifetime migration and all five previously failing CI suites
passed locally (167 tests). Same-token gateway A → B → A and six
broker/launcher boundary tests passed.
- Remote callback, launcher, sandbox, and runtime contract tests passed.
Both native and legacy Codex completed actual Daytona executions on the
integrated revision ([campaign
results](https://github.com/paperclipai/paperclip/actions/runs/34155056509)).
The remote package-manager shim staging regression also passed locally.
## Risks
- Deploy the migrations, server broker, launchers, and runner artifacts
together. Existing processes finish with their original contract. New
managed processes need the broker endpoint for GitHub operations.
- Finish or stop new managed executions before rolling application code
back. Keep the additive schema and identity history during rollback.
- Scripts that require a persistent raw GH_TOKEN must use managed git,
gh, or GitHub gateway tools. Run capabilities authorize code executing
within that run to acquire its current identity; this is not
hostile-code isolation within one execution principal. Managed commands
prevent automatic credential carryover; arbitrary code deliberately
copying a credential is outside that boundary.
- Uncertain steering acknowledgement deliberately holds new credential
acquisition until reconciliation. Already-started operations retain
their captured identity.
- GitHub private access and provider outages can still fail the specific
operation that needs them. Dedicated grant failure does not fall back to
personal access.
## Model Used
OpenAI GPT-6 through Codex assisted implementation, review, shell
execution, and browser acceptance. The exact model variant and
context-window size are not exposed in this session. Tool use included
TypeScript and Rust tests, database integration tests, GitHub CLI, and
authenticated browser control.
## 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 and their work.
> - The new runner gives agents dedicated tools for common tasks.
> - Some API operations and parameters have no dedicated tool.
> - Agents need a controlled way to find and use those operations.
> - This pull request adds API search and calls through the real server
routes.
> - Existing tools remain the preferred path. The new tools are disabled
by default.
> - Paired tests measure correctness, tool choice, cost and time.
## Linked Issues or Issue Description
**Subsystem affected**
Paperclip Runner contracts, production tool authority and the server API
catalog.
**Problem or motivation**
The runner cannot use much of the API described by the old Paperclip
skill. A generic HTTP client would also let agents bypass runner control
rules.
**Proposed solution**
Add `search_api` and `call_api`. Resolve calls from the mounted API
catalog. Use server-held, run-bound credentials. Preserve route checks
and runner lifecycle rules. Keep the tools disabled until an operator
enables selected companies.
**Alternatives considered**
A dedicated tool for every endpoint would add a large initial prompt. An
unrestricted HTTP tool would weaken authorization and replay controls.
**Roadmap alignment**
This extends the native runner tooling. The repository owner requested
this design and implementation. The roadmap and related open PRs were
checked. No duplicate API escape-hatch PR was found.
## What Changed
- Register two compact fallback tools in canonical contracts and
provider projections.
- Build deterministic API discovery from OpenAPI, mounted experimental
routes and the old skill reference.
- Execute bounded JSON, text, file and download requests through
authenticated HTTP routes.
- Recheck active runs, company access and work modes. Block runner
lifecycle, scheduling, credential and approval bypasses. Keep routine
annotation collaboration available.
- Retain mutation receipts. Report uncertain outcomes without blindly
repeating writes.
- Add a company rollout gate and a durable eval worker with complete
cost accounting checks.
- Record child-task creation in the activity log with the agent and run.
- Add contract, authorization, file, replay and real runnerd/PRP/HTTP
tests.
- Document rollout gates and paid coverage limits. The companion eval
repository retains immutable attempts and reports.
## Verification
- Final app commit `da58370524c3626a744eec20164397c5fb6ba9ef`: all 32
checks passed; the unrelated Storybook visual check was skipped.
Greptile 5/5; no unresolved review threads.
- Full Linux build and recursive typecheck passed. Repository tests were
run by project and serialized shard; all 143 serialized server suites
passed.
- Runner TypeScript: 1,599 passed, two skipped. Rust release: 451
passing test reports. Conformance and replay parity passed. The required
API check passed 837 tests, including runnerd → PRP → authority → real
HTTP.
- Bindings cannot enable API tools without the explicit deployment flag.
Unit and real-authority tests prove the default-off boundary.
- The standalone API check builds and stages its own binary. It passed
after existing staged and debug binaries were removed from the test
container.
- UI and CLI tests passed. Initial environment failures (missing jq,
Docker overlay file identity, and parallel linker memory pressure) and
focused passing reruns are retained. The macOS full runner suite has
platform-specific failures; Linux is the qualified full-check platform.
- Eval harness: 27 tests passed; existing CI discovery ran 86 tests with
two unrelated skips. Credential export rejection is tested against the
actual report command.
- Luna and OpenRouter Sonnet each passed 60 common-workflow runs: ten
workflows, three repetitions per arm, zero unnecessary API fallback.
- Sonnet passed 11 selected capability/contract cases after fixes.
Gemini passed three smoke cases. DeepSeek exceeded the 120-second limit
and remains unqualified.
- Luna's two cost flags received focused follow-up. The original flags
and a later n=1 latency flag remain visible. Sonnet had no cost or
latency increase above 20%.
- The catalog contains 785 entries; 58 were exercised across all stages.
Most operation probes remain unrun and some need additional fixtures.
Authored probes do not establish successful coverage.
- Total conservative accounted cost: $9.875960. Active paid-campaign
time: 88.16/90 minutes. No missing accounting. Later security and
harness fixes have provider-free verification; no paid validation is
claimed for those revisions.
- Inspect the [qualification
report](https://github.com/paperclipai/paperclip-evals/blob/codex/seach-call-api-tools/evals/runner-api-tools/reports/2026-09-07-production/READINESS.md)
and [verification
record](https://github.com/paperclipai/paperclip-evals/blob/codex/seach-call-api-tools/evals/runner-api-tools/reports/2026-09-07-production/verification.json).
## Risks
- This is a broad authenticated API surface. Keep the default-off gate
until an operator selects initial rollout companies.
- Paid coverage is incomplete. Small regression samples do not prove all
workflows are unchanged.
- A timeout or server failure can follow a committed mutation. The
result reports an unknown outcome and requires state inspection.
- The new definitions add prompt tokens. The report retains cost flags
and cache variation.
- No database migration is required.
- Repository rules require code-owner approval before merge. Technical
CI and automated review are complete.
## Model Used
OpenAI Codex based on GPT-6 assisted with code, tests and review. The
exact serving model ID and context window are not exposed in this
session. It used reasoning, tool calls and code execution.
Eval models: `gpt-5.6-luna` with low reasoning,
`openrouter/anthropic/claude-sonnet-5`,
`openrouter/google/gemini-3.8-flash`, and
`openrouter/deepseek/deepseek-v4-flash-0731`. Attempts retain runtime
versions, model identity, usage and source provenance.
## 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 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 is the open source app people use to manage AI agents for
work.
> - Runner direct live evals test semantic tools against a mock control
plane.
> - The first complete AWS campaign exercised 358 cells.
> - It exposed setup differences from the working full-stack harness.
> - This pull request corrects those direct-harness differences.
> - It preserves production permission defaults and the full-stack
workflow.
## Linked Issues or Issue Description
**What happened?**
Native Codex cells could not find a global codex executable. ACPX denied
unattended tool requests and lost valid provider usage receipts.
AgentCore hit a 30-second facade timeout while its worker allows 120
seconds for delivery. Several models reported a native run result
without updating the separate mock task state.
**What did you expect?**
The direct harness should use the pinned executable, explicit test
permissions, and a timeout compatible with the provider delivery
contract. Its instructions should explain which operation changes mock
task state.
**Steps to reproduce**
Run the full Runner Direct Live Protocol Evals workflow. Baseline
campaign:
https://github.com/paperclipai/paperclip/actions/runs/34059009921.
**Paperclip version**
Master at b55ce03e86. Related public
change: #12932. No duplicate fix was found.
## What Changed
- Resolve native Codex from the pinned Codex ACP dependency, as the
full-stack launcher does.
- Select explicit unattended ACPX permissions only in the isolated
direct eval harness.
- Add a configurable bounded turn-admission wait. AgentCore direct evals
use up to 125 seconds, capped by their turn budget. Other callers retain
the current default waits.
- Explain mock task-state operations separately from native run-result
reporting. Keep all scoring assertions unchanged.
- Bind the direct turn's scope to the current request, rather than stale
shared fixture notes. A provider's end-of-turn result does not authorize
an unrequested mock task completion or extra completion comment.
- Normalize the qualified Claude/Codex usage semantics without
double-counting reasoning or inventing unknown billable categories.
Forward ACPX's persisted terminal prompt-response usage for the exact
current turn; reject stale or ambiguous receipts.
- Preserve the known numeric ACPX token-counter aliases through durable
redaction. Continue redacting strings and credential-shaped values. A
regression test reproduces the previously redacted usage before
normalization.
- Add regression tests and operating documentation.
## Verification
- Passed Runner TypeScript build.
- Passed 88 focused tests across the eval request contract, provider
setup, and Runner transport.
- Passed 93 focused ACPX adapter, sidecar lifecycle, and
usage-accounting tests, including persisted receipt identity and
missing-field regressions.
- Passed 97 additional runtime-host, in-process ACPX driver, OpenCode
proxy, and MCP bridge tests.
- Reproduced the ACPX counter-redaction failure, then passed 36 Rust
decoder/normalizer and durable-redaction tests after the fix.
Credential-shaped strings and objects remain redacted.
- Native Codex get-task-context passed a real provider probe with no
global CLI dependency.
- Local ACPX probes reached a platform startup rejection on macOS;
qualification therefore used the intended AWS Linux fleet. Complete
campaign 34060573948 reached 341/358 passing with zero infrastructure
failures (up from 250/358 and 100 infrastructure failures). ACPX Claude
reached 34/35, ACPX Codex 7/8, native Codex 34/35, and AgentCore 35/35.
Remaining failures were retained in the canonical report:
https://d1p6rlowie26tp.cloudfront.net/runner-protocol-evals/campaigns/gha-34060573948-1/.
- Final complete campaign 34062394019 exercises the current-request
scope clarification with all 358 cells; results pending.
- Passed git diff --check.
- Repo-wide typecheck, build, and tests are delegated to PR CI. They
were not repeated on this machine.
## Risks
- ACPX approval is scoped to the operator-requested isolated fixture
harness. Production defaults do not change.
- AgentCore admission can wait longer, but remains within the whole-turn
budget.
- The mock fixture has no production task lifecycle service. Explicit
task-state instructions describe that boundary; they do not relax
scoring.
- This change does not modify the browser full-stack workflow or its
test fixtures.
- ACPX receipt corrections apply to all native ACPX sessions. They rely
on the pinned qualified server contracts; unknown or ambiguous receipts
stay unknown, and a final accounting-read error never overwrites the
provider's terminal result.
## Model Used
- OpenAI GPT-6 (`gpt-6-astra`) through Codex, with reasoning, tool use,
and code execution. The session does not expose an exact context-window
limit.
## 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 runs AI agents through qualified provider backends.
> - The direct live eval workflow builds one immutable Runner runtime
for every matrix cell.
> - The workflow reinstalled the packed Runner with npm.
> - That install discarded pnpm patches and selected provider
dependencies outside the qualified lock.
> - The first pnpm deployment model also placed its virtual-store marker
at the wrong level; a real deployment keeps `.pnpm` beside the scoped
Runner package.
> - AgentCore enforced the current context-aware harness but the direct
eval CLI did not supply the production v3 runtime context that harness
requires.
> - This pull request preserves the qualified dependency graph, resolves
the real deployment layout, and makes direct evals exercise the
production runtime-context contract.
> - The benefit is that live eval cells reach their provider turn with
the same artifacts and context contract that Paperclip qualified.
## Linked Issues or Issue Description
Refs: #12931
**What happened?**
The full direct live eval campaign failed every ACPX cell during
`session.open`. The portable runtime had an incorrect dependency root.
Its npm install also discarded the qualified ACP server patches.
AgentCore cells first failed because Runner enforced
`aws-agentcore-harness-v1` while the provisioned stack and eval profile
use `aws-agentcore-harness-context-v2`; after aligning that revision,
the direct eval CLI still omitted the required v3 runtime context.
**Expected behavior**
The direct eval runtime must preserve the frozen pnpm dependency graph
and patched provider bytes. Runner, server validation, OpenAPI, and the
deployed AgentCore stack must use one qualification revision. Direct
eval attempts must supply the same immutable native runtime-context
contract as production.
**Steps to reproduce**
1. Dispatch `Runner Direct Live Protocol Evals` from `master`.
2. Select an ACPX Claude, ACPX Codex, or AgentCore roster.
3. Observe a pre-turn provider bootstrap failure.
**Paperclip version or commit**
`d96452db059338b329b458ba8fe359fef72f1363`
**Deployment mode**
GitHub Actions on the RunsOn Linux x64 fleet.
## What Changed
- Build the reusable direct-eval runtime with `pnpm deploy --prod`.
- Resolve ACPX dependencies from the actual scoped-package layout of a
self-contained pnpm deployment.
- Align AgentCore configuration and qualification checks on
`aws-agentcore-harness-context-v2`.
- Materialize a minimal immutable v3 runtime context for each isolated
direct eval attempt.
- Add workflow, package-authority, runtime-context, Rust, and server
regression coverage.
- Document the qualified packaging, runtime-context, and AgentCore
revision contracts.
## Verification
- `pnpm --filter @paperclipai/paperclip-runner typecheck`
- `pnpm --filter @paperclipai/server typecheck`
- `pnpm --filter @paperclipai/paperclip-runner exec vitest run
src/live/runnerd-codex-transport.test.ts` (70 tests)
- `pnpm --filter @paperclipai/paperclip-runner exec vitest run
src/cli/eval-session-contract.test.ts` (14 tests)
- Focused Runner contract tests (36 tests)
- Focused server profile tests (47 tests)
- Focused Rust managed-provider and native-selector tests (19 tests)
- `node --test
packages/paperclip-runner/scripts/runner-protocol-eval-workflow-security.test.mjs`
- `actionlint .github/workflows/runner-protocol-live-evals.yml`
- A local `pnpm deploy --prod` produced both qualified ACP server
digests.
- A Linux reproduction of the first follow-up smoke identified the real
deployment root and the missing AgentCore runtime context.
## Risks
The AgentCore revision change rejects profiles that still use the
obsolete v1 value. This is intentional because the provisioned
context-aware harness and current eval profile use v2. Direct eval
prompts now receive the same fixed runtime-context preamble as
production, so behavior scores may move; that is the intended
qualification surface. The workflow package layout changes, but tests
assert the new entrypoint and dependency root. This change does not
modify the browser full-stack E2E workflow.
> 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. See `CONTRIBUTING.md`.
## Model Used
OpenAI Codex with GPT-5.6. The context-window size is not exposed in
this session. The model used extended reasoning, repository tools, code
execution, Docker-based Linux reproduction, and GitHub Actions
diagnostics.
## 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 (for example, `docs/...` or
`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
> - Daytona preserves a stopped sandbox filesystem, but deleting or
replacing a sandbox removes its only remote copy.
> - Warm reuse therefore improves latency but cannot be Paperclip's
durability boundary.
> - The host execution workspace must remain authoritative after every
successful turn, while same-run recovery must avoid overwriting
unexported remote work.
> - Result proposal, workspace export/merge, and terminal completion
need a durable, replayable ordering so a crash never starts a duplicate
provider turn.
> - A paid browser acceptance suite must exercise both legacy Codex and
Runner Codex for three real turns on one continuously warm Daytona
sandbox.
## Linked Issues or Issue Description
Refs #12901.
Runner Codex did not previously export successful Daytona workspace
changes back to the authoritative host workspace. That made warm reuse
depend on Daytona's remote filesystem and left deleted/replacement
sandboxes without a reliable reconstruction path. The existing paid
fixture also lacked a focused three-turn continuity case for both Codex
adapters.
## What Changed
- Persist versioned, atomic native workspace-sync descriptors and
durable seeds in `PAPERCLIP_HOME`, without credentials or a database
migration.
- Classify fresh, warm, replacement, and same-run-recovery workspace
preparation explicitly; ambiguous lease/root/digest evidence fails
closed.
- Finalize native workspace export/merge after semantic result proposal
and before run completion, with idempotent replay that never submits a
second provider turn.
- Surface legacy Codex workspace restoration failures instead of masking
them, while preserving an earlier provider error when both fail.
- Keep healthy reusable Daytona leases warm for legacy and native
adapters, stamp finalized workspace generations, and retain existing
cleanup behavior for per-turn or unhealthy leases.
- Preserve Runner Codex's provider process/session across warm turns,
including bounded post-terminal tail draining and exact authority
rotation.
- Add the exact paid `daytona-warm-continuity` matrix:
- `legacy-codex × daytona × warm-three-turn`
- `runner-codex × daytona × warm-three-turn`
- Drive all three turns through the browser, verify ordered file
continuity and stable lease/workspace/runtime identities, capture
per-turn timings, and delete the sandbox immediately after assertions.
- Document `pnpm test:e2e:runner -- --suite daytona-warm-continuity`; no
package script was added.
## Verification
- `pnpm typecheck` — passed, including migration safety (no migration
added)
- Focused server/runner Vitest coverage — 144 passed
- `pnpm test:e2e:runner:unit` — 114 passed
- `pnpm test:e2e:runner:typecheck` — passed
- `pnpm --filter @paperclipai/paperclip-runner test:codex` — 66 passed,
1 helper ignored
- `native-session-executor.test.ts` — 139 passed, including safe
fail-closed cleanup after remote runner identity capture failure
- Paid local browser acceptance, exact post-rebase Linux/amd64 runner
binary:
- Runner Codex — passed in 1.7m; 3 runs; lease outcomes `created,
resumed, resumed`; 10/10 matchers; cleanup passed
- Legacy Codex — passed in 2.7m; 3 runs; lease outcomes `created,
resumed, resumed`; 10/10 matchers; cleanup passed
- [Protected paid GitHub Actions
campaign](https://github.com/paperclipai/paperclip/actions/runs/34026735033)
against `7da42a91b95fa7fb2df126668ef7e37afb3b2b9d` — passed 2/2:
- Runner Codex — 3 runs; lease outcomes `created, resumed, resumed`;
evidence and cleanup passed
- Legacy Codex — 3 runs; lease outcomes `created, resumed, resumed`;
evidence and cleanup passed
- Merge/enforcement, S3 history, and Pages publication jobs passed
- Paid result artifacts were scanned for both provider credentials;
neither secret was present.
- Current PR checks — 31 passed, 1 expected Storybook skip; Greptile
5/5; Superagent security scan passed
- `git diff --check origin/master...HEAD` — passed
- Confirmed no `package.json`, lockfile, migration, or SQL changes.
## Risks
- Workspace synchronization now sits on the terminal-success path, so a
remote export failure deliberately prevents false success. Retryable
state retains its lease/seed; loss of the only unexported remote copy
fails closed.
- Warm provider reuse has strict identity and quiescence checks.
Mismatched or ambiguous evidence blocks reuse rather than risking
concurrent provider work.
- The paid suite incurs Daytona and Codex cost only in the existing
protected scheduled/manual workflow and explicitly destroys its sandbox
after each cell.
## Model Used
OpenAI Codex with GPT-5 agentic reasoning, repository inspection, real
browser E2E execution, Rust/TypeScript test execution, and GitHub
Actions diagnostics.
## 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 issue or described the issue
in-PR
- [x] I have not referenced internal/instance-local Paperclip issues or
links
- [x] My branch name contains no internal ticket id
- [x] I have run tests locally and they pass
- [x] I have added or updated tests where applicable
- [x] I have documented the dedicated suite invocation without adding a
package script
- [x] I have considered and documented risks above
- [x] All Paperclip CI gates are green on the current revision
- [x] Greptile is 5/5 with no open P2s, recommendations, or follow-ups
on the current revision
- [x] I will address all reviewer comments before requesting merge
## Thinking Path
> - Paperclip is the open source app people use to manage AI agents for
work.
> - The Runner executes agents through native and managed provider
drivers.
> - The direct live eval layer had drifted from the current Runner
contracts.
> - The old local workflow did not provide a complete parallel campaign
or durable report history.
> - The Runner also needed current native OpenCode and OpenRouter
qualification.
> - This pull request restores the direct campaign, corrects the runtime
gaps that the campaign found, and adds safe hosted Evalbook history.
> - The benefit is repeatable model comparison against an immutable
Runner and eval source revision.
## Linked Issues or Issue Description
Refs #11297
Refs #11634
**What existing behavior does this improve?**
This improves the direct live `paperclip-runner` eval workflow, provider
execution contract, and static Evalbook reporting path.
**Current behavior**
The direct evals do not have one maintained full campaign on current
`master`. OpenCode has no qualified multi-model OpenRouter roster.
Parallel provider bursts can compact committed events before the
transport observes them. Local reports do not have a separate safe S3
history index.
**Proposed behavior**
Run one immutable roster-plus-case matrix. Use the shared paid AWS
runner fleet. Keep raw artifacts access-controlled. Publish a sanitized
canonical Evalbook report under the separate `runner-protocol-evals` S3
prefix. Keep immutable campaign directories plus root history, latest,
and latest-green pointers.
**Reason and benefit**
Maintainers can compare native Codex, native OpenCode, ACPX, Claude
Managed, and AWS AgentCore behavior over time. They can inspect failures
without mixing this direct protocol layer with browser full-stack E2E.
**Breaking changes**
None. The new workflow and S3 prefix are additive. The existing Runner
full-stack E2E workflow and report remain separate.
## What Changed
- Added a trusted two-shard direct live workflow for up to 393
roster-plus-case cells.
- Reused the numeric actor allowlist, protected paid environment, and
RunsOn fleet controls from Runner full-stack E2E.
- Added immutable Runner and eval revision resolution, exact credential
boundaries, bounded retries, and cost ceilings.
- Added a public report projection that removes sessions, transcripts,
tool payloads, state, traces, raw failures, remote profile identities,
and credential-shaped values.
- Added additive S3 history under `runner-protocol-evals`, with
immutable campaigns and mutable root index pointers.
- Added native OpenCode model injection and current OpenRouter pricing
contracts.
- Fixed direct eval completion, workflow execution, semantic discovery,
warm-attach state reset, executable binding, and event-burst handling.
- Kept Runner browser full-stack E2E behavior and publication separate.
- Documented local and hosted direct eval operation.
## Verification
- `pnpm --filter @paperclipai/paperclip-runner
test:runner-protocol-eval-publish` — 15 passed.
- `pnpm --filter @paperclipai/paperclip-runner build:typescript` —
passed.
- `actionlint .github/workflows/runner-protocol-live-evals.yml
.github/workflows/runner-full-stack-e2e.yml` — passed.
- Local current matrix at the revision in
[paperclip-evals#17](https://github.com/paperclipai/paperclip-evals/pull/17)
— 323 cells across 10 enabled configurations completed.
- Final local current matrix — 269 passed, 11 behavior failures, and 43
expected macOS-only ACPX platform failures.
- Targeted Runner checks — 13/13 eval-session tests, 15/15
publisher/security tests, and package typecheck passed; complete PR CI
is green, including all browser E2E shards.
## Risks
- Paid live campaigns can consume provider budget. Actor authorization,
exact per-cell ceilings, protected environments, and explicit schedule
enablement bound this risk.
- Public reports can leak provider data. The workflow publishes only a
separately projected report and validates every file before upload.
- The new workflow cannot publish until it is present on the default
branch. This pull request does not change the existing
`runner-full-stack-e2e` publication path.
- The campaign is large. It uses two GitHub matrices and caps combined
concurrency at the shared fleet limit.
> 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 on GPT-5. The exact deployment ID and context-window size
are not exposed. The model used reasoning, code editing, browser
inspection, repository tools, and live provider 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
## Thinking Path
> - Paperclip is the open source app people use to manage AI agents for
work.
> - Managed GitHub connections resolve a responsible user's or dedicated
agent's identity into an audited, run-scoped credential projection.
> - An enrolled instance could retain a hidden managed setup method
after Cloud stopped advertising it, producing a blank, disabled setup
step.
> - Native runner processes also dropped the resolved GitHub projection
before the provider shell, so `gh` and Git could not use the selected
identity in Daytona.
> - Daytona already provides the outer isolation boundary. Applying
Codex's inner Linux sandbox there both duplicated containment and failed
because nested user namespaces are unavailable.
> - This change repairs setup fallback, carries only the bounded GitHub
projection across each runner boundary, and allows only a
controller-selected managed sandbox transport to act as the outer
sandbox.
## Linked Issues or Issue Description
**What happened?**
An enrolled self-hosted instance could show a blank GitHub setup step
when its managed profile was unavailable. Separately, a native Codex run
in Daytona could resolve a managed GitHub connection on the Paperclip
host but lose it before the provider shell. Once projected, Codex's
nested sandbox failed before commands could run because Daytona does not
expose the user-namespace operation used by the inner sandbox.
**Expected behavior**
Setup must select an advertised customer method when the managed method
is unavailable. A Daytona run must receive the exact managed GitHub
identity selected for that run, support `gh` and HTTPS Git, and rely on
Daytona as its outer sandbox without weakening local or SSH execution.
**Steps to reproduce**
1. Enroll a self-hosted instance while Cloud does not advertise the
managed GitHub profile and open GitHub setup.
2. Observe the blank second step and disabled action.
3. Configure a native Codex agent with a Daytona environment and a
responsible-user GitHub grant.
4. Run `gh api user` or HTTPS Git from the agent shell.
5. Observe missing GitHub environment projection or nested-sandbox
startup failure.
**Paperclip version or commit**
The setup bug reproduces on `1dceee9a4`; the runner proof was developed
from the same branch and verified at the latest head below.
**Deployment mode**
Self-hosted Paperclip enrolled with Paperclip Cloud, using the Daytona
sandbox-provider plugin and native Paperclip runner.
## What Changed
- Wait for connector enrollment hydration, retain a hidden managed
method only while enrollment is needed, and otherwise select an
advertised customer fallback.
- Add a single bounded GitHub credential-environment projection for
`GH_TOKEN`, `GITHUB_TOKEN`, the process-only Git helper token, GitHub
commit identity, and at most 32 controller-generated Git config entries.
- Forward that projection through the durable controller, Codex
app-server transport, and Rust provider child without placing token
values in arguments or config.
- Allow Codex shell inheritance only for the exact projected GitHub keys
and enable provider network access only when the managed credential
exists.
- Derive outer-sandbox authority exclusively from a managed `sandbox`
transport; strip the same flag from configured, host, local, and SSH
environments.
- Define a named external-sandbox permission profile that Codex resolves
to `dangerFullAccess` for default-mode Daytona turns while plan mode
remains read-only.
- Add regression tests for setup fallback, credential projection,
local/SSH/sandbox authority separation, provider forwarding, and
permission-profile selection.
## Verification
- `pnpm exec vitest run ui/src/pages/apps/AppsConnect.test.tsx` — 96
passed.
- `pnpm exec vitest run src/drivers/codex/codex-security-config.test.ts
src/drivers/codex/app-server-transport.test.ts
src/control-plane/durable-prp-control-plane.test.ts` from
`packages/paperclip-runner` — 31 passed.
- Focused native-session executor tests — 3 passed.
- `pnpm --filter @paperclipai/paperclip-runner typecheck` — passed.
- `pnpm --filter @paperclipai/server typecheck` — passed.
- `cargo test --manifest-path
packages/paperclip-runner/runner/Cargo.toml --locked -p
paperclip-runner-core --lib` — 194 passed.
- Real Codex app-server configuration probe accepted
`paperclip-runner-external-sandbox` and reported
`sandbox.type=\"dangerFullAccess\"` while using the named profile.
- [Signed Daytona image
workflow](https://github.com/paperclipai/paperclip/actions/runs/33995270328)
built commit `86571f7997e7100e47bd131aac1f1e773112a0ce`; the isolated
environment was pinned to
`sha256:ecef21105f8de382d75787e59439d936be239b77ae74a31c8ed3a17cde39b023`.
- Live isolated Daytona proof passed: the three projected token
variables were non-empty and equal; the host-scoped Git credential
helper returned the same token without printing it; `gh api user`
resolved `cryppadotta`; authenticated `git ls-remote
https://github.com/paperclipai/paperclip.git HEAD` returned
`1dceee9a4e75b13456760bb54c752deb2dba1d79`; no repository mutation
occurred.
- The persisted 28,476-byte run log contains no GitHub token shape,
bearer header, credential-bearing URL, or private-key marker.
- Latest-head pull-request CI and reviews provide the remaining
full-suite gate.
## Risks
- This deliberately gives shell Git and `gh` access to the run's
resolved GitHub identity. It is the audited class-3 behavior required by
the GitHub connection design and is outside per-tool Ask-first controls.
- The credential source is the trusted broker projection, which
overwrites configured environment values. The helper is scoped to HTTPS
`github.com`, revalidates protocol and host, and never places its token
in command arguments, URLs, or files.
- Managed Daytona sandboxes become the containment boundary for
default-mode provider commands. Local and SSH targets retain the inner
Codex workspace sandbox, and plan mode remains read-only everywhere.
> 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, with reasoning, browser control, shell access,
and code execution. The product did not expose the context-window size.
## 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 agent work and the execution state for each
task.
> - Remote agents run in sandbox environments such as Daytona.
> - Daytona keeps files while a sandbox is stopped, but deletion removes
those files.
> - Runner Codex did not copy successful remote workspace changes back
to the host workspace.
> - A warm sandbox could therefore hide data loss until Daytona replaced
or deleted the sandbox.
> - This pull request makes the host workspace durable after every
successful turn and keeps verified reusable sandboxes warm.
> - The benefit is reliable multi-turn work across warm reuse, restart,
stop, and sandbox replacement.
## Linked Issues or Issue Description
**What happened?**
A successful native Codex turn in Daytona could leave workspace changes
only in the remote sandbox. A later warm turn appeared to work because
it reused that filesystem. A replacement sandbox could start from stale
host data and lose the successful changes.
**Expected behavior**
Paperclip must merge each successful remote turn into the authoritative
host workspace before it completes the run. A verified warm lease may
reuse its remote files. A replacement lease must reconstruct the exact
durable workspace seed.
**Steps to reproduce**
1. Run Codex in a reusable Daytona environment.
2. Write a file during one successful turn.
3. Replace the Daytona sandbox before the next turn.
4. Observe that the next turn can start without the prior file on the
unpatched code.
Related remote workspace foundation: #10070.
## What Changed
- Added explicit `host_current`, `durable_seed`, and `adopt_remote`
workspace preparation modes.
- Added atomic, versioned native workspace descriptors and seed archives
under `PAPERCLIP_HOME`.
- Added real native sandbox export and three-way host merge before
terminal result completion.
- Added workspace-only recovery after a proposed result. Recovery does
not submit another provider turn or consume the provider retry budget.
- Added fail-closed handling when a sandbox with unexported changes is
gone.
- Kept healthy reusable Daytona sandboxes started for legacy Codex and
Runner Codex.
- Kept the Runner Codex process and provider session across verified
warm turns.
- Added the paid `daytona-warm-continuity` browser suite. It contains
exactly the legacy Codex and Runner Codex cells. Each cell performs
three measured turns.
- Documented `pnpm test:e2e:runner -- --suite daytona-warm-continuity`.
No package script was added.
- Added no database migration. The metadata format is backward
compatible and idempotent.
## Verification
- `pnpm typecheck`
- `pnpm test:e2e:runner:unit` — 114 passed
- Native workspace, finalizer, session, and environment tests — 232
passed
- Daytona provider tests — 150 passed
- Workspace staging and merge tests — 98 passed
- Runner transport tests — 63 passed
- Legacy Codex restore tests — 5 passed
- Rust format and compile checks pass through root typecheck
- The paid Daytona suite was not run locally because the required
Daytona, OpenAI, and immutable image credentials are not present.
## Risks
- The main risk is an incorrect workspace identity or merge after a
crash. Durable descriptors bind the run, workspace, lease, provider
lease, local root, remote root, and baseline digest. Ambiguous evidence
fails closed.
- The host merge may conflict with concurrent host edits. The existing
three-way merge and exclusion rules handle this case and surface
failures.
- A deleted sandbox cannot recover unexported bytes. Paperclip now
blocks with `workspace_sync_out_unrecoverable` instead of reporting
success or rerunning the provider.
- There is no database migration. Descriptor writes and recovery are
atomic and idempotent.
## Model Used
OpenAI Codex with GPT-5. The run used agentic reasoning, repository
inspection, code execution, test execution, Git, and GitHub CLI tools.
## 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
## Thinking Path
> - Paperclip manages AI agents and their work.
> - The runner executes agent turns on local and remote providers.
> - A remote per-turn session must save its state before Paperclip
releases its sandbox.
> - The session runtime returned after 100 milliseconds while the remote
checkpoint still ran.
> - The next turn also checked the local state path instead of the
verified remote backup.
> - This pull request waits for the bounded remote close and accepts
only a verified suspended backup.
> - The benefit is reliable multi-turn execution without weaker identity
checks.
## Linked Issues or Issue Description
**What happened?**
A successful remote agent turn released its sandbox before the runner
saved the verified continuation backup. The next turn failed with
`runner_state_identity_mismatch`.
**Expected behavior**
Paperclip must finish the bounded remote checkpoint before it releases
the sandbox. A later turn must validate and restore the digest-matched
suspended backup.
**Steps to reproduce**
1. Run a native ACPX Claude Plan test in a non-reusable Daytona sandbox.
2. Reject the first plan to start a second turn.
3. Observe that the second turn fails before provider execution.
**Paperclip version or commit**
The failure reproduced at `13775a90b078ff64872f50961ea1b83d575e7bc6`.
**Deployment mode**
GitHub Actions with a Daytona sandbox.
## What Changed
- Wait for the internally bounded remote runner close and checkpoint
before the host returns.
- Preserve the existing short cleanup bound for other providers.
- Validate remote continuation lifecycle from a complete digest-verified
backup when local runner state is absent.
- Keep corrupt, non-suspended, mismatched, and unverified state
fail-closed.
- Make native Plan completion and accepted-Plan wake prompts
deterministic.
## Verification
- A prior 45-cell local campaign passed 44 cells. The only failure was
the OpenCode Plan prompt variance fixed here.
- A focused OpenCode local Plan rerun passed.
- ACPX Claude Daytona message and question cells passed.
- Focused regressions cover delayed checkpoint close and verified remote
backup lifecycle.
- GitHub Build and the focused ACPX Claude Daytona Plan cell will
validate this exact head.
## Risks
Remote runnerd sessions now wait for their internally bounded
close/checkpoint path before returning; generic provider cleanup retains
the existing 100 millisecond bound. Durable run success still cannot be
reversed. The environment release guard still blocks sandbox destruction
when no verified backup stamp exists.
## Model Used
OpenAI Codex, GPT-5.6, extended reasoning, with code execution and
GitHub Actions inspection.
## 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 task
id
- [ ] I have run tests locally and they pass
- [x] I have added or updated tests where applicable
- [ ] 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 findings
- [ ] 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.
> - Paperclip Runner owns durable provider sessions and streams their
work to the control plane.
> - Pull request #12845 added native restart recovery for live and dead
local runners.
> - A real browser test found three live-adoption gaps after that pull
request merged.
> - Lazy runner process ownership was not always stored before restart.
> - The old controller did not release its PRP authority without closing
the provider turn.
> - Reconnect events could arrive before the active provider turn was
restored.
> - This pull request closes those gaps and proves the same turn
completes after a UI hot restart.
## Linked Issues or Issue Description
Refs #12845
Related search results: #12646 covers indeterminate command results
after a runner restart. It does not cover controller adoption or
active-turn rebinding. No open duplicate pull request was found.
## What Changed
- Store lazy runnerd process ownership after provider session creation,
read, and resume.
- Detach native PRP controller authority during coordinated hot
shutdown. Keep the live provider turn running.
- Restore the exact checkpointed provider session when bounded PRP
identity events have been compacted.
- Restore the active provider turn before reconnect events are replayed.
This prevents `turn_binding_mismatch`.
- Keep exact live ownership by the current controller out of generic
orphan recovery.
- Add driver, transport, and server regression tests for these paths.
## Verification
- Ran 12 Codex driver lifecycle tests.
- Ran 53 runnerd transport tests.
- Ran 143 recovery and orphan-reaper server tests.
- Ran all 8 real-process restart recovery scenarios.
- Ran all 96 existing runner E2E unit tests.
- Ran runner TypeScript typecheck.
- Ran server TypeScript typecheck.
- Ran the migration replay test and migration safety checks.
- Tested the board UI on an isolated local instance. A real local
Codex-backed turn entered a 120-second terminal wait. The UI `Restart
now` action replaced the server and kept the same runner PID, process
start time, run ID, native session ID, runner ID, provider session ID,
and active turn. The original turn then completed.
- Confirmed one heartbeat run, no retry row, one result, one
proposed-result event, one terminal event, no protocol errors, no active
recovery state, and no surviving runner or provider process.
## Risks
- A live runner can continue provider work while no server owns the
control route. Recovery fails closed when the process fingerprint or
durable identity is ambiguous.
- Provider identity can be restored from the database only for an exact
verified adoption claim. An authenticated live `session.snapshot`
validates that identity before the driver can resume.
- The new detach path applies only to native sessions that expose
restart detachment. Other adapters keep their existing shutdown
behavior.
- This follow-up does not change the database migration or
`package.json`. The migration in #12845 remains replay-safe through `ADD
COLUMN IF NOT EXISTS` and its embedded-Postgres idempotence test. The
dedicated real-process command remains in `doc/DEVELOPING.md`.
## Model Used
- OpenAI Codex based on GPT-5. The exact serving build and
context-window size are not exposed. The run used extended reasoning,
repository tools, shell execution, and in-app 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 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
## Thinking Path
> - Paperclip is the open source app people use to manage AI agents for
work.
> - The Paperclip Runner keeps durable run and provider state outside
one server process.
> - A server restart can leave that runner alive or can interrupt it
after a provider checkpoint.
> - The old startup path used handoff intent and PID evidence, but it
did not reconstruct native ownership.
> - That gap could block the issue, create a replacement run, or start
duplicate provider work.
> - This pull request adds durable same-run recovery for coordinated and
uncoordinated restarts.
> - The benefit is exact recovery of the run, runner, session, provider,
steering, and finalization state.
## Linked Issues or Issue Description
Refs #9628. That pull request added earlier local-adapter hot-restart
work. This change adds native PRP authority reconstruction and same-run
provider resume.
Refs #10935. That pull request handles missing hot-restart snapshots.
This change also supports hard restarts with no snapshot.
Refs #11624. That pull request prevents unsafe retry after an adopted
legacy process exits. This change reconciles native terminal evidence
before provider recovery.
Refs #12070. That pull request improves process liveness checks. This
change also binds recovery to a process-start fingerprint and fails
closed on ambiguity.
**What happened?**
The server could record hot-restart intent, but startup did not rebuild
native runner ownership. A live runner could not re-register its PRP
authority. A dead runner could not resume the exact native and provider
session on the same heartbeat run. Generic recovery could then block the
issue or create replacement work.
**Expected behavior**
A live native runner must reconnect with the same PID and logical
identities. A dead runner must resume the same durable session and
heartbeat run with only a new operating-system PID. A proposed or
terminal result must finalize once before any provider turn starts.
Ambiguous process or session evidence must stay blocked without a signal
or duplicate spawn.
**Steps to reproduce**
1. Start a Paperclip Runner heartbeat and wait for an active provider
turn.
2. Restart only the Paperclip server, with or without a hot-restart
marker.
3. Observe that the old startup path does not reconstruct the native
control-plane authority.
4. Kill both the server and runner after a provider checkpoint.
5. Observe that the old path cannot resume the exact native session on
the original heartbeat run.
**Paperclip version or commit**
The defect was reproduced from commit
`1991f31fd53e7f7794d5c2e4b93be384ade2b41d`. This branch is rebased onto
the current `master`.
**Deployment mode**
Local development and self-hosted server deployments that use the local
Paperclip Runner.
## What Changed
- Added correlated hot-restart requests and version-compatible native
handoff fields.
- Added controller boot identity, process-start identity, controller
generation, recovery state, request id, and bounded history to the
native finalization ledger.
- Added transactional recovery claims for live-runner reattach,
dead-runner resume, and incomplete bootstrap.
- Added fail-closed ownership takeover rules and process identity
validation.
- Added live runner adoption to the local runner transport without a
duplicate spawn.
- Added same-run provider checkpoint resume and legacy retry-row
compatibility.
- Reconciled proposed and terminal results before runner or provider
recovery.
- Bound the HTTP and PRP listener before startup recovery and delayed
scheduling and generic reapers until classification completes.
- Added restart-aware health diagnostics, run-log recovery transitions,
durable runner diagnostics, and bounded shutdown finalizer draining.
- Moved restart-survivable diagnostics into runner-owned, pre-redacted
bounded writes; raw stdout and stderr are never persisted.
- Added process-start fencing for controller, runner, and provider PIDs;
startup classifies every candidate without an implicit cap.
- Added crash-recoverable, contention-safe development restart-request
coordination and failed-startup listener cleanup.
- Added a credential-free real-process restart suite for eight restart,
scale, and identity scenarios.
- Documented native restart operation, persistence, diagnostics, and
verification.
## Verification
- The documented native restart commands passed. They ran eight
real-process/database recovery scenarios and the live runner adoption
transport test.
- Native executor tests passed: 111 tests.
- Heartbeat recovery tests passed: 124 tests.
- Hot restart, health, and shutdown tests passed: 52 tests.
- The broader affected server suite passed: 350 tests.
- Focused native recovery and startup tests passed: 49 tests.
- Runner transport and control-plane tests passed: 63 tests.
- Runner-owned diagnostic tests passed for write-time bounding,
credential redaction, private file modes, and raw stream
non-persistence.
- Development restart coordination tests passed: 11 tests.
- Database migration checks and the partial-application/replay
regression test passed.
- Server, database, and Paperclip Runner typechecks passed.
- `git diff --check` passed.
- Full Paperclip PR CI passed, including build, canary, all five general
server shards, all five serialized server shards, all three browser E2E
shards, workspace suites, and release-registry verification.
- Greptile completed at 5/5 with no outstanding findings,
recommendations, follow-ups, or open review threads.
## Risks
- Moderate risk. This changes startup ordering and ownership transfer
for active native runs.
- The migration adds nullable columns and does not rewrite existing
rows.
- Recovery fails closed when process or durable session identity is
incomplete or contradictory.
- The first implementation supports the local Paperclip Runner. Remote
targets keep their existing behavior.
- The real-process suite covers cleanup and asserts that no runner or
provider process survives each test.
> 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 with GPT-5. The runtime did not expose a more specific
model revision or context-window size. Repository editing, shell
execution, database tests, and real-process test 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 (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
## 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 ACPX runtime host tests manage credentials and sandbox
operations.
> - These tests poll operations that can join quarantine recovery.
> - Recovery uses real backoff and directory synchronization, so the
default poll deadline can expire while the operation makes progress.
> - Three tests also stage a contender before the kernel lease release
completes.
> - This pull request binds every relevant poll and staging call to the
documented retry envelope.
> - The benefit is more stable tests and error output that names the
last observed cause.
## Linked Issues or Issue Description
**What happened?**
Under concurrent test load, ACPX runtime host tests failed while
credential recovery still made progress. Three tests also saw an active
lease after they removed `auth.json`.
**Expected behavior**
The tests must wait for the documented retry envelope before they report
a failure. They must stage a contender only after the credential lease
becomes available.
**Steps to reproduce**
1. Run `npx vitest run src/drivers/acpx/` from
`packages/paperclip-runner`.
2. Run the suite under high concurrent load.
3. Observe intermittent timeout or active-lease failures in
`runtime-host.test.ts`.
**Paperclip version or commit**
`865b4854fb44d3689f1c0ff17e3e715d52aaea73` base commit.
**Deployment mode**
Built from source.
**Installation method**
Built from source.
**Agent adapter(s) involved**
ACPX Codex runtime host tests.
**Database mode**
Not database-related.
**Access context**
Unclear / not applicable.
**Node.js version**
Not recorded in the handoff.
**Operating system**
Not recorded in the handoff.
**Relevant logs or output**
Under concurrent load, the failure included `Timed out in waitFor!`
after 1157 ms and `Managed Codex credential home already has an active
lease`.
**Relevant config (if applicable)**
Not applicable.
**Additional context**
The change touches test code only. It adds no test, removes no test, and
weakens no assertion. The file keeps 28 tests and 146 assertions.
## What Changed
- Add a test-local wait helper with an explicit 10-second deadline.
- Apply the helper to every credential and sandbox poll in
`runtime-host.test.ts`.
- Report the last observed error when a poll reaches its deadline.
- Guard the three credential staging calls that could race with lease
release.
- Set a 20-second timeout on tests that use the long wait.
## Verification
- `npx vitest run src/drivers/acpx/runtime-host.test.ts` passes all 28
tests on the change branch.
- A 40-run concurrent comparison produced zero `runtime-host.test.ts`
failures on the change branch.
- The base comparison produced 13 `runtime-host.test.ts` failures across
40 runs.
- The broader ACPX suite still has a separate
`codex-credentials.test.ts` flake on both arms.
- CI and Greptile results will provide the remaining merge checks.
## Risks
Low risk. The change affects test synchronization only. It increases
selected test wait limits and does not change product behavior.
## Model Used
OpenAI Codex, GPT-5. The model used tool calls and code execution. The
runtime did not provide a context window value.
## 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 recorded the separate ACPX suite
flake 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>
> - Paperclip is the open source app people use to manage AI agents for
work
> - Paperclip uses local adapters to connect agent sessions to the
control plane
> - The Codex adapter emits turn events from response and notification
channels
> - A terminal notification can arrive before the turn/start response
> - This pull request gates the terminal event on turn.accepted
> - The result keeps the event order stable for consumers and tests
## Thinking Path
> - Paperclip is the open source app people use to manage AI agents for
work
> - Paperclip uses local adapters to connect agent sessions to the
control plane
> - The Codex adapter emits turn events from response and notification
channels
> - A terminal notification can arrive before the turn/start response
> - This pull request gates the terminal event on turn.accepted
> - The result keeps the event order stable for consumers and tests
## Linked Issues or Issue Description
**What happened?**
The Codex harness session emitted `turn.accepted` only after the
`turn/start` response resolved. A terminal notification could arrive
before that response and reach consumers first.
**Expected behavior**
The Codex driver must emit `turn.accepted` before any terminal event for
the same turn.
**Steps to reproduce**
1. Start a Codex harness session.
2. Keep the `turn/start` response pending.
3. Send `turn/started` and `turn/completed` notifications.
4. Observe the event order.
**Paperclip version or commit**
`afbcd28dae9e51108738c4258929b95ca359186c`
**Deployment mode**
Built from source with the Codex driver test harness.
**Agent adapter(s) involved**
Codex.
## What Changed
- Add session state that tracks a pending `turn/start` operation.
- Resolve the state when `turn/start` succeeds or fails.
- Wait for that state before the terminal notification handler emits its
event.
- Add a regression test that delivers a terminal notification while
`turn/start` remains pending.
## Verification
- The regression test failed 5 of 5 times before this change and passed
5 of 5 times after it.
- The Codex driver suite passed 189 of 189 tests.
- The affected live transport test file passed 46 of 46 tests on 10
consecutive runs.
- The TypeScript check exited with status 0.
- Continuous integration must pass before merge.
## Risks
The change affects only Codex turn event ordering. It adds no sleep,
retry, or timeout. The main risk is a provider path that does not settle
`turn/start`; existing provider response handling still controls
completion.
## Model Used
OpenAI GPT-5. The exact deployment identifier is not exposed in this
environment. Tool use and code execution assisted this change.
## 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 and their work.
> - The runner starts ACPX sessions and controls their resources.
> - An aborted admission can leave sandbox preparation active after the
opening promise rejects.
> - Test teardown can then remove the sandbox directory before that work
ends.
> - This pull request retains and observes each unfinished admission
stage.
> - The change gives runtime resources and temporary directories one
deterministic cleanup owner.
## Linked Issues or Issue Description
**What happened?**
Under full test load, an aborted admission test can end before sandbox
preparation settles. Test teardown then removes the temporary session
directory. The active preparation can report an unhandled `ENOENT`
error.
**Expected behavior**
An aborted admission must observe and retain all active preparation
work. Test teardown must wait until that work settles.
**Steps to reproduce**
1. Run the complete `@paperclipai/paperclip-runner` test suite under CI
load.
2. Abort runtime admission during credential or sandbox preparation.
3. Observe an intermittent test timeout or an unhandled
missing-directory error.
**Paperclip version or commit**
The failure occurred on a branch based on commit `b1f4910ee`. This fix
is based on current `master` commit `4d30efa8e`.
**Deployment mode**
The failure occurred in GitHub Actions on a source build.
## What Changed
- Retain each unfinished abortable admission stage in the global
runtime-host cleanup set.
- Notify the embedding lifecycle when an aborted stage needs deferred
cleanup.
- Make test teardown abort and await all active opening and cleanup
promises before directory removal.
- Replace time-based stage detection with exact deferred stage signals.
- Add a deterministic regression test for an abort during sandbox
preparation.
## Verification
- Ran the focused runtime-host file in 20 separate processes. All 20
runs passed without an unhandled error.
- Ran `pnpm --filter @paperclipai/paperclip-runner exec vitest run
src/drivers/acpx/runtime-host.test.ts` after the rebase. All 27 tests
passed.
- Ran `pnpm --filter @paperclipai/paperclip-runner check:all` after the
rebase. The full command passed.
- The final TypeScript test stage passed 127 files and 1,490 tests. All
Rust checks, tests, and parity checks passed.
- Greptile reviewed two heads. The final review is 5/5 with no open
comments.
- All latest-head CI and security checks passed. One unrelated workspace
test passed on its permitted rerun.
## Risks
- Risk is low. An aborted stage now delays final runtime-host cleanup
until its active operation settles.
- A stage that never settles can delay embedding shutdown. The existing
stage operations have bounded or controlled owners.
- The regression test holds sandbox preparation and confirms the new
cleanup order.
> 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 with GPT-5 assisted this change. The environment did not
provide the exact deployment ID or context size. The model used
reasoning, shell tools, code editing, and 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
- [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 control plane for agents that perform work.
> - Paperclip Runner connects durable provider sessions to individual
task runs through PRP.
> - Provider continuity and per-run authority are different lifetimes.
> - The existing implementation mixed those lifetimes and lost event
metadata between provider frames, runnerd, persistence, API
sanitization, and the task thread.
> - That caused failed continuation, missing progress and Plans,
duplicate replies, hidden failures, and unsafe recovery.
> - This repair gives every heartbeat fresh authority, preserves
qualified provider-session continuity, and restores one lossless
presentation path without changing direct adapters.
## Linked Issues or Issue Description
**What happened?**
A second native heartbeat could reuse tickets, leases, command receipts,
sequence state, and run identity from the first heartbeat. Provider
phase and item identity could be lost before the UI read them. Redaction
could corrupt protocol discriminators while still missing malformed
credential tails. The task thread could fold progress into the final
response, hide failures, or show more than one final answer. Native
Codex also exposed approval modes that do not yet have a durable
approval bridge.
**Expected behavior**
Each heartbeat uses a new PRP authority epoch. Codex and OpenCode
preserve exact qualified provider sessions; ACPX emits an explicit
continuity event when its qualified process-replacement policy is used.
Every accepted provider event is presented, classified as internal, or
surfaced as unsupported. The task page shows chronological progress,
reasoning summaries, activity, Plans, interactions, terminal failures,
and exactly one final reply. Direct adapters retain their existing path.
**Steps to reproduce**
1. Enable the unified experimental Paperclip Runner setting.
2. Create a local native Codex, OpenCode, ACPX Claude, or ACPX Codex
agent.
3. Run response, Plan, structured-question/resume, restart,
cancellation, and failure scenarios.
4. Reload the task while active, waiting, failed, and settled.
5. On the old implementation, observe stale run authority, missing
classifications, incomplete output, or duplicated/folded replies.
**Paperclip version or commit**
The repair is based directly on `master` at
`87d05e194b643810d16d20612115acd01d735d43`.
**Deployment mode**
Local development with the embedded database.
Related work: Refs #12616, #12646, #12666, #12685, and #12700.
## What Changed
- Rotates PRP control-plane, outbox, ticket, lease, command, receipt,
and sequence authority for each heartbeat while carrying forward only a
validated provider-session identity.
- Reads `control-plane-state.json`, validates both durable schemas and
lifecycle values, resumes coherent current runs, archives qualified
settled authority, and quarantines malformed or mismatched scoped state
without moving ambiguous live legacy state.
- Preserves Codex provider phase and stable item identities so
commentary remains progress and only `final_answer` becomes final.
- Adds raw OpenCode HTTP/SSE boundary coverage and canonical reasoning
lifecycle mapping.
- Makes ACPX normalization lossless for visible reasoning, tool
lifecycle metadata, stable bounded identities, Plan revisions,
structured requests, failures, and qualified process replacement. Only
the compatible terminal assistant message is promoted as final.
- Applies schema-aware redaction before generic JWT-shaped detection and
scans every diagnostic string leaf. Malformed raw/escaped quoted
credential tails are redacted in both server and durable Rust state.
- Restores snapshot-style chronological task presentation, expandable
tool activity, inline Plan cards, visible waiting/resume/cancel/failure
states, and exactly one final answer.
- Makes `never` the only qualified native Codex permission mode and
rejects unsupported persisted native modes with remediation. OpenCode
and ACPX policies remain intact.
- Keeps the unified experimental Runner setting as the only enablement
flag. Onboarding and direct Codex, Claude, and OpenCode stay on their
legacy execution/finalization paths.
- Adds cross-language goldens, authority/recovery/fault coverage, exact
response/count assertions, and native plus legacy acceptance scenarios.
## Verification
- Pull-request GitHub Actions run Rust formatting/tests, TypeScript
checks, server/UI tests, builds, protocol drift checks, browser E2E, and
security scans.
- A separate workflow-only validation ref is pinned directly on this PR
head and runs the 35-cell paid local matrix: three core scenarios plus
structured-question resume and restart/resume for native Codex, native
OpenCode, ACPX Claude, ACPX Codex, and direct Codex/Claude/OpenCode.
Run: https://github.com/paperclipai/paperclip/actions/runs/33682434315
- Acceptance requires exact single visible replies, monotonic sequences,
matching envelope discriminators, one semantic terminal, one run
terminal, no unresolved interaction, no duplicate mutation, no secret
leakage, provider continuity, and zero native rows for direct adapters.
- Per maintainer direction, tests are running in GitHub Actions rather
than on the slower local host. Only formatters and static diff checks
were run locally.
## Risks
- Recovery from old or partial filesystem state is sensitive. The repair
fails closed, preserves active or unverifiable authority, and
quarantines only state whose scoped ownership is safe to move.
- Provider event formats can change. Closed validators and boundary
goldens turn new or malformed events into visible diagnostics instead of
silent drops.
- Shared task presentation could affect direct adapters. Runtime-fact
gating plus the direct-adapter matrix protect the existing path.
- Managed and remote providers are not qualified here. Shared code
continues to compile and fail safely, but live qualification is
deferred.
> 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. The exact deployed snapshot and
context-window size are not exposed to this task. It used agentic
reasoning, repository inspection, code editing, Git, parallel subagents,
and GitHub Actions.
## 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
- [ ] I have run tests locally and they pass (intentionally deferred to
GitHub Actions)
- [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
- [ ] The paid local-provider matrix is 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 control plane for AI-agent companies.
> - Agent outputs must remain visible after a run and easy to inspect
from a task.
> - The thread and artifact inventory need one consistent rich-card
vocabulary.
> - Run uploads also need durable artifact registration and
producing-run context.
> - Reviewers need deterministic examples for each rich-card kind and
state.
> - This pull request adds the shared presentation, registration,
inventory, and Storybook review coverage.
> - The benefit is a complete output path that reviewers can inspect
without seeded data.
## Linked Issues or Issue Description
**What existing behavior does this improve?**
This change improves work-product presentation in task threads and the
task Artifacts tab.
**Subsystem affected**
The change affects shared work-product contracts, the runner diff path,
server attachment and work-product services, GitHub metadata refresh,
the React board UI, and Storybook.
**Current behavior**
The thread used generic cards. Some files uploaded by a run existed only
as message attachments. The Artifacts tab showed a flat list without run
context or filters. Storybook showed only one resting card per kind.
**Proposed behavior**
The thread uses rich cards for supported work-product types. Each
run-produced file registers one attachment-backed artifact work product.
The Artifacts tab groups outputs by run and supports filters. Storybook
shows every kind and requested state, PR lifecycle states, stats
variants, truncation, mobile layout, and message-tail media.
**Reason and benefit**
Users can identify outputs quickly. Reviewers can inspect all card
permutations without creating task data.
**Breaking changes**
None. The metadata fields and automatic artifact registration are
additive. Existing attachments and work products keep their current
behavior.
## What Changed
- Added a shared rich work-product card with kind-specific content and a
compact inventory variant.
- Added pull-request and commit diff metadata plus bounded GitHub state
refresh.
- Added media strips and typed file chips to message-tail attachments.
- Registered each run-produced attachment as an artifact work product in
the same server transaction.
- Grouped task artifacts by run with agent and timestamp headings.
- Added type and run filters, image thumbnails, compact cards, and a
company Artifacts link.
- Added a Storybook kind-by-state matrix with stats variants for all
eight visual kinds.
- Added PR open, draft, merged, and closed examples, long-title
truncation, an exact 375-pixel viewport, and message-tail overflow
coverage.
- Closed reconciled runtime work products when the linked runtime stops
or disappears, so the card shows `Stopped` instead of `Unhealthy`.
### Screenshots
Before: one resting card per kind.
After: the kind and state matrix.
After: message-tail media at 375 pixels.
[Open the Storybook evidence
viewer](https://pages.paperclip.ing/rich-work-product-storybook-20260902/).
The earlier artifact inventory comparison remains available in the
[artifact inventory
viewer](https://pages.paperclip.ing/rich-artifacts-inventory-proof-20260902/).
## Verification
- `pnpm --filter @paperclipai/ui typecheck` passed.
- `pnpm check:token-gates` passed.
- `pnpm build-storybook` passed.
- `pnpm exec vitest run
server/src/__tests__/work-product-runtime-reconciliation.test.ts` passed
with 5 tests.
- Chromium visual checks passed at desktop and 375-pixel widths.
- All 30 latest-head GitHub checks passed. One unrelated annotation test
was flaky and passed on its single retry.
- Greptile passed at 5/5 with zero unresolved threads.
## Risks
- Low risk. The Storybook change adds review fixtures only. The runtime
fix changes read-time reconciliation without database writes.
- The matrix is intentionally large so every permutation stays visible
in one review surface.
> I checked `ROADMAP.md`. This work does not duplicate planned core
work.
## Model Used
- OpenAI Codex with GPT-5 and GPT-5.6-sol across this pull request.
Reasoning, tool use, and code execution were enabled. The context-window
size is 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 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 public branch name describes the change and contains no
internal task id
- [x] I have run tests locally and the changed-path tests 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
- [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 subsystem executes agent work across local and managed
provider backends.
> - The lower pull requests restore the task runtime, provider backends,
and managed-provider control plane.
> - The restored system needs repeatable full-stack checks before it can
ship safely.
> - Paid live checks also need clear access, cost, and secret controls.
> - This pull request adds acceptance, live evaluation, chaos, and
release gates for the restored runner stack.
> - The benefit is measurable runner parity with safer release
decisions.
## Linked Issues or Issue Description
**Subsystem affected**
Cross-cutting. This change covers runner tests, release workflows,
server contracts, and evaluation tools.
**Problem or motivation**
The runner stack did not have one complete acceptance surface for native
Codex, ACPX, Claude Managed, and AWS AgentCore. Release checks could
miss provider drift, task-view regressions, cost-policy errors, and
destructive cleanup errors.
**Proposed solution**
Add a 57-cell full-stack catalog, a Daytona image, and opt-in paid
workflows. Add live evaluation, chaos, cost-limit, redaction, and
release contract checks. Add AWS AgentCore infrastructure and guarded
provisioning tools. Keep the native runner experimental flag off by
default.
**Alternatives considered**
We considered manual smoke tests only. They do not give repeatable
evidence and they do not protect release branches. We also considered
one large pull request. The stacked pull requests keep each review below
the Greptile file limit.
**Roadmap alignment**
This work supports the shipped Cloud / Sandbox agents milestone and the
shipped Agent evals & feedback milestone in `ROADMAP.md`.
Related stack:
- #12699 adds managed provider backends and lifecycle support.
- #12691 adds qualified OpenCode and ACPX provider backends.
- #12685 restores task runtime rendering and steering.
## What Changed
- Add the runner full-stack harness with 57 catalog cells and 60 unit
tests.
- Add a Daytona runner image with digest-pinned base images and
base-aware image-content checks.
- Add guarded live evaluation and chaos workflows with a fixed
40-execution matrix; live and full-stack paid schedules now run only on
Sundays or by manual dispatch.
- Add in-flight reported-usage cost stops, post-turn cost caps,
exact-threshold failure classification, secret redaction, retry
classification, and actor authorization.
- Reattach stream and hard-budget listeners before restart-recovery
continuations so restored paid sessions cannot bypass in-flight
interruption.
- Preserve OpenCode usage and cost across tool-loop messages and turns
while exposing an explicit current-run delta to durable accounting.
- Keep PNG/WebM evidence in access-controlled artifacts only, reject
SVG, and publish only pruned inert structured per-attempt evidence.
- Add AWS AgentCore infrastructure, provisioning checks, and smoke
tools; reject unsafe model identifiers, require exact stack ownership
markers, and make failed-stack replacement explicit.
- Add evaluation-session contracts and capability reports.
- Add release workflow checks for immutable action pins, frozen
dependency installs, exact weekly cron shape, paid-run guards,
provider-secret isolation, and chaos test paths.
- Reauthorize the original and triggering numeric actor IDs as the first
step of every provider-secret job, including partial reruns, before
checkout or provider access.
- Give each full-stack matrix cell only its matching provider
credential, expose Daytona only to Daytona cells, and disable shared
dependency caches anywhere paid credentials or OIDC write access are
present.
- Protect the legacy manual E2E workflow with the same default-branch,
allowlist, environment, and per-job authorization boundary.
- Rotate live-eval candidates by week and retain 120 days of compatible
history so the seven-week trend window remains viable.
- Restore the root runner-acceptance commands and reconcile reported
snapshots,
raw receipts, and terminal usage without double counting or losing late
usage.
- Mark ACPX token deltas exact only when every budget field is present,
keep
cumulative cost/request authority separate, reject non-USD cost
labeling,
and include thought tokens in output-token budgets.
- Keep `enableNativeRunner` off by default. The acceptance harness
enables it only in its isolated test instance.
## Verification
Passed locally:
- `pnpm --filter @paperclipai/paperclip-runner typecheck`
- `pnpm test:runner-acceptance:typecheck`
- `pnpm test:runner-acceptance` (19 tests)
- focused OpenCode proxy, driver, runnerd transport, live-session, and
turn-stream tests (106 tests)
- `pnpm --filter @paperclipai/paperclip-runner exec vitest run
src/live/clean-room-server.test.ts` (22 tests)
- `pnpm test:e2e:runner:typecheck`
- `pnpm test:e2e:runner:unit` (62 tests)
- `node --test scripts/__tests__/release-verify-workflow.test.mjs`
- `pnpm --filter @paperclipai/paperclip-runner
test:runner-workflow-evals` (22 tests)
- `pnpm -r typecheck`
- `pnpm build`
- `node --test
packages/paperclip-runner/scripts/aws-agentcore-provisioning.test.mjs`
(6 tests)
- `git diff --check`
- `cargo test --manifest-path
packages/paperclip-runner/runner/Cargo.toml -p paperclip-runner-core
--lib --locked` (161 tests)
- focused ACPX provider-event tests (10 tests)
- The rebased PR changes 92 files. `pnpm-lock.yaml` is unchanged.
I did not run paid live provider jobs or provision AWS resources. Those
checks need credentials and can create cost.
## Risks
The paid workflows can create provider cost. They require an allowlisted
original and triggering actor, the protected `runner-e2e-paid`
environment, explicit opt-in variables, and cost limits. The four
provider credentials exist only in that master-only environment, which
requires allowlisted reviewer approval and disables administrator
bypass; repository and organization Actions scopes contain no copies.
Provider usage arrives after a billable request, so the live guard
cannot prevent one request from crossing a threshold. It interrupts
immediately on the first reported threshold hit and permits no
continuation.
Visual evidence can contain secrets rendered as pixels. PNG/WebM remain
only in access-controlled workflow artifacts; SVG and per-attempt XML
are excluded, and S3/Pages receive a pruned structured dashboard.
The AWS scripts can create cloud resources. They use explicit commands,
least-privilege roles, KMS encryption, saved nonsecret metadata, and
explicit teardown.
This pull request does not enable the experimental native runner for
existing instances.
> 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 with GPT-5. The model used extended reasoning, tool use,
code execution, and parallel subagents.
## 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
## Thinking Path
> - Paperclip is the open source app people use to manage AI agents for
work.
> - The Paperclip Runner provides durable, provider-neutral agent
execution.
> - The current stack supports qualified local providers but omits the
managed provider paths from the integration branch.
> - Claude Managed Agents and AWS AgentCore need explicit profile
qualification, durable recovery, usage accounting, and cleanup controls.
> - This pull request adds those managed backends as the third part of
the Runner parity stack.
> - The benefit is managed execution without weakening the default-off
Runner rollout gate.
## Linked Issues or Issue Description
**Subsystem affected**
Cross-cutting: Runner, server orchestration, database profiles, CLI, and
adapter configuration UI.
**Problem or motivation**
The current Runner stack cannot select or execute the managed Claude
Agents API or AWS Bedrock AgentCore Harness backends. It also lacks
qualified profile storage and recovery checks for those remote
resources.
**Proposed solution**
Add qualified managed and remote profiles, API and CLI management, exact
provider selection, durable lifecycle handling, cumulative usage
accounting, bounded cleanup, and retention acknowledgement. Keep
`enableNativeRunner` default-off.
**Alternatives considered**
A direct copy of the old integration branch was rejected because its
provider contracts, model values, credential flow, and migration history
no longer match the current base. A single large parity pull request was
also rejected because stacked review keeps each subsystem bounded.
**Roadmap alignment**
This continues the existing Runner architecture and rollout work. It
does not introduce a separate execution system.
**Additional context**
This pull request is based on the merged #12691 and #12685 stack. It
also closes the delayed security-review findings reported on #12691 by
binding qualified ACPX and OpenCode launch artifacts to the bytes
actually executed. A GitHub search for managed agent, AgentCore, and
Claude managed work found no duplicate public issue or pull request.
## What Changed
- Add Claude Managed Agents and AWS AgentCore provider executors to
runnerd.
- Add qualified managed and remote profile storage, routes, OpenAPI
contracts, CLI commands, and migration 0237.
- Validate profile ownership, enabled state, exact qualified revision,
model, agent version, and secret binding before persistence and
recovery.
- Persist durable provider session and owned skill state for
restart-safe cleanup.
- Reconcile uncertain create responses and delete remote sessions before
owned skills.
- Track cumulative provider usage and enforce positive session spend
caps.
- Recover interrupted AgentCore usage at the next turn boundary by
charging the prior invocation ceiling exactly once; keep the session
gated until an explicit monotonic budget raise.
- Isolate AgentCore AWS configuration from host profiles and
credential-process/SSO configuration while preserving workload identity.
- Require OpenCode 1.18.17 and fixed build-owned provider-pack artifact
paths; remove the ambient executable override.
- Snapshot and content-verify ACPX and OpenCode commands, scripts, and
provider executables before launch. Linux executes sealed inherited
descriptors; macOS uses authenticated private snapshots with retry-safe
rematerialization at the spawn boundary.
- Persist canonical ACPX and OpenCode launch-profile digests, reject
drift across fresh recovery, and make recovery failures sticky.
- Close and journal unsafe ACPX active-turn recovery before any provider
bootstrap or reconnect.
- Add managed provider fields to the Runner configuration UI and
permission projection.
- Preserve the default-off `enableNativeRunner` experimental flag.
## Verification
- `pnpm -r typecheck`
- `pnpm build`
- Focused managed server, database, CLI, Runner TypeScript, Rust,
Claude, AgentCore, ACPX, OpenCode, process-supervisor, and
durable-recovery tests passed.
- `cargo test -p paperclip-runner-core --lib --locked` (160 tests)
- `cargo check --workspace --all-targets --locked`
- Native Codex integration tests passed (60 tests); native provider
tests passed (7 tests); server native-runtime tests passed (87 tests).
- Verified-launch replacement, nested-spawn retry, exact-version,
profile-drift, sticky-failure, and no-bootstrap active-recovery tests
passed.
- `git diff --check`
- The PR changes 91 files. `pnpm-lock.yaml` is unchanged. The Rust
workspace lockfile adds the approved `rustix` dependency used for safe
descriptor handling while `#![forbid(unsafe_code)]` remains enabled.
## Risks
- The provider APIs can change while they are in beta. Exact
qualification and fail-closed recovery checks limit drift.
- Remote cleanup can fail after a partial create. Durable ownership
inventories and retry-safe deletion preserve recovery state.
- Migration 0237 adds profile tables. The generated migration and
snapshot pass the repository migration checks.
- Managed execution can incur provider cost. Positive default spend caps
and explicit retention acknowledgement limit accidental use.
- An interrupted AgentCore invocation without final metadata is
conservatively charged to its active session ceiling. This can overstate
cost, but cannot undercount it; later work requires an explicit budget
increase.
- Linux qualified launches use sealed memory descriptors. macOS lacks
executable-descriptor APIs, so the runner uses owner-only private
snapshots and minimizes linked-path lifetime; hostile same-UID processes
remain outside the documented local-host trust boundary.
- The global Runner feature remains default-off.
> 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, with tool use, code execution, and subagent 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 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] 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
## Thinking Path
> - Paperclip is the open source app people use to manage AI agents for
work.
> - Paperclip Runner is the experimental native runtime for governed
agent work.
> - The runtime contracts already describe Codex, OpenCode, and ACPX
providers.
> - The merged control plane still rejected OpenCode and ACPX for new
runner agents.
> - Runnerd also selected only the Codex provider implementation.
> - This pull request activates the qualified OpenCode and ACPX paths
from the form to runnerd.
> - The benefit is one durable runner path with provider-specific
permissions and recovery.
## Linked Issues or Issue Description
Refs #12685
**Subsystem affected**
This change affects the runner package, server orchestration, adapter
configuration, and UI configuration.
**Problem or motivation**
Paperclip Runner stores provider contracts for OpenCode and ACPX. New
agents cannot select those providers. Runnerd cannot execute those
stored provider descriptors. The UI also shows only Codex.
**Proposed solution**
Accept the qualified OpenCode 1.18.17 profile and the fixed ACPX Claude
and Codex profiles. Route them through runnerd. Keep provider selection,
model selection, permissions, credentials, events, and recovery inside
closed provider-specific boundaries.
**Alternatives considered**
One option was to keep the contracts dormant. That option leaves stored
configuration and runtime behavior out of sync. Another option was to
enable every ACPX agent. That option is not safe because Pi does not yet
have the same verified launch path.
**Roadmap alignment**
This change supports the completed cloud and sandbox agent milestone. It
also supports self-healing runs and governed agent execution. It does
not add a new roadmap surface.
## What Changed
- Add one server profile resolver for Codex, OpenCode, and qualified
ACPX descriptors.
- Keep `adapterConfig` as the provider and permission authority for
fresh runs.
- Add Paperclip Runner provider, ACPX agent, and provider-specific
permission controls to the UI.
- Reset the model to a compatible qualified value when the provider
changes.
- Route Codex, OpenCode, and ACPX through the durable runnerd provider
selector.
- Add a durable ACPX executor with bounded state, recovery, events, tool
receipts, and identity checks.
- Remove Codex labels from OpenCode events, results, evidence, and
recovery diagnostics.
- Pass only provider-specific credential names to child processes.
- Keep ACPX Pi unavailable and reject it before process launch.
- Keep the existing Paperclip Runner experimental flag unchanged.
## Verification
- `pnpm exec vitest run
packages/paperclip-runner/src/backends/native-backend-factory.test.ts
packages/paperclip-runner/src/live/runnerd-codex-transport.test.ts
packages/adapters/codex-local/src/ui/build-config.test.ts
ui/src/adapters/codex-local/config-fields.test.tsx
server/src/__tests__/adapter-registry.test.ts
server/src/__tests__/adapter-routes.test.ts
server/src/__tests__/agent-adapter-validation-routes.test.ts
server/src/__tests__/company-portability.test.ts
server/src/services/native-runtime/runtime-mode.test.ts
server/src/services/native-runtime/native-session-executor.test.ts
server/src/services/heartbeat-runner-provider-config.test.ts`
- The focused TypeScript, server, and UI suites passed 274 tests.
- `cargo test -p paperclip-runner-core --test native_provider_backend`
- The executable native provider integration suite passed 4 tests.
- `cargo test -p paperclip-runner-core --lib`
- The Rust unit suite passed 91 tests.
- `pnpm -r typecheck`
- `pnpm check:token-gates`
- `pnpm build`
- `git diff --check codex/runner-parity-task-runtime...HEAD`
## Risks
- This changes provider process selection and durable recovery. The
experimental flag still gates every fresh Paperclip Runner run.
- OpenCode requires a model in `provider/model` form and stays pinned to
version 1.18.17.
- ACPX accepts only exact Claude and Codex profile versions and models.
Pi stays unavailable.
- ACPX steering stays unavailable and reports that limit through the
driver capabilities.
- Child processes receive explicit environment allowlists. They do not
inherit the full server environment.
- This pull request has no database migration.
## Model Used
OpenAI Codex, GPT-5, with tool use, code execution, and subagent 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 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 Paperclip Runner package manages live sessions and durable
command recovery
> - Capability live-session tests can fail when a turn-timeout rejection
has no handler
> - A resumed session can also stall when the durable control plane
rejects an indeterminate command result
> - These failures make valid tests fail or hide the turn that stalled
> - This pull request captures timeout rejections early and accepts
indeterminate recovered commands
> - The benefit is stable tests and clearer timeout failures after a
runner restart
## Linked Issues or Issue Description
**What happened?**
Capability live-session tests failed intermittently on loaded CI hosts.
A timer could reject a turn promise before the test attached its
assertion. A resumed session could also stall after a runner restart
because the durable control plane rejected the indeterminate command
status.
**Expected behavior**
The test must handle a timeout rejection at promise creation. The
durable control plane must accept an indeterminate recovered command and
allow the session to continue. A configured timeout must persist in the
checkpoint and identify the stalled turn.
**Steps to reproduce**
1. Run the capability live-session test file on a loaded host.
2. Create a turn promise with a timeout and delay before attaching its
assertion.
3. Resume a session after a runner restart with a journaled but
unconfirmed command.
4. Observe the unhandled rejection or the stalled resumed session.
**Paperclip version or commit**
Commit `ede642e57e22ea3fb0a73590fca8bcc994f1a47f` on `master`.
**Deployment mode**
Built from source.
**Installation method**
Built from source.
**Agent adapter(s) involved**
Not adapter-specific. The tests use the Paperclip Runner package.
**Database mode**
Not database-related.
**Additional context**
Pull request #12646 also updates durable recovery for indeterminate
command results. If it lands first, this pull request must retain the
compatible behavior without duplicate edits.
## What Changed
- Add a helper that captures a turn rejection before any await step.
- Update three live-session test sites to assert the captured rejection
value.
- Add a helper test that waits past the turn timeout before it asserts.
- Accept indeterminate as a terminal recovered-command status.
- Add tests for acceptance, duplicate absorption, and reload from
persisted state.
- Add an optional turnTimeoutMs value to resume and pin its checkpoint
behavior.
## Verification
- `npx vitest run src/live/live-session.test.ts` from
`packages/paperclip-runner`: 19 passed, 1 skipped.
- `npx vitest run src/control-plane/durable-prp-control-plane.test.ts`
from `packages/paperclip-runner`: 5 passed.
- The live-session file passed 10 of 10 runs with 30 competing workers
on a 32-core host.
- TypeScript reported five pre-existing errors in
`src/eval/workflow-harness.ts`.
- CI must run `pnpm --filter @paperclipai/paperclip-runner check:all`.
## Risks
The durable control plane now accepts one additional terminal recovery
status. The change affects only recovered command handling and
capability live-session tests. The main risk is overlap with pull
request #12646 if that pull request lands first.
## Model Used
OpenAI Codex, GPT-5, 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 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 subsystem pairs a Rust runner process with a durable
control plane in TypeScript. The control plane is the authority for
every command the runner executes
> - That pair has a crash-recovery contract. If the runner dies between
journaling a command and confirming the command's effect, it must not
run the command twice
> - The runner keeps its side of the contract. On restart it promotes
such a command to the `indeterminate` status and reports that status
back
> - The control plane did not accept `indeterminate`. It closed the
connection without a diagnostic, the runner reconnected and replayed the
same result, and the loop repeated forever
> - This pull request accepts `indeterminate` as a terminal command
status
> - The benefit is that a session survives a runner crash during a tool
call, instead of hanging until a 30 second deadline expires
## Linked Issues or Issue Description
No public issue exists for this defect, so it is described here.
**What happened?**
A live session cannot resume after the runner process is killed during a
governed tool call. The resumed transport waits for the provider
identity for
30 seconds and then fails with `runnerd did not report its provider
identity`.
`packages/paperclip-runner/src/live/live-session.test.ts` covers this
exact
sequence in "terminates real runnerd after a durable receipt and resumes
its
exact provider thread". That test has a 15 second budget, so it reports
the
defect as `Test timed out in 15000ms` and reads like a flake.
**Expected behavior**
The resumed control plane accepts the runner's recovery report, the
runner
reports its provider identity, and the session resumes on its original
provider thread.
**Steps to reproduce**
Build the runner binary, then run the test:
```
cargo build --manifest-path packages/paperclip-runner/runner/Cargo.toml --locked --workspace --bins
cd packages/paperclip-runner
npx vitest run src/live/live-session.test.ts -t "terminates real runnerd"
```
It fails every time on an idle machine. It also fails at `560e7e48b`,
the
commit that added the test, so the defect is not a recent regression.
**Paperclip version or commit**
Reproduced on `master` at `0a422fda5`, which is the base of this branch.
**Deployment mode**
Local development, running the package test suite.
**Root cause**
`DurablePrpControlPlane.#commandResult` accepted only `completed`,
`failed`
and `rejected`. The runner reports a journaled-but-unconfirmed command
as:
```json
{ "status": "indeterminate",
"result": { "code": "execution_indeterminate",
"message": "runner recovered after journaling this command; it will not execute twice" } }
```
That status fell through to a silent `connection.close()`. The runner
reconnected after 250 ms, replayed the same result, and was closed
again. No
durable event ever reached the control plane, so the transport never saw
`harness.ready`.
`indeterminate` is a deliberate part of the runner's contract. See
`reconcile_pending_commands` in
`packages/paperclip-runner/runner/crates/runner-core/src/durable/state.rs`.
The rest of the TypeScript code already models the status; only this
control
plane did not.
## What Changed
- `DurablePrpControlPlane.#commandResult` accepts `indeterminate` as a
terminal command status.
- The persisted-state validation accepts `indeterminate`, so a control
plane
restarted over the same directory can read its own saved state back.
Without this, accepting the status would make the next restart throw.
- `DurableRecoveryCoreCommand.status` includes `indeterminate` in both
declarations of that interface.
- Added an integration test that drives the exact recovery frame the
runner
sends. It asserts the connection stays open, the next command is
delivered,
the status is persisted, a restarted control plane reloads it, and a
replayed duplicate is absorbed rather than treated as a conflict.
## Verification
All commands run from `packages/paperclip-runner`.
- New test fails before the change and passes after it. Before:
`expected null to match object { kind: 'command' }` — `null` is the
closed
connection.
`npx vitest run src/control-plane/durable-prp-control-plane.test.ts`
→ 4 passed.
- The live runner test that exposed this reproduced
**deterministically** on an
idle machine before the change, and now passes in 3.3 s, well inside its
existing 15 s budget. Ran it 10 times in a row: 10/10 pass, 0 failures.
`npx vitest run src/live/live-session.test.ts -t "terminates real
runnerd"`
- Full package suite: `npx vitest run` → 1298 passed, 1 failed. The one
failure is `src/mock-core/local-runner.test.ts > cleans up the harness
process group when the controller closes`. It fails identically on an
unmodified checkout in the same container, so it is a pre-existing
environment issue and not related to this change.
- Typecheck: `tsc -p tsconfig.json --noEmit` → clean.
I did **not** raise the test's timeout. The budget was never the problem
—
with a 600 s budget the same test still failed, at 31 s, with the real
error.
## Risks
Low risk, and it widens rather than narrows what is accepted.
- Behaviour only changes for a status that is currently rejected, so no
previously working path is affected.
- `indeterminate` is terminal, not successful. A caller waiting on such
a
command still receives an error from the transport, which is correct:
the
effect is genuinely unconfirmed. This change does not make an
unconfirmed
command look like it succeeded.
- The persisted-state change only widens an allow-list, so existing
state
files stay valid.
Open topics for a reviewer:
- The control plane closes connections without any diagnostic. That
silence is
why this defect looked like a flaky test. Adding a diagnostic channel is
a
larger change and is not included here.
- `DurableRecoveryProcessedCommand` in
`src/contracts/durable-recovery.ts`
drifts from the Rust `StoredCommandResult` by more than this status: it
declares `commandDigest` and `logicalEffectCount`, which Rust does not
have,
and omits `commandType`, which Rust does. That is a separate correction
and
is deliberately not folded in here.
## Model Used
Claude Opus 5 (`claude-opus-5`), extended thinking, with tool use and
code
execution.
Depends-on: none — this is a self-contained fix with no dependent
changes.
## 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: zannis <1011451+zannis@users.noreply.github.com>
Co-authored-by: Claude Opus 5 <noreply@anthropic.com>
## Thinking Path
> - Paperclip is the open source app people use to manage AI agents for
work.
> - Paperclip Runner needs repeatable evaluation contracts.
> - Evaluation code must stay separate from provider launch and
production orchestration.
> - Offline fixtures need stable compatibility, scoring, traceability,
and report rules.
> - Published Runner consumers need only the supported evaluation
contract surface.
> - This pull request adds offline evaluation tooling and a
workspace-private matrix kernel.
> - The benefit is deterministic evaluation without credentials or paid
provider calls.
## Linked Issues or Issue Description
Refs #11297
This pull request extracts the offline evaluation unit from the earlier
aggregate Runner work.
## What Changed
- Add a workspace-private, provider-neutral evaluation matrix kernel.
- Add the public `@paperclipai/paperclip-runner/evals` compatibility and
native execution contracts.
- Add fail-closed runnerd artifact and protocol compatibility checks.
- Add deterministic workflow catalogs, scoring, traceability, and report
generation.
- Add sanitized Codex, OpenCode, and ACPX fixtures.
- Add package-boundary and clean-consumer checks.
- Add the eval package manifest to the Docker dependency stage.
- Add the generated protocol fixture digest without changing the
lockfile.
## Verification
GitHub Actions must run:
- Runner TypeScript and Rust type checks.
- Runner unit and protocol tests.
- Evaluation kernel tests.
- Workflow traceability checks.
- Clean-consumer and package-boundary checks.
- Repository test, type-check, build, policy, and security gates.
No local test command was run. The repository owner requested
GitHub-only verification.
## Risks
This is a large greenfield review surface with 51 files. The code does
not launch a live provider or load credentials. Package and protocol
drift fail closed. The workspace lockfile remains under the existing
CI-owned process.
## Model Used
OpenAI Codex with the GPT-5 agent model. The work used high reasoning,
repository inspection, tool use, and parallel code 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 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
- [ ] 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.
> - Paperclip Runner currently enables only the Codex production path.
> - The package also contains dormant OpenCode and ACPX provider
boundaries.
> - Dormant boundaries must still fail safe before later activation
work.
> - Provider children must not inherit unrelated server secrets or host
homes.
> - Permission defaults must require interaction instead of broad
automatic approval.
> - This pull request hardens those boundaries without activating them.
> - The benefit is a safer base for later provider-specific runnerd
work.
## Linked Issues or Issue Description
**What existing behavior does this improve?**
This improves the inactive OpenCode and ACPX provider boundary in
Paperclip Runner.
**Subsystem affected**
The adapter permission contract, Runner provider environment, and native
execution input builder.
**Current behavior**
Dormant OpenCode code can inherit the full server environment. Its
default permission mode allows operations. ACPX also defaults to broad
approval. The provider guard can accept inherited object property names.
**Proposed behavior**
Use exact provider identifiers. Use interactive defaults. Allow only
required OpenCode environment keys. Reject invalid proxy permission
modes.
**Reason and benefit**
This reduces accidental authority and secret exposure before future
provider activation.
**Breaking changes**
No production provider is activated. Codex runtime selection and Codex
credential-home discovery do not change. Dormant OpenCode and ACPX
callers that omit permission modes now receive safer defaults.
## What Changed
- Change dormant OpenCode and ACPX permission defaults to interactive
modes.
- Reject prototype property names as provider identifiers.
- Default dormant ACPX input to the qualified Codex agent profile.
- Add an explicit OpenCode runner environment allowlist.
- Exclude host homes, server credentials, database values, and Node
injection options.
- Add a fail-closed OpenCode proxy permission parser.
- Add focused tests for defaults, filtering, and invalid values.
## Verification
GitHub Actions must run:
- Adapter utility tests.
- Paperclip Runner tests, type checks, and build.
- Server native runtime tests.
- Repository test, type-check, build, policy, and security gates.
No local test command was run. The repository owner requested
GitHub-only verification.
## Risks
Future OpenCode credential providers must add required variables to the
allowlist through review. The safer defaults can pause dormant internal
scenarios that relied on implicit broad approval. Production Codex
behavior is unchanged.
## Model Used
OpenAI Codex with the GPT-5 agent model. The work used high reasoning,
repository inspection, tool use, and parallel security 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 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
- [ ] 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.
> - Paperclip supports direct adapters and an experimental native Runner
path.
> - These paths need one stable compatibility matrix.
> - The matrix must not launch providers or load credentials during
normal tests.
> - Result handling must reject incomplete output and sensitive values.
> - This pull request adds a credential-free acceptance foundation.
> - The benefit is a reviewable contract for later end-to-end executors.
## Linked Issues or Issue Description
**What existing behavior does this improve?**
This improves verification for direct adapters and Paperclip Runner
providers.
**Subsystem affected**
Cross-cutting test infrastructure for adapters, the server runtime, and
the task thread.
**Current behavior**
The repository has subsystem tests. It does not have one declarative
matrix for direct and native compatibility.
**Proposed behavior**
Add a pure acceptance catalog, result validator, redaction helpers, and
failure classification. Keep all execution authority outside this
change.
**Reason and benefit**
The matrix makes legacy isolation and native recovery requirements
explicit. The helpers let later executors report safe and complete
results.
**Breaking changes**
None. This change does not alter production runtime selection or start
any provider.
## What Changed
- Add a catalog for built-in direct adapters and qualified native
provider profiles.
- Add compatibility cases for runtime selection, task threads,
questions, and flag-change recovery.
- Add pure redaction and transient-failure classification helpers.
- Add fail-closed Markdown and JUnit report aggregation.
- Add isolated test and type-check commands.
- Document the credential-free boundary and deferred live execution
work.
## Verification
GitHub Actions must run:
- `pnpm test:runner-acceptance`
- `pnpm test:runner-acceptance:typecheck`
- The repository test, type-check, build, policy, and security gates.
No local test command was run. The repository owner requested
GitHub-only verification.
## Risks
Low production risk. The change adds test-only files and root scripts.
The catalog can drift when a built-in adapter changes. Its validation
fails closed on that drift.
## Model Used
OpenAI Codex with the GPT-5 agent model. The work used high reasoning,
repository inspection, tool use, and parallel code 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 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
- [ ] 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.
> - Administrators need bounded controls for experimental native
execution.
> - The lower stack adds remote Codex execution and the task workspace.
> - Operators need to configure Codex safely and inspect provider
traces.
> - Unsupported providers must not appear as runnable choices.
> - This pull request adds Codex-only administration and observability.
> - The benefit is a default-off operational surface for production
diagnosis.
## Linked Issues or Issue Description
Refs #12640.
Refs #12616.
Refs #12352.
**Subsystem affected**
Agent configuration, instance experimental settings, run ledger,
provider trace inspector, and administrator actions.
**Problem or motivation**
The native runner lacks one safe operator surface for Codex permissions,
lifecycle, raw trace capture, and run inspection. The integration branch
also contains provider choices that the production backend cannot
execute yet.
**Proposed solution**
Expose only the qualified Codex controls. Keep Paperclip Developer Mode
and runner preview ingress off by default. Gate raw trace actions by
administrator access and existing trace authorization.
**Alternatives considered**
Exposing unfinished providers would create configurations that fail at
runtime. Always-on tracing would increase sensitive data and storage
risk.
**Roadmap alignment**
This work supports governed Cloud and Sandbox agents and production
diagnostics.
## Stack
- Base PR: #12640.
- Lower PRs: #12639 and #12638.
- This PR contains only its 54-file administration and observability
delta.
- This is the final feature PR in the Codex production stack.
## What Changed
- Added Codex-only Paperclip Runner permission and lifecycle controls.
- Added bounded warm idle configuration.
- Kept the provider field fixed to Codex.
- Added administrator-only one-run raw trace requests.
- Added a persistent future-run raw trace toggle.
- Added trace status, metadata, ledger, and canonical runner inspection.
- Added JSON-RPC request-origin grouping and finalization lineage.
- Restored the stateful PRP transcript parser and focused projection
tests required by trace inspection.
- Added default-off Paperclip Developer Mode.
- Added Honeycomb run links for authorized developer mode.
- Disabled the legacy operational skill for `paperclip_runner`.
- Did not expose OpenCode, ACPX, Pi, Claude Managed, or AWS runner
choices.
- Did not change migrations, workflows, dependencies, or
`pnpm-lock.yaml`.
## Verification
- GitHub Actions will run UI tests, server tests, repository typecheck,
build, browser tests, security, and policy gates.
- Tests cover Codex configuration defaults and bounds, administrator
trace actions, persistent settings, ledger inspection, trace lineage,
and Honeycomb links.
- Existing server trace authorization and retention tests remain the
backend authority.
- Local tests were not run. The requested verification policy uses
GitHub Actions for this series.
- `git diff --check runner/task-workspace-experience...HEAD` passes.
- The delta contains 54 files.
## Risks
- Raw provider traces can contain sensitive provider data.
- Existing server authorization controls access, reveal, download,
retention, and deletion.
- The UI gates trace actions by administrator access and developer mode.
- All new instance settings remain off by default.
- Fresh Paperclip Runner configuration remains Codex-only.
- Direct adapters and legacy task behavior do not change in this PR.
> 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 with GPT-5.6. The work used high-reasoning agent mode,
repository tools, GitHub tools, and parallel code-audit agents.
## 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 or 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
Dotta
Devin Foley
Nicky Leach
Zannis Kalampoukis
