## Thinking Path
> - Paperclip is the open source control plane people use to coordinate
AI agents and their work.
> - The heartbeat recovery subsystem detects successful runs that leave
assigned issues `in_progress` without a durable disposition or
continuation path.
> - The existing corrective wake used a cheap, status-only model
profile, so the assignee could not perform missing verification or
deliverable work before choosing the issue disposition.
> - The existing wake prompt also omitted the original issue context and
the agent's own final report, making an honest finish/blocked/continue
decision harder.
> - This pull request keeps the structural handoff guards and
one-attempt loop bound, but wakes the assignee on its normal model lane
with context-rich instructions.
> - The benefit is that Paperclip asks the responsible agent to inspect
its own evidence, perform the smallest missing verification when needed,
and then record a real disposition without server-side prose
classification.
## Linked Issues or Issue Description
Related prior approach: #10154 (closed; this PR intentionally does not
reuse its regex classifier or route-level gate).
**Problem**
A succeeded agent run can leave its issue `in_progress` with no valid
disposition. Paperclip already detects this structurally and queues a
corrective handoff, but that wake currently runs as cheap/status-only
recovery and receives little context. The assignee may be unable to
create deliverables or verify the work, and the prompt does not quote
the report that caused the ambiguity.
**Expected behavior**
The corrective wake should use the assignee's normal model and adapter
settings, include the issue identifier/title/description, quote the
agent's own final report, include any recorded next action, preserve the
four disposition options, and explicitly require concrete verification
before marking the issue done.
**Scope**
This change does not classify run prose, add a route-level disposition
gate, alter run-liveness classification, or change the one-attempt
handoff loop bound.
## What Changed
- Switched successful-run corrective handoff payloads and context
snapshots from `status_only` to `normal_model`, removing cheap-model and
status-only guard hints.
- Added issue description, final-report, next-action, and
detected-progress fallback context to the handoff decision and
instruction builder.
- Reworked the instruction into clear "supposed to do / what happened /
options / what to do" sections with bounded description/report excerpts
and verbatim blockquotes.
- Added unit and heartbeat integration coverage for normal-lane
payloads, context plumbing, evidence quoting, fallback behavior, and
truncation while preserving structural skip tests.
## Verification
- `cd server && pnpm exec vitest run
src/services/recovery/successful-run-handoff.test.ts` — 24 tests passed.
- `cd server && pnpm exec vitest run
src/__tests__/heartbeat-process-recovery.test.ts -t "queues one
finish-handoff wake when a successful run leaves in-progress work
without a next action"` — 1 passed, 90 skipped.
- `pnpm --dir server typecheck` — passed.
- `git diff --check` — passed.
## Risks
- Low-to-moderate behavioral risk: an ambiguous successful run now
consumes the assignee's normal model rather than a cheap profile and may
perform verification or finish work before disposition.
- Prompt excerpts are bounded to approximately 1,200 description
characters and 2,000 report characters; very long context is
intentionally ellipsized.
- The existing structural skip guards, idempotency key, and single
corrective attempt remain unchanged to prevent loops.
> 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 using `gpt-5.6-sol`, high reasoning effort, with
repository/tool execution. Context-window size was not exposed by the
runtime configuration.
## 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>
Co-authored-by: Claude Fable 5 <noreply@anthropic.com>
7014e46e5b