mirror of
https://github.com/paperclipai/paperclip.git
synced 2026-10-07 07:23:08 +02:00
3124dd0f1e
## Thinking Path
> - Paperclip is the open source app people use to manage AI agents for
work
> - When a run is stranded (process lost, adapter failure, a finished
run with no disposition, an over-eager inactivity kill), the harness
opens a *recovery action* and wakes an owner to recover it
> - Recovery volume regressed sharply in one week — 3.26% of all runs vs
a ~1.2% monthly norm, 5–8x the prior volume — and nobody noticed until
it was ~194 actions deep, because there was no way to *see* the recovery
rate
> - We also could not see which causes drive recovery, nor how often a
manager ends up doing the deliverable work themselves instead of handing
it back to the original owner (the product goal is that managers doing
the work stays rare)
> - This pull request adds a recovery-observability report + API
endpoint: weekly rate normalized per run, a threshold alert, the cause
taxonomy live from the ledger, and the handed-back vs owner-completed
ratio and per-cause routing outcomes
> - The benefit is that a recovery regression like that week is caught
by a threshold instead of by a human noticing it by feel, and each
recovery playbook row can be verified in production
## Linked Issues or Issue Description
**Feature.**
**Problem or motivation**
Recovery takeovers are a first-class exception path
(`issue_recovery_actions`), but there is no aggregate view of them. A
week where the recovery rate tripled went unnoticed until it was deep.
There is no signal for (a) the per-run recovery rate over time, (b)
which cause + run error code drives it, or (c) whether the recovery
owner hands the task back to the original assignee or ends up doing the
deliverable work themselves.
**Proposed solution**
A read-only report service and `GET
/companies/:companyId/recovery-observability` endpoint that surfaces the
weekly rate, a threshold alert, the cause taxonomy, the hand-back ratio,
and per-cause routing outcomes.
**Alternatives considered**
Adding `handed_back` / `owner_completed` to the recovery-action outcome
vocabulary and writing them at resolution time. Rejected for this
change: the distinction is derivable from the recovery owner, the
recorded return owner, and where the source issue actually landed, so
the report works against all historical data without a backfill.
**Roadmap alignment**
Implements the recovery-observability line of the approved
recovery-takeover plan (make regressions visible via a threshold rather
than by human feel); no schema or write-path change.
## What Changed
- Add `server/src/services/recovery-observability.ts`:
- `recoveryObservabilityService(db).report(companyId, { weeks,
thresholdPercent, now })` returns weekly rates (recovery actions / runs,
Monday-anchored to match the retrospective), a `cause` +
`latestRunErrorCode` breakdown, a handed-back vs owner-completed
summary, and per-cause routing outcomes.
- `evaluateRecoveryRateAlert(weekly, thresholdPercent)` — a pure
function (default threshold 2% of runs) returning the breached weeks and
whether the latest week regressed.
- `classifyRecoveryHandoff(...)` — a pure classifier deriving
`self_recovery` / `handed_back` / `owner_completed` from the recovery
owner, return owner, and final issue landing.
- Add `GET /companies/:companyId/recovery-observability` (optional
`weeks` and `threshold` query params) to the existing dashboard router.
- The `weeks` window is bounded (`MAX_WINDOW_WEEKS = 104`,
service-authoritative and re-clamped at the route) so a large query
value can't over-allocate the per-week array.
- Add tests: unit coverage for the alert and the classifier, plus an
embedded-Postgres integration test that seeds synthetic runs and
recovery actions crossing 2% and asserts the alert fires and the
hand-back ratio is computed.
## Verification
- `CI=1 NODE_ENV=development npx vitest run
server/src/__tests__/recovery-observability.test.ts` — 10/10 pass
(includes the synthetic 2%-crossing alert case and the hand-back ratio
case).
- Rendered against a live database of 300+ recovery actions: the weekly
rates reproduce the retrospective (e.g. 1.37% / 1.53% / 0.65% / 0.86% /
1.37% for early-June weeks), the alert fires on the two most recent
weeks (3.15% and 3.05%, both over 2%), and the hand-back summary shows
owner-completed ≈ 73% vs handed-back ≈ 27% — matching the observed
"managers keep ~80% of takeovers".
## Risks
- Low risk. Read-only: adds one GET endpoint and a service; no schema,
migration, or write-path changes. The hand-back classification reads the
source issue's current assignee/status, so a much-later reassignment
could reclassify a historical action — acceptable for an aggregate trend
view.
## Model Used
- Claude, `claude-opus-4-8` (Opus 4.8), extended thinking, tool use /
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
- [x] My branch name describes the change and contains no internal
Paperclip ticket id
- [x] 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 P2s, recommendations, or follow-ups
- [ ] I will address all Greptile and reviewer comments before
requesting merge
🤖 Generated with [Claude Code](https://claude.com/claude-code)
---------
Co-authored-by: Paperclip <noreply@paperclip.ing>
Co-authored-by: Claude Opus 4.8 <noreply@anthropic.com>