mirror of
https://github.com/paperclipai/paperclip.git
synced 2026-10-06 21:05:21 +02:00
DottaandPaperclip
78e0034498
## Thinking Path > - Paperclip is the open source app people use to manage AI agents for work. > - Product E2E evals check real hiring and delegated task completion. > - The hiring fixture requires three requested CEO turns and two coder executions. > - The server can also wake the CEO when each delegated task completes. > - Two exact-five-run guards rejected these valid completion turns in all four retained cells. > - This pull request validates bounded completion turns in both guards. > - The benefit is accurate workflow grading while all actual runs and coverage failures remain visible. ## Linked Issues or Issue Description Refs: #14985, #14948, #14961. **What happened?** The original hiring comparison reports Codex Fail → Fail and Claude Fail → Fail. Each cell has seven successful runs. The five requested work turns are accompanied by two server task-completion notifications. All six other delivery checks pass. **Expected behavior** Require exactly three distinct user-requested CEO turns and one coder execution for each of two known tasks. Admit at most two strictly attributed server completion turns, including one turn that batches both tasks. Reject unknown, duplicate, failed, retried or extra-work runs. **Steps to reproduce** Inspect the retained four-cell report linked below. Each original result fails `five-successful-turns`. The same exact count was also enforced by the final chat-flow guard. ## What Changed - Add one typed lifecycle helper shared by the hiring scorer and the hiring-only final chat guard. - Validate public run ledgers, company/user/account identity, request attribution, task origins, completion deliveries, timing and replies. - Keep exactly five required work turns; declare seven maximum total turns for cost and timeout planning. - Count all actual runs, including notification runs and unexpected resets. Keep other chat count guards unchanged. - Version the hiring grader as v3 (turn accounting v2) and include the helper and chat guard in its definition digest. - Keep source-read, exact coder-body and all six other delivery checks unchanged. - Add 144 focused helper/scorer/settlement calibrations and separately versioned exact retained-input replay reports. - Retry complete bracketed observations, await both owed callbacks and attributed replies, and refresh the final guard consistently. - Reject unrelated completion writes and failed mutation attempts using exact canonical/native action IDs. Missing identity mapping is uncomparable action coverage. ## Verification - All 977 credential-free E2E support tests pass across 64 files, including 144 focused lifecycle/action/scorer/settlement calibrations. - E2E typecheck, ordinary plugin SDK and Runner TypeScript dependency builds, capability contract/inventory checks and the existing two-cell hiring discovery pass. - [Executable replay report](https://github.com/paperclipai/paperclip/blob/fed1729018cc100f5f4bbfb692777e49009c423b/doc/plans/2026-10-02-hiring-executable-accounting-replay.md) pins current code revision `e4077ade1818d98b9862ae79ee1d49a007dcf9c1`, v3 definition digest, exact source/input hashes and each original/new check. - The stricter replay verifies both Codex variants through both executable guards. ACPX Claude action attribution remains unresolved/uncomparable because provider execution IDs cannot be exactly joined to native request IDs; guards fail closed. No notification writes are observed. All six other outcomes and every original source/template coverage check stay unchanged. Original files and Fail → Fail machine verdicts remain preserved; zero providers are called. - Full attempts remain uncomparable in both profiles. Historical Claude also keeps its six-backtick exact-template mismatch. This grading repair does not prove model-performance equivalence. - The limited sidecar-v1 and initial executable-v2 passes checked notification-created tasks but could miss unrelated document writes. Those assessments remain preserved and do not prove harmless notifications. The stricter v3 replay is separate. - [Original measurement and separate sidecar](https://github.com/paperclipai/paperclip/blob/8eb517ca1497687237163bdef4dfc4d3332ea916/doc/plans/2026-10-02-hiring-template-live-comparison.md) retain 28 actual runs, eight automatic notifications, four successful cleanups and unknown actual model charges. No models are rerun. - The branch is replayed on master `59c07ede7`. Intervening master changes are UI-only; eval source bytes and replay verdicts match. The four-cell provider-free replay was repeated against the reachable code revision. - Initial-head normal CI retained browser failures in agent-run denial feedback and touch-picker scroll position. Those browser paths and imports were unchanged, but their cause was not established. The necessary review-fix head passes both browser checks; no blind rerun was requested. - Local full repository typecheck/test/build were not repeated. Exact-head normal CI passes the required repository gates, including typecheck, tests, build and browser shards. Fresh Greptile review completed on `fed1729018cc100f5f4bbfb692777e49009c423b` with 5/5 and zero unresolved threads. An independent rerun of the 144 focused helper/scorer/settlement tests passes on the unchanged head. **Merge readiness:** This PR repairs the evaluator. Its positive and negative calibrations pass, both guards reject missing action attribution, current-head CI and review pass, and there are no merge conflicts. The retained ACPX cells remain uncomparable because their action IDs cannot be joined. That coverage limit remains a separate follow-up; it does not require relaxing this grader or changing the old results. No model calls, production instructions, carrier changes, or historical regrades are part of this readiness update. ## Risks - Missing or inconsistent public lifecycle evidence fails the bounded helper. The focused calibrations reject plausible false positives and malformed observations. Unmatched action IDs fail closed and are reported as uncomparable rather than a model task regression. - Source-read evidence remains incomplete. This PR does not change provider event carriers or relax the coverage oracle. - The versioned count check differs from original v1 results. Reports retain both versions and exact input hashes. ## Model Used OpenAI Codex, GPT-6 family as identified by this session. The exact deployment ID and context-window size are not exposed. The assistant used reasoning, repository tools, code execution and delegated calibration work. ## 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 normal CI gates are green (exact head `fed1729018cc100f5f4bbfb692777e49009c423b`; fresh review tracked separately below) - [x] Greptile is 5/5 with no open P2s, recommendations, or follow-ups (completed exact-head review; zero unresolved threads) - [x] I will address all Greptile and reviewer comments before requesting merge --------- Co-authored-by: Paperclip <noreply@paperclip.ing>