mirror of
https://github.com/paperclipai/paperclip.git
synced 2026-10-17 20:09:05 +02:00
d4bb28e0f4d00a98cc1d86181cda2c09c0de1ebf
4
Commits
| Author | SHA1 | Message | Date | |
|---|---|---|---|---|
 DottaandPaperclip
|
428b617fc9 |
test(evals): qualify native question resume paths (#15616)
## Thinking Path > - Paperclip lets agents pause tasks for human answers and continue the work. > - Native questions can yield a run or pause inside the provider's running turn. > - Those paths have different run identities and different forms. > - The documentation-placement test requires a semantic three-turn journey. > - A valid provider question therefore needs separate behavioral qualification. > - This pull request adds an opt-in two-answer journey with exact source and resume checks. > - Qualification exposed a Vite startup crash in idle-handler traversal, so the branch also distinguishes Connect mount paths from Express Route objects. ## Linked Issues or Issue Description **What existing behavior does this improve?** Product E2E evaluation of native task questions and answer delivery. **Current behavior** The semantic documentation case rejects the provider adapter's optional Other field before answering. Its three-run expectation also cannot prove a provider question that resumes within the same run. **Proposed behavior** Keep that semantic test intact. Add a separate opt-in case that classifies each recorded question by authoritative identities. Accept the optional Other field only for a verified provider question. Require both user answers, the correct continuation, and one saved final document. Related: #15564 added strict lifecycle checks. #15522 introduced idle-request tracking; this PR corrects its Connect/Vite compatibility after reproducing the startup crash. #14591 changes ACP question drafts and provider support; this PR has no overlapping files with that change. ## What Changed - Add the explicit-only `question-resume` suite for native Codex and ACPX Claude. Reuse the existing neutral choice-then-text task request. - Bind each question to its company, task, agent and run. Require an applied semantic tool receipt or the exact provider request identity. - Require a provider answer to continue the same run. Require a semantic answer to start one new run with the matching interaction and source-run wake fields. - Preserve question forms and both real board answers through the final checkpoint. Reject missing inputs, duplicate answers, extra runs and substituted identities. - Keep the existing lifecycle, output, task ownership and budget checks. Limit the new case to one attempt and one to three runs. - Add mutation calibrations and document the separate qualification claims. Keep production guidance, scheduling, provider code and the historical semantic case unchanged. - Fix startup with Vite middleware: treat Connect string mount paths separately from Express Route objects when traversing idle-request handlers. A real-Vite regression verifies startup and retained async work through an idle hold. ## Verification - 81 focused continuation, question and calibration tests passed before review. After the reference-preservation correction, all 50 question/resume tests pass, including two new negative cases that failed against the previous grader. - Final eval support suite passes: 1,921 Vitest tests and 128 Node tests, with one intentional skip. - Eval and workspace typechecks pass. Workspace build passes. - Catalog discovery finds only the two declared opt-in cells. Existing continuation remains 23 cells; default paid scope is unchanged. - Original [campaign 37831726391](https://github.com/paperclipai/paperclip/actions/runs/37831726391), source `1ac4b2873466896ff2892f49be82904927b6c57d`, failed during server startup before any test or provider run. Its original `transient_infrastructure` result and `cleanup: not_started` are preserved. Local reproduction identifies the Connect string route traversal crash; the question grader was never reached. - Setup-corrected [campaign 37833872898](https://github.com/paperclipai/paperclip/actions/runs/37833872898) measures source `9781a772389d41421df55759004676e1b905e512` with trusted workflow `8e59efc50b162126f33816841e06509634eba65c`. The workflow bytes are unchanged from the first dispatch. Same single selected native Claude cell, one attempt, at most three runs, ten-minute cell deadline, and 1,000-cent company/agent hard stops. Original result **PASS**, 24/24 checks and cleanup pass. Exactly three successful native Claude runs: two applied Paperclip `request_human_input` questions and one finishing run. Both real board answers persist, exact interaction/source-run response wakes match, one revision-1 task document contains Afternoon and the supplied reference, and final task state is done with no active lock, retry, recovery, monitor, pending interaction or child task. No model retry. - Current review head `3c08cb7d26e7f9ac16469eeb07d1f31d828ec02e` differs from the measured source only in the question/resume grader and its tests (11 insertions, 3 deletions). The saved document must contain the complete submitted reference, including its prefix and punctuation. Separate provider-free replay of the retained successful campaign passes all 24 continuation/lifecycle checks under this stricter grader. The original result and its measured source remain unchanged; no new provider run was made. - Observed paths were **semantic → semantic**. The provider built-in question/optional Other path was not exercised by this new live attempt. Its form acceptance has retained-original calibration; same-run answer/resume has deterministic positive/negative calibration. This neutral-task pass alone does not qualify the provider path; the subsequent explicit bridge trial below is separate evidence. Neither trial regrades the old Claude failure or establishes a causal behavior/performance comparison. - Subsequent explicit provider-path [campaign 37841107684](https://github.com/paperclipai/paperclip/actions/runs/37841107684) measures the final PR source `3c08cb7d26e7f9ac16469eeb07d1f31d828ec02e`, using trusted workflow `2a5f65c9ae501b49b2a38ce7edd04209b806c9d0`. Exact cell: `continuation.runner-acpx-claude.local.provider-question-bridge`, Claude `claude-sonnet-5`, suite fingerprint `fc77203bf85fd99b06ebc53cd52982186c0ae0dc369fcc38a9ec3c737e9a31e5`. **Original PASS, 15/15 checks, cleanup passed, one actual provider run and zero eval rerolls.** The built-in question produced one choice plus its optional Other companion. The browser selected the requested reference, left Other blank, and submitted. Exact runtime request/response IDs and the selected option match the saved interaction; the same paused native run created one revision-1 document and finished Done with no pending interaction, lock, retry, recovery, monitor or child task. Independent evidence and runtime-receipt audits pass; marked initial/final screenshots were visually checked. [Original public report](https://d1p6rlowie26tp.cloudfront.net/runner-e2e/campaigns/gha-37841107684-1/index.html). - This separate provider-directed trial qualifies one choice-answer bridge and traversal of an empty optional Other field. It does not establish natural tool selection, a typed Other answer, two consecutive built-in questions, free-text-only input, restart recovery or general reliability. No production, prompt or grader change was made for this dispatch. The old failure and the separate neutral semantic-path result retain their original grades. - Preserve within-run friction: one `write_task_document` input-schema rejection and two HTTP 409 “Document does not exist yet” responses preceded a successful `write_document` call. These are three unsuccessful tool calls inside the same run, not three extra provider runs; no flawless document-tool behavior is claimed. Billing reports token usage but remains unpriced/incomplete, so actual charges are unknown and local/hosted runtime is unmetered. Both company and agent hard stops were 1,000 cents, with one attempt, one selected cell, concurrency one and a ten-minute cell deadline. - Billing records all three model runs with token usage, but no priced dollar receipt (`unpriced`, `complete: false`). Actual charges are unknown; local/hosted runtime is unmetered. Numeric zero reported cost does not mean free. - Startup repair: real-Vite reproduction passes; 35 targeted idle-tracking tests pass, 34 database-dependent checks skip because embedded PostgreSQL is unavailable locally. Workspace typecheck and build pass again after the repair. - Full local repository database tests were not repeated because embedded PostgreSQL was unavailable in the preceding workspace verification. Full Linux CI now passes on the final review head. - Final-head [CI run 37837804284](https://github.com/paperclipai/paperclip/actions/runs/37837804284) passes. Complete current-head audit: 53 successful check-runs, two intentional Storybook skips, and separate Snyk success. The ready transition’s contributor and security checks also pass (security scan: no findings). [Fresh review](https://github.com/paperclipai/paperclip/pull/15616#issuecomment-6068092601) is 5/5 on `3c08cb7d26e7f9ac16469eeb07d1f31d828ec02e`; zero unresolved threads and no merge conflicts. ## Risks This is a new behavioral definition. The neutral two-question trial records semantic-tool use; the separate provider-directed trial records one built-in question. They do not establish consistent tool selection, documentation placement, default hiring, remote environments, crash recovery or general reliability. The separate explicit bridge trial covers one provider choice, an empty optional Other companion and same-run completion. Typed Other, two consecutive built-in questions and free-text-only provider input remain unqualified. The observed document-tool schema rejection and two 409 responses are retained as a separate follow-up, without attributing their cause. The historical case and its original grades remain intact. The reference question must still be text-only; the oracle does not turn an arbitrary extra question into an optional companion. Missing or unfamiliar identity evidence fails closed. The only production change distinguishes Connect string mount paths from Express route objects during idle-handler traversal. Unknown route objects still fail closed. No API, schema, migration or workflow change is included. ## Model Used OpenAI Codex, GPT-6-based, with tool use and code execution. The exact serving model ID 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> |
||
|
DottaandPaperclip
|
24a15c5afa |
test(evals): verify durable waiting across continuation checkpoints (#15564)
## Thinking Path > - Paperclip manages agents and durable tasks across provider runs. > - Human questions must preserve task ownership and stop work until a real answer arrives. > - The continuation eval checks the final work, but its lifecycle oracle misses several broken waiting states. > - A correct final answer can hide a stale execution lock or a lost intermediate answer receipt. > - This pull request checks each wait and retains every question and run identity. > - A source audit records which waiting operations belong to native runtimes and which still require legacy API calls. ## Linked Issues or Issue Description **What existing behavior does this improve?** The existing Product E2E continuation oracle for human question and approval journeys. **Current behavior** A pending interaction can pass the waiting check even when its task has the wrong status, a stale lock, or a retry. Only the first pending question and first run receipts are checked at the final checkpoint. **Proposed behavior** Require a healthy wait on the same task and assignee. Preserve all intermediate run receipts and every pending question's answered identity. Allow a paused native provider question only when its pending runtime request identifies the running native run and execution lock. Related: #15548 and #15554 cover earlier bookkeeping slices. #15544 changes production continuation summaries; this PR changes the eval oracle and does not overlap that fix. ## What Changed - Check waiting state, execution ownership, and all answer/run identities in the existing lifecycle oracle. - Add negative calibrations for broken records and positive coverage for both semantic waits and paused provider questions. - Retain task activity at each checkpoint for inspection of successful persisted mutations. - Verify 1,000-cent company and agent budget limits before continuation work. Disable automatic cell rerolls. - Document runtime ownership, existing coverage, and the remaining instruction decision. Keep production instructions unchanged. ## Verification - `pnpm test:e2e:runner:unit`: 1,867 Vitest tests pass, one skips; 128 Node tests pass. - `pnpm test:e2e:runner:typecheck`: passes. - `pnpm -r typecheck`: passes. - `pnpm build`: passes. - Negative calibration: 27 added cases fail against the prior oracle and pass with these checks. - Four existing local continuation cells are selected for a separate bounded live canary. Live results are pending; no behavioral pass is claimed here. - The full local `pnpm test:run` suite was not repeated because embedded PostgreSQL was unavailable in the preceding workspace verification. Required Linux PR CI must pass before readiness. ## Risks The stronger oracle can expose existing product or fixture defects. A paused provider question and a terminal semantic wait have distinct valid states. The four-cell canary does not qualify approval/review, dependency unblock, crash races, remote execution, or general task quality. Task activity records successful persisted writes, not failed API attempts; repeated progress comments are not automatically defects. Historical eval grades remain unchanged. No production scheduling, prompt, tool, schema or migration changes. ## Model Used OpenAI Codex, GPT-6-based, with tool use and code execution. The exact serving model ID 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 - [ ] 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> |
||
|
DottaandPaperclip
|
43acbcc398 |
fix(runner): preserve sessions and complete question and approval continuations (#13655)
## Thinking Path > - Paperclip is the open source app people use to manage AI agents for work. > - The native runner connects task state to provider sessions. > - Follow-up turns must retain provider memory and carry new user direction. > - Lost session IDs caused repeated context and extra input tokens. > - Native question answers and approval races could leave valid work blocked. > - This pull request repairs those paths and adds regression coverage. > - Agents can continue accepted work without repeating the conversation or losing the user's answer. ## Linked Issues or Issue Description Refs #13574. That merged PR shortened continuation prompts and moved question instructions into tool documentation. This change preserves sessions and fixes failures exposed by broader testing. Related runtime work: #13408 and #13410. **What happened?** Native follow-up turns could lose the provider session ID. Completion guidance could replace the original task with its latest comment. Claude native questions could remain pending after the user answered. Approval during a running tool call could suspend the run before the tool response arrived. Onboarding and chat handoff instructions also caused repeated planning or missing plan documents. **Expected behavior** Reuse a valid provider session. Send only new events when that session already has the history. Preserve the task requirements and apply later user direction. Store the question answer and deliver it to the waiting run. Finish governed tool responses before suspending. Execute the accepted plan without asking for the same approval again. **Steps to reproduce** Run the continuation, local-session-integrity, first-task, and agent-chat suites with native Codex and Claude. Include provider-question-bridge, accept-while-running, and plan-handoff. **Paperclip version or commit** This branch is based on master |
||
|
DottaandPaperclip
|
e26d787928 |
Shorten continuation prompts and verify question tool guidance (#13574)
## Thinking Path > - Paperclip is the open source app people use to manage AI agents for work. > - Agents must continue tasks using user answers without losing earlier requirements or approval gates. > - The wake prompt mixed human decisions with prior tool evidence and repeated detailed question instructions. > - Those instructions belong with the question tool, with a short routing hint in the wake. > - The Runner evals need to prove that answers, approvals, and completed work survive later turns. > - This PR shortens the prompts, separates authenticated answers, and adds continuation tests with useful screenshots. ## Linked Issues or Issue Description Refs #13517. This is a follow-up to the merged onboarding skill and Runner E2E work. Related #13539 covers responses received while a run is active; this PR preserves its cases and adds continuation coverage. Existing continuation/recovery and question PRs were searched; none covers this prompt/documentation and eval change. **What existing behavior does this improve?** The instructions sent when an agent continues a task, the native human-input tool documentation, and the evidence captured by Runner full-stack E2E. **Current behavior** The wake repeats a long question-tool guide. Human answers appear alongside untrusted prior results. Screenshot capture can finish at DOM load while the task still shows a spinner, even when backend behavior checks pass. **Proposed behavior** Keep earlier requirements unless the user changes them. Treat clarification as distinct from approval. Give authenticated human responses a scoped field. Keep tool and agent results as evidence. Put detailed question behavior in the tool descriptor and retain one routing sentence in the native wake. Wait for the correct task and loaded conversation before taking screenshots. **Reason and benefit** Reduce repeated prompt text and make authority boundaries clear. Test that real question cards, later answers, approval gates, and completed child tasks still work. Make screenshots useful for human review. ## What Changed - Shorten shared continuation instructions for legacy and native runners. Separate authenticated user responses from tool results and agent summaries. - Remove the detailed question guide from native wake prompts. Keep its behavior in the canonical `request_human_input` descriptor and existing payload schema. Regenerate semantic contracts and fixture hashes. - Add five continuation cases across four local profiles. Add a dedicated choice-then-text case for native Codex and native Claude. All 22 cells join the shared full E2E campaign. - Cover revised scope, clarification without approval, hostile instructions in a handoff file, and reuse of a completed child after restart. Keep production instructions and fixed user facts. - Capture continuation screenshots only when the intended task and conversation have rendered. Add provider-free browser regressions for loaders and wrong-task capture. - Preserve current master’s extra tool and onboarding cases. The default campaign now contains 166 cells; 35 manual everyday cells remain separate. ## Verification - `pnpm -r typecheck`: passed after replay on current master. - `pnpm test:e2e:runner:unit`: 340 passed. Harness typecheck passed. - `PAPERCLIP_PLAYWRIGHT_CHANNEL=chrome pnpm test:e2e:runner:browser-support`: 4 passed. These tests failed against immediate screenshot capture and passed after the fix. - Focused continuation and native-input tests: 36 passed locally. The tool-authority suite could not initialize embedded PostgreSQL locally, including one isolated retry; its 17 assertions did not run locally. The full remote server shards passed on this PR commit. - `pnpm build`: passed after replay on current master. `pnpm test:run` was attempted locally but hit the same embedded PostgreSQL initialization failure; the remaining local run was stopped after complete remote CI passed. This is not claimed as a full local test pass. - [Full PR CI](https://github.com/paperclipai/paperclip/actions/runs/35232755685): passed on `6a22128c14f4552d0613a6d9a25955db4a1ed02f`. All server/chat/workspace/serialized shards, browser shards, Runner checks, typecheck, build, canary and policy checks passed. The isolated native Runner build and security checks also passed: 57 successful checks, with two expected Storybook skips. - Greptile reviewed the exact PR head at 5/5, with no findings or unresolved review threads. The PR has no merge conflicts. - [Live question-docs report](https://pages.paperclip.ing/runner-e2e-question-docs-35227647794/): 3/3 passed at source `83dd132f2` before replay on master. Native Codex and Claude each asked a choice, waited, asked a text question, and saved both answers. Claude also passed a completed-child restart case. All three native turns are checked for absence of the old question block. - [Earlier continuation report](https://pages.paperclip.ing/runner-e2e-continuation-35154943615/): all five continuation cases passed on native Claude. The report retains campaign and revision provenance and separately shows two unresolved onboarding behavior failures. - [Before/after prompt report](https://pages.paperclip.ing/runner-prompt-comparison-20260917/): full text, current recorded Claude inputs, and reproducible reference-token counts. The controlled wake comparison removes 401 reference tokens; the net counted input reduction is 339 after charging the larger tool description. These are text-size estimates, not measured billing savings. ## Risks - Prompt wording affects model behavior. Live results cover the stated cases, not every provider or conversation. Legacy profiles are registered but were not rerun for this change. - The optional continuation field changes prompt data only; there is no database migration or new production API. - Authenticated answer projection excludes generated summaries and agent-resolved interactions. It preserves the answer’s question or approval scope. - The screenshot guard can expose UI loading failures that earlier runs hid. Backend grading alone no longer makes those captures valid. - The two prior onboarding failures remain separate product issues: work before acceptance and a missing saved plan. This PR does not claim the entire onboarding suite passes. ## Model Used OpenAI Codex, GPT-6, with reasoning, repository tools, code execution, and browser verification. The exact deployed 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 — targeted tests above; the full local database-startup limit is documented - [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> |