mirror of
https://github.com/paperclipai/paperclip.git
synced 2026-10-29 04:25:52 +01:00
Nicky LeachandPaperclip
a2fbce9305
## Thinking Path > - Paperclip preserves task state when agents receive follow-up work. > - Plugin computers keep their process state outside the controller. > - A local receipt proves that a previous plugin run suspended. > - A cold resume left that receipt in place after the session became live. > - The next warm follow-up rejected the stale receipt as an identity mismatch. > - A separate follow-up path also treated a fresh task directory as a Git worktree. > - This PR consumes the obsolete receipt after admission and preserves the bound directory on follow-up. ## Linked Issues or Issue Description Related: #17820. This change retains the controller restoration added by #18018. **What happened?** A plugin session resumed from a suspended checkpoint and completed a turn. Its next warm follow-up failed with `runner_state_identity_mismatch: prior_owner_terminal_state_indeterminate`. The journal had the completed run's identity. The local suspension receipt still had the preceding run's identity. Two separate task histories showed this sequence. After a fresh retry completed in a repository-free plugin directory, an ordinary follow-up could instead fail with `The selected plugin environment does not support Core Git worktree preparation.` The saved task settings contained only `mode: isolated_workspace`, which introduced the default Git strategy on the next admission. **Expected behavior** Keep the same admitted provider session during warm follow-ups. Reuse the saved plugin task directory without Core Git preparation. Retain ownership checks and explicit strategy settings. **Steps to reproduce** 1. Complete a warm plugin task and let its session suspend. 2. Resume the checkpoint and complete another turn. 3. Send a follow-up while that session remains warm. 4. Separately, complete a fresh retry in a repository-free plugin directory and send an ordinary follow-up. **Paperclip version or commit** Both regressions reproduce against base `820c45846` with the added tests. **Deployment mode** Hosted plugin execution. Local reproduction uses the real filesystem and a disposable PostgreSQL 18 database, with synthetic provider execution. ## What Changed - Remove the local plugin suspension receipt after provider recovery and durable run admission publish a live session. Admission still checks the receipt first. Normal suspension writes a new receipt. - Recognize a saved fresh plugin directory through its company, task, workspace, owner, attempt and path. Use the existing directory-selection path. Explicit strategy settings still apply. - Add cold-to-warm, database-backed workspace-selection and ownership regressions. Update execution semantics and session verification recipes. ## Verification - Red: the cold-to-warm regression failed on unchanged executor code with `prior_owner_terminal_state_indeterminate` on the follow-up. - Red: the PostgreSQL regression selected a Git strategy on unchanged environment-selection code. The explicit-strategy denial case already passed. - Green: all 95 tests in `workspace-preparation.test.ts` and `provider-fresh-workspace.test.ts` pass against a disposable local PostgreSQL 18 database, selected with the existing `PAPERCLIP_WORKSPACE_PREPARATION_TEST_DATABASE_URL` test seam. - Green: 92 focused executor cases pass, including plugin checkpoint recovery, warm rotation and denial cases. The full `plugin-runner-task.test.ts` suite also passed (38 tests). - Green: direct server check, `node node_modules/typescript/bin/tsc -p server/tsconfig.json --noEmit`. - Green: `node scripts/check-maps.mjs`, `git diff --check`, and a local diff scan for secrets, personal data and live identifiers. - Full local native suites: 988 of 1,004 cases passed. Sixteen integration cases could not start their bundled PostgreSQL cluster. The failure persisted outside the sandbox. These cases are not claimed as passing. An earlier sandboxed attempt also denied process and loopback operations; those denials disappeared when run outside the sandbox. - The first broad local commands could not locate Cargo on `PATH`. Rust is installed under `~/.rustup`; its Cargo executable was verified. Per the requested scope, subsequent local verification stays targeted to changed TypeScript paths; no Rust source changed. CI Typecheck + Release Registry and Rust verification passed on the initial commit. The complete repository test suite was not run locally. - Review regression: an explicit agent-level Git strategy reproduced the incorrect directory fallback before correction. After including the raw agent strategy in the existing check, all three targeted cases (implicit, task-level explicit, agent-level explicit) pass. The server TypeScript check passes after this correction. - CI on `a81668c4b`: the first workspace-a 2/2 attempt failed in unchanged UI `AgentActionButtons.test.tsx` (`calls clearError and refreshes agent-related queries`, zero mock calls). [Original failed job](https://github.com/paperclipai/paperclip/actions/runs/40717870938/job/122921463517). That shard passed on `294c5dfc3`. The failure is retained here. The UI shard passed on final commit `687000287`; no unrelated UI edits or local UI suite runs were added. - CI also exposed three existing local-directory recovery failures on `a81668c4b`: the review correction applied the raw agent strategy guard to local directories. [Original failed server job](https://github.com/paperclipai/paperclip/actions/runs/40717870938/job/122921463229). Narrowed that guard to adapter-managed directories, preserving local recovery behavior. All 20 targeted cases (three plugin selection cases and 17 local recovery scenarios) pass afterward, and server TypeScript verification passes. - Final commit `687000287d3339b7e15e35d95515b3c75c273d7a`: [hosted CI is green](https://github.com/paperclipai/paperclip/actions/runs/40720169342), including the corrected server shard, all other server/workspace shards, Runner checks, typechecking and build. All 55 non-skipped PR checks pass. Greptile scores this exact commit 5/5, with no unresolved review threads. GitHub reports no merge conflicts. - Pre-merge hosted acceptance was requested. The maintained deployment CLI stopped during artifact discovery: no published private Cloud composition exists for `687000287`. The current private build requires canonical Core master readiness and standard-image provenance. The Core-only branch preview omits the private computer plugin needed for the journey. No campaign or tenant update was started; private-preview deployment support is a separate prerequisite. No live deployment or provider acceptance run is claimed. ## Risks The receipt is removed only after the existing runtime publishes an admitted session. Earlier admission failures keep their proof. A later normal suspension creates a new receipt. The change preserves existing identity, lifecycle and controller checks. Directory selection remains limited to the exact bound, active, repository-free plugin recovery directory. Foreign owners, changed paths, wrong attempts and explicit Git strategies do not gain this fallback. No schema, API, credential, permission or fleet-default change is required. Existing failed runs and recovery holds remain history. This PR does not edit quarantined state or replay an interrupted task. Rollback restores the previous code and can reintroduce the two admission failures. ## Model Used OpenAI GPT-6 through Codex, with reasoning, tool use and code execution. A more specific 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>