wangxuming/makelore
1
0 Fork 0
Code Issues Pull Requests Actions Packages Projects Releases Wiki Activity
Files
d189181ceee2cf66dfcb52c68eaf3f7f7fc42002
makelore/.project-docs/30-worklog/tasks/20260825-pi-session-reopen-installer-8c3e7a91.md
T

8.9 KiB
Raw Blame History

Task: Fix persisted Pi session reopen and build Windows installer

Identity

  • Task ID: 20260825-pi-session-reopen-installer-8c3e7a91
  • Mode: Feature
  • Branch: codex/20260825-pi-session-reopen-installer-8c3e7a91-pi-session-reopen-installer
  • Worktree: D:\Datas\OthersProjects\makelore-pi-session-reopen-installer-8c3e7a91
  • Base commit: e74b59a674
  • Owner: codex-root
  • Status: Ready for Integration

Scope

  • Fix the packaged Pi runtime path that opens generation 1 without the Conversation's already persisted Pi session binding after an application restart or later re-entry.
  • Add focused regressions proving the managed opener reuses the Registry binding when PiWorkerPool does not provide an explicit existingSession, while a genuinely new Conversation still creates one binding.
  • Build and verify a Windows x64 NSIS installer from the cumulative hotfix base e74b59a67479a32619a9e3dbef476a65a6ad071f plus this repair.

Intent And Constraints

  • Preserve the user's existing Pi JSONL and Conversation metadata. Do not delete, rewrite, migrate, or silently replace a persisted binding.
  • Derive the effective reopen binding from the explicit worker-pool binding or the Main-owned Registry record, in that order. Keep the existing CODING_SESSION_UNREADABLE mismatch checks fail-closed.
  • Keep the repair inside the existing Pi managed opener/Registry boundary; do not add a fallback runtime, compatibility layer, or product migration.
  • Preserve new-Conversation allocation and fork semantics, and report a Registry-backed reopen as warm telemetry.
  • The previously diagnosed thinking-level submission stall is a separate task with no accepted implementation. Do not mix an unverified fix into this installer.
  • Use pinned pnpm 10.33.4, focused tests, full repository checks, build:vite, Windows Electron E2E, formal Windows packaging, and final artifact/runtime verification.
  • Real Provider verification remains Explicitly Waived / Accepted Risk with realTurnVerified=false; loopback/package evidence is not a real Provider Pass. macOS and native non-WSL Linux evidence status is unchanged.

Project Context Loaded

  • Concurrent Task Gate: Passed in the isolated worktree owned by codex-root; the task record, branch, worktree, and base commit match the registry.
  • Planning Gate: Passed after reading the project entry/memory indexes, current state, architecture/data-flow/evidence/commitment records, the installed session-unreadable diagnosis, the cumulative Windows proxy hotfix, the thinking-stall diagnosis, PI-150 packaging evidence, and all peer records still marked planning or blocked.
  • Canonical current-state documents predate the cumulative Pi delivery. The current source, diagnostic evidence, hotfix task record, and cumulative release-proof record are authoritative for this repair.
  • Other planning tasks are on older or unrelated AI Design/Robot/OpenCode baselines. The main-worktree OpenCode integration task is blocked and does not package; no unresolved semantic conflict or code-write overlap changes this plan.
  • No subagents are used, per user direction and repository policy.

Plan

  1. Add a focused red regression for a Registry-persisted binding with no explicit existingSession, including session-id argv and warm telemetry.
  2. Make the managed opener use one effective binding consistently for session selection, fork rejection, reopen mismatch validation, and telemetry.
  3. Run focused Pi tests, typecheck, lint, full unit tests, build:vite, and Windows Electron E2E; fix only task-caused failures.
  4. Commit the clean candidate, build the Windows NSIS, and run final Windows artifact/Pi/runtime/package smoke checks, including persisted-session reopen and fresh-Conversation binding behavior.
  5. Record the final installer path, size, SHA-256, exact evidence, and remaining accepted/deferred risks; complete the task documentation gate.

Outcome

  • Implemented one effective reopen binding in createPiManagedWorkerOpener: explicit worker-pool binding first, otherwise the persisted Registry binding. Session-key selection, reopen/fork validation, Pi session-id mismatch checks, and cold/warm telemetry now use that same binding.
  • Preserved the existing post-open Registry mismatch guard and session-file containment checks. No user session, Conversation schema, or migration path changed.
  • Extended the managed opener regression to simulate a process restart with a fresh PiSessionRegistry: the persisted binding is reused without invoking the new-session key generator, while the original first-open path still creates exactly one binding.
  • Committed the reviewed repair as 862d1b6d921f421ff0eaeb2f1df16274445dbc01 and built the Windows x64 product from that clean commit.
  • Generated the NSIS installer release/Makelore-2.0.0-win-x64.exe. The final unpacked app runs Main from app.asar, loads the packaged Pi 0.84.2 production closure, and passes the existing managed extension/subagent/Host-proxy product proof.
  • Ran a task-local, non-retained packaged restart proof against an isolated copy of the exact user Conversation and Pi session reported in the incident. Two separate launches of the final executable each hydrated 18 Conversation nodes with worker ready, run idle, and no public error. The source project/session and installed application were not modified.

Verification

  • Frozen install with pinned pnpm 10.33.4: passed; lockfile unchanged.
  • Red regression before implementation: failed at the exact existing Pi session binding does not match the Conversation registry branch.
  • Focused managed opener/Registry/runtime/worker-pool suite: 4 files / 23 tests passed after the repair.
  • corepack pnpm run typecheck: passed.
  • corepack pnpm run lint:check: passed with 0 errors and five pre-existing Renderer warnings.
  • corepack pnpm test: passed; 178 primary files / 1512 passed / 2 skipped, plus the isolated REN-008 pressure test 1/1.
  • corepack pnpm run build:vite: passed with only existing advisory warnings.
  • corepack pnpm run test:electron:windows: 2 files / 4 tests passed.
  • git diff --check, task ownership drift, and required project-doc checks: passed before packaging.
  • corepack pnpm run package:win: passed from clean candidate 862d1b6; staged 130 Windows x64 Pi production packages and six runtime assets, then produced the NSIS and blockmap.
  • Final installer: 211,888,196 bytes; SHA-256 00B474114B5B398DCA8374238A5C144E69A17E9419B21615143FCD59E4C62658. Blockmap SHA-256: AD2F44E242BF0A6C2C9C08FD3E068DBF4E2431D62663A93188E95D3DCA388439. Final app.asar SHA-256: 6F5B95F653ED346DD1616C403584EC61F23A2AE6907DB47D27FE28EE5DA4AF09.
  • corepack pnpm run verify:publish-runtime: passed; bundled npm 11.6.2.
  • corepack pnpm run verify:artifact:win: passed and bound the artifact to commit 862d1b6; Electron 43.4.0 / Node 24.18.1, bundled Python/uv/npm, msgpackr, and Canvas native assets loaded from the final product.
  • corepack pnpm run verify:artifact:pi -- --samples 2: passed; Pi 0.84.2, CLI, all 130 expected packages, six runtime assets, five unpacked native assets, managed Skills, extension markers, session persistence, local worker overlap/abort isolation, cleanup, and cold/warm budgets were verified. The nested report correctly remains partial-pass for the pre-existing external platform and real-Provider evidence boundaries.
  • corepack pnpm run test:pi-subagent:packaged: passed from final win-unpacked/Makelore.exe. Main executed from app.asar; cold/warm managed turns, a real packaged parent and ephemeral child, 4+4 process pressure, authenticated Host-proxy loopback, failure-safe cleanup/retry, UI interactivity, token non-disclosure, and zero released resource counters all passed with realTurnVerified=false.
  • Final persisted-session packaged proof: two independent launches both returned worker ready, run idle, 18 hydrated nodes, and no error. The temporary verifier and isolated user-data copy were removed after the run.
  • Final Windows process query found zero residual processes from this worktree.
  • The local installer has no Authenticode signature (NotSigned); no signed release claim is made and Windows may show a publisher/SmartScreen warning.

Follow-ups

  • The unsupported-thinking-level / indefinite 正在提交… issue remains a separately diagnosed defect. This package does not claim that unimplemented fix.
  • Real external Provider validation remains explicitly waived accepted risk, with realTurnVerified=false; packaged loopback evidence is not a real Provider Pass.
  • macOS and native non-WSL Linux release evidence status is unchanged by this Windows-only repair.

Promotion Candidates

  • Promote the invariant that generation-1 worker open must consume an existing Main Registry binding, plus the packaged restart/re-entry regression, into the Pi release runbook when the cumulative Pi branch is integrated.
Reference in new issue View Git Blame Copy Permalink