fix(pi): serialize write-leased tool batches

This commit is contained in:
inman
2026-09-02 12:07:13 +08:00
parent 6073bd6f4c
commit 49112b688d
7 changed files with 304 additions and 4 deletions

View File

@@ -0,0 +1,84 @@
# Task: Fix Bash tool execution bridge
## Identity
- Task ID: 20260902-fix-bash-tool-bridge-9c4e7a12
- Mode: Feature
- Branch: codex/20260902-fix-bash-tool-bridge-9c4e7a12-fix-bash-tool-bridge
- Worktree: /Users/inmanx/Documents/makelore-fix-bash-tool-bridge-9c4e7a12
- Base commit: 6073bd6f4ce269b067b13e1eed0371b38dc2c054
- Owner: codex
- Status: Completed
## Scope
- Reproduce and fix the Makelore-managed Pi tool-batch deadlock that delays
parallel Bash calls until the five-minute HTTP idle timeout.
- Preserve project-scoped write-lease isolation while making Bash/edit/write
tools execute sequentially inside one assistant tool batch.
- Add real Agent Server coverage for two Bash calls, including a command-level
timeout, and verify the same resource path used by packaged builds.
## Intent And Constraints
- Keep Pi `0.84.2` as the sole runtime and keep Renderer isolated from Pi wire.
- Do not replay an accepted mutation or weaken the cross-Conversation/project
write lease; only remove the same-batch pre-execution self-deadlock.
- Retain parallel execution for batches containing only read-only tools where
Pi supports it.
- Treat packaging as a verification path, not as a substitute for the source
correction.
## Outcome
- Reproduced the stall with the real Pi Agent Server process. Pi `0.84.2`
prepares every tool in a parallel batch before it executes any of them, while
Makelore acquires the project write lease in `tool_call` and releases it in
`tool_result`. Two Bash calls in one assistant response therefore self-deadlocked:
the first held the lease without starting and the second waited for that lease.
- Confirmed that the apparent five-minute Bash hang came from the Agent Server
HTTP idle timeout around the managed-extension request, not from a missing
shell dependency or the Bash command timeout. Packaging alone would retain the
same behavior because it ships this Agent Server resource and extension bundle.
- Marked built-in `bash`, `edit`, and `write` tools as sequential whenever an
Agent thread is rebound. Marked dynamically declared product tools that require
the same project write lease as sequential in the generated Makelore extension.
Read-only tool batches keep Pi's parallel behavior.
- Advanced the managed extension artifact from v4 to v5 and synchronized the
release proof and worker-opening expectations.
- Added a real-process regression that emits two Bash calls in one assistant
response. The first exercises its explicit command timeout and the second must
still run successfully without waiting for the HTTP idle timeout.
## Verification
- Pre-fix regression: the new real-process test failed with the second Bash result
equal to `fetch failed`, reproducing the production failure under a 250 ms HTTP
idle timeout.
- Focused regression: `pnpm exec vitest run
tests/unit/pi-agent-server-process-real.test.ts
tests/unit/pi-extension-bundle.test.ts
tests/unit/pi-managed-worker-opener.test.ts --maxWorkers=2` passed 3 files and
9 tests.
- `pnpm run typecheck` passed.
- `pnpm test` passed 213 files and 1,733 tests; 3 tests were skipped.
- `pnpm run lint:check` passed with 0 errors and 5 existing warnings in
`src/pages/Home/index.tsx` and `src/pages/Makelore/index.tsx`.
- `pnpm run build:vite` passed for Renderer, Electron Main, Preload, and the
release utility worker, proving the fixed Agent Server resource is included by
the production build path.
- `git diff --check` passed.
## Follow-ups
- Integrate this feature branch into local `main`, update the canonical current
state/README, restart Makelore, and use the real-process regression as the
non-destructive smoke check instead of replaying the user's accepted turn.
## Promotion Candidates
- Promote the write-lease batch rule to README/current state: write-leased Pi
tools are serialized within one assistant batch, while read-only tools may
remain parallel.
- Record the full-suite, production-build, and real-process regression results in
the evidence index during integration.