fix: pin Codex compact-kill waiter rejection

[ladydd]

Related issue

Closes #550

Depends on LodyAI/acp-extension-codex#38. Review this host PR, then merge the adapter first and this pin second. Do not ship a desktop release until that adapter merge exists.

Problem / pressure

/compact waits on a thread/compacted notification. Turn waiters already reject when the Codex process exits; compaction waiters were resolve-only. If app-server dies after thread/compact/start, the ACP prompt never finishes. Host Stop does not clear it. The next prompt is already active. Healthy compact still works. Mid-turn process death already returns -32603. This is the compact waiter hole, not the fork hang (#543 / #544) and not adapter #30 (stop during compact).

Summary

Pin acp-extension-codex to 31c5ecc (adapter #38, rebased onto current adapter main / 314b3823) so close/dispose rejects compaction waiters. Other submodules stay on current main.

Visual explanation

sequenceDiagram
  participant Host
  participant Adapter
  participant Codex
  Host->>Adapter: prompt /compact
  Adapter->>Codex: thread/compact/start
  Codex-->>Adapter: start ok
  Note over Codex: process exits
  Note over Adapter: before: wait forever for thread/compacted
  Host->>Adapter: session/prompt
  Adapter-->>Host: already active
  Note over Adapter: after: close rejects waiter, prompt returns -32603

Loading

Before / after

Before	After
/compact + process death stays compacting; Stop then already active	Prompt returns -32603; next prompt is not already active
Healthy /compact	Unchanged
Main pins Codex 314b3823 (host #275 / adapter cancel-compaction)	Codex 31c5ecc (#38 on top of that main)

Test plan

• Adapter unit: vitest run src/__tests__/CodexAppServerClient.test.ts --no-file-parallelism --retry=0 — 7 passed, including close while start is still in flight.
• Independent adapter check (Codex 0.153.4, synthetic model, rebuilt bundle of this patch): compact-kill returned -32603 in 10ms with no already active; restart + loadSession then history, a normal prompt, and another compact all completed. Lifecycle 8/8 included real vscode-jsonrpc close while start was pending. Related adapter regression 130 passed; the known /review slash-command timeout is a pre-existing baseline skip.
• Host change is the gitlink plus Agent Notes. pnpm run docs check reported no errors on the new note pair; the command still fails on pre-existing broken links in unrelated notes/specs. Full pnpm check was not re-run on this pin-only host worktree.
• Not claimed: in-window Electron compact-kill on this host pin. Acceptance was adapter protocol plus host -32603 classification.

Context handoff

Instructions for reviewing agents

• Review focus: The Codex gitlink 314b3823 → 31c5ecc and that Core / other submodules are unchanged.
• Decisions to challenge: Pinning unmerged adapter Fix/electron local file browser #38 (rebased onto adapter main after fix: avoid PR polling for direct local projects #30) versus waiting for that merge; whether this overlaps adapter fix: avoid PR polling for direct local projects #30 cancel-during-compact.
• Plausible failures / evidence gaps: CI submodule fetch of 31c5ecc before Fix/electron local file browser #38 merges; no in-window desktop compact-kill on this pin.

Authoring context

• User goal / directives: Land the host pin for the Codex compact-kill hang so desktop builds pick up adapter Fix/electron local file browser #38 after that adapter merges, same two-PR shape as feat: upgrade public workspace to React 19 #37 / fix: forked Codex sessions hang on the first prompt #544.
• Constraints / non-goals: No other submodule refresh; no adapter publish; no desktop release from this PR; do not mix fix(cli): do not let Codex explicit titles replace generated names #536.
• Risk-bearing decisions: Host points at an open adapter PR SHA (31c5ecc / Fix/electron local file browser #38).
• Destructive or irreversible behavior: Gitlink move only. No data migration, no history rewrite.
• Deliberately not done or tested: In-window Electron compact-kill on this host pin; full host pnpm check on this worktree.
• Unknowns / confidence: High on the adapter path from independent real Codex evidence; merge order must keep adapter Fix/electron local file browser #38 first.

Original user prompt

Codex compact-kill should reject waiters instead of hanging.

[ladydd]

Retargeted the gitlink onto current main: 5f0aab0f (#554) → a67f231 (adapter #38 rebased onto that main). Notes updated. Still depends on adapter #38 merging first; other submodules unchanged.

[ladydd]

Rebased onto current main. Codex gitlink is now 314b3823 → 31c5ecc (adapter #38 rebased onto adapter main after #30). Adapter tests: CodexAppServerClient 7 passed.