- log the worker thread's exception when a workspace op raises during
caller cancellation, so a persistently failing setup surfaces instead
of being buried behind CancelledError.
- raise on a timed-out (124) git symbolic-ref probe in the repo-exists
path, matching the rev-parse probes, instead of silently accepting the
caller-supplied branch.
- assert the subprocess timeout is passed in the two _chown_workspace
test fakes so a refactor cannot silently drop the bound.
Op: correct
Restores: spec:indeterminate-git-probes-raise-not-silently-proceed
Workspace setup/teardown (git clone/fetch, worktree add/remove, chown)
ran synchronously on the asyncio dispatcher loop, so one stalled
subprocess froze the entire process.
- Offload every ensure_workspace/remove_workspace call to a worker thread
via a new _run_workspace_op helper that drains the thread to completion
on cancellation, so a cancelled event cannot reap/release a slot the
setup thread still owns.
- Serialize same-repo setup with a per-repo threading.RLock while letting
distinct repos run concurrently.
- Bound the direct git/chown subprocesses with a 120s timeout
(returncode 124); treat a timed-out branch probe as an error rather
than "branch absent" to avoid silently rebasing a follow-up onto the
default branch and losing the PR's commits.
- When a timed-out worktree remove leaves the checkout behind, rmtree it
and run `git worktree prune` so the pool's dangling registration cannot
trip a later worktree add for the same path.
Adds regression tests for event-loop liveness, cancellation-safe offload,
per-repo lock serialization, subprocess timeout mapping, the branch-probe
timeout guard, and worktree-prune after a failed remove.
Op: correct
Restores: spec:dispatcher-event-loop-never-blocks-on-workspace-io
- Restored CustomInputRow.priority field dropped in 3b80dc01d ask row budgeting.
- Narrowed dereferenced schema properties via isRecord in yield-assembly and output-schema-validator instead of untyped object access.
- Renamed stale advisorReadOnlyTools to advisorTools in advisor parity test.
- Narrowed AgentMessage content access in session-loader-stream test.
- Reformatted browser-schema test to satisfy biome.
- Removed a duplicated branch.reverse() left by the PR #3862 merge in
SessionEntryIndex.pathTo(), which returned branches leaf-to-root and made
getLastModelChangeRole() read the oldest model change instead of the
newest — pinning the ctrl+p cycle to one slot and breaking session model
restore.
- Hardened getRoleModelCycle() to trust the recorded role only while its
resolved model still equals the active model, falling back to matching by
model after switches through alt+m, /model, or retry fallback.
- Added mutation-verified regression tests for branch ordering and the
stale-role fallback.
- Consolidated `browserOpenSchema`, `browserCloseSchema`, and `browserRunSchema` into a single `browserSchema`.
- Simplified the `action` type definition to accept `'open' | 'close' | 'run'`.
- Updated schema validation tests to reflect the unified schema definition.
Dropped rebuilt OpenAI Responses assistant item IDs when a replayed turn lacks its matching reasoning item while preserving text and call_id pairing. Added regression coverage for message, function_call, and custom_tool_call replay.\n\nFixes #4173
Fixes issue #2115: ConPTY hosts truncated a >1MB single full paint at a logical-line boundary, corrupting large-session resume. Bounds the bytes #emitFullPaint emits (retain tail) while keeping full logical state for render planning/commit; guarded to ConPTY hosts only. Includes the issue-2115 repro test. The PR head's original pre-planning truncation was obsolete against the refactored render pipeline, so the evaluator-approved approach (fix 44b88a2f5) was ported onto current #emitFullPaint.
Ports only the thinking double-format fix: resolveThinkingDisplay reuses block.thinking when rawThinking is set (buildDisplayMessage already formatted it), plus a single-entry memo in formatThinkingForDisplay and a rawThinking regression test. The PR's incremental reveal slicing is superseded by the already-merged #3848 (memoized grapheme slicing).
Cherry-pick of the reserve-budget clamp only (resolveBudgetReserveTokens + no-op compaction guard): applies compaction.ts + agent-session.ts + compaction/shake/progress-guard tests. Excludes unrelated Julia prelude timeout and ai/test churn from the PR head.
/quit and /exit hung for many seconds because AgentSession.dispose()
awaited MnemopiSessionState.dispose() unconditionally, and that path
runs consolidate() (state.ts:421) which fires a fresh LLM fact
extraction for the just-retained transcript and then awaits
flushExtractions() per owned bank. One LLM round-trip per shutdown,
no upper bound, no visible status.
- Add a timeoutMs option to MnemopiSessionState.dispose. When the cap
fires the in-flight consolidate is detached to the background and the
SQLite handles close once it settles, so writes never race a closed
handle.
- AgentSession.dispose passes SHUTDOWN_CONSOLIDATE_BUDGET_MS = 1_500 on
the user-visible shutdown path. Per-turn maybeRetainOnAgentEnd has
already retained earlier turns, so the worst case is losing episodic
promotion for the last few turns. State-replacement disposes
(mnemopiBackend.start) stay unbounded.
- InteractiveMode.shutdown surfaces a 'Closing session…' status before
dispose runs so the brief pause is explained rather than mysterious.
Two regression tests in memory-tools.test.ts cover (1) dispose returns
within the budget when flushExtractions stalls and the deferred close
still runs once consolidate settles, and (2) unbounded dispose still
runs the full #2320 consolidate-then-close pipeline.
Fixes#3641
metaphorics