[Juliusolsson05]

Closes #676.

Problem

ClaudeQueueState.decisions was append-only for the life of a Claude session. Every time a prompt left the queue (delivered, consumed, popped, stale), a QueueDecision was appended with a content preview and evidence. Each append copied the whole array ([...state.decisions, …], at 7 sites), so a long session paid growing memory and O(n) per departure.

Nothing reads the old decisions:

• the per-session state lives in useIpcSubscriptions' module-level map, not in SessionRuntime, so debug bundles don't carry it;
• QueueStrip renders pending and stale, not decisions.

Change

• A bounded log instead of an append-only one.
  • While the queue is empty, a session keeps its QUEUE_DECISION_WINDOW = 500 most recent decisions. The corpus already has a 175-decision session, so 200 was too thin.
  • While any queue item is pending, the log may grow to QUEUE_DECISION_CEILING = 2 000. Evidence about a row that is still stranded then survives ordinary churn (review c).
  • The ceiling keeps a forever-stranded row from bringing back the unbounded growth this issue fixes.
  • The opening of a pending episode survives the ceiling (review round 2, c). An episode runs from an item entering an empty queue to the queue emptying again. Its first decisions move into episodeHead (up to 200) instead of being evicted. The head shares the ceiling, and is released and counted when the episode closes.
• droppedDecisions counts evicted decision records, so kept + evicted is every decision ever recorded.
  • That is not the number of items (nextSeq): a pending item has no decision yet, and a stale-unattributed mark is an extra decision (reviews a, b, c).
  • Only a session that drained every item exactly once and never went stale has the two equal, and the tests assert the equality only there.
• One helper, withDecisions(state, added, pendingAfter), owns appending and trimming, and replaces the 7 hand-written spreads. Each append copies at most the bound.
• markStaleWhenIdle's "has anything departed" guard counts evicted records too.
• Plan: docs/plans/2026-09-26-claude-queue-decision-window.md (with a review section).

Tests

src/renderer/src/session-runtime/claudeQueue/reconcile.test.ts:

• Fail-first: 1 000 queue/consume cycles keep the newest 500 and count 500 dropped. On origin/main the session keeps all 1 000.
• A stranded episode: the recorded divergence-stranded-background-commands fixture plus 600 churn cycles still holds every original decision while its rows are pending.
• A forever-stranded row: the log is still bounded at the ceiling.
• Boundaries: a single 600-item batch trims to the window, and a popAll departure is bounded like any other.
• Episode opening: after 2 200 unrelated departures, the fixture's episode opening is still in episodeHead. Stale marks at the ceiling are bounded, and by the pending queue.
• Mutations killed:
  • a batch counted as one (a);
  • popAll bypassing the bound (b);
  • no pending ceiling (c);
  • no episode head, the head outside the ceiling, and stale marks unbounded or bounded by the empty window (round 2, c).
• The corpus replay's conservation assertion is scoped to what that corpus is (drained, never stale).
• Stated residual: an episode's middle (after its first 200 decisions, before its most recent ~1 800) can still be evicted.

npx tsc -b prints 0 lines, and the branch is merged with current main.

🤖 Generated with Claude Code

[Juliusolsson05]

Review disposition (round 1 at 3079ef8c, round 2 at 5313a7b8; all codex)

#	Finding	Verdict	Disposition
a1 / b1 / c2	decisions.length + droppedDecisions === nextSeq is false for valid states (pending items, stale marks)	valid	Fixed: droppedDecisions is restated as evicted decision records. The equality is asserted only for drained, never-stale sessions, and the test says why.
c1	A recent-only window evicted every decision around rows still stranded in the queue	valid	Fixed: while any item is pending, the log may grow to a 2 000 ceiling; with the queue empty, it trims to the window, raised to 500. A fixture-replay test and a forever-stranded bound test pin it; the no-ceiling mutation is red. Residual: evidence older than 2 000 later decisions.
a2	A batch trim is untested (+1 mutant survived)	valid	Fixed: a single 600-item batch test; the mutant is red.
b2	The popAll path could bypass the bound	valid	Fixed: a popAll boundary test; the bypass mutant is red.

Also fixed: markStaleWhenIdle's "anything departed" guard now counts evicted records. Verification pass: see Round 2 below.

Round 2 (verification at 5313a7b8: a and b MERGE-READY, c FIX-BEFORE-MERGE):

• c major — the ceiling only postponed the loss (after 2 200 departures, 0 of the 164 episode decisions were kept). Valid. Fixed in 94d4d5ca: while items are pending, the episode's first 200 decisions are kept in episodeHead within the ceiling.
• c's two surviving idle-path mutations: pinned.

Round 2 is the cap: any further finding is the manager's call.