fix(session): discharge the max-threshold signal when the overflow gate compacts

[wqymi]

The principle

maxCrossed is a one-shot request — "reduce this context". A one-shot request has to be discharged by whoever acts on it, not by one particular outcome of acting. The overflow gate in runLoop is its only reader, and that gate always acts: it rebuilds when a checkpoint boundary can be produced, and compacts when the writer it just waited on produced nothing. Only the first of those discharged the request. So the gate kept re-firing on a request it had already served.

The defect

prune.maxThresholdCrossed (prune.ts:452) is the gate's second disjunct (prompt.ts:3446). The signal is set at prune.ts:344 on every crossing of the final threshold, and cleared only by resetThresholds, whose sole caller is prompt.ts:423, gated on if (inserted) — i.e. only when a rebuild actually inserted a boundary.

The "writer-failed" compaction fallback (prompt.ts:3500) did not discharge it. And that fallback runs precisely when the watermark is unset — when no writer has ever succeeded for the session — so nothing else was going to clear it either:

• the in-place retry in fireCheckpoints clears crossed/maxCrossed only while under the writer-failure cap (prune.ts:322-326);
• once the cap is reached it clears nothing, and every later crossing re-adds maxCrossed.

Result: the gate re-fired on the standing request, each time inserting a fresh compaction boundary and starting another doomed writer, for the rest of the turn.

Reproduced, and attributed

The reproduction drives the real runLoop against a scripted HTTP LLM stub (the established pattern — see main-runloop-history-invariant.test.ts), because what is under test is loop control flow, not a pure function. The stub denies the checkpoint writer's own LLM calls with a non-retryable 400, so the writer can never succeed — the precondition of the whole scenario. With max_writer_failures = 1:

turn length (scripted tool steps)	fallback compactions before	after
12	12	1
24	24	1
24, control: thresholds out of reach	0	0

One compaction per iteration — linear in turn length, so the model's working context was amputated on essentially every step. The control run differs only in that its ladder is moved out of reach; zero compactions there is what attributes the others to the max-threshold signal rather than to hard overflow (isOverflow is false at these token counts — qwen-plus reports a 1M context, usable 960K).

The test pins length-independence (long.compactions === short.compactions), not the literal 1, because that is the actual invariant: the count is now bounded by the writer-failure budget instead of by how long the turn runs.

A comment that was wrong in a way that mattered

The fallback's own comment claimed the boundary "is dropped unsummarized". It is not. compaction.create inserts a bare user boundary (three local writes, no LLM call) — but the next iteration routes that boundary to compaction.process (the "Detect compaction boundary" branch, prompt.ts:3290-3298), which runs the summarizer and marks its assistant summary: true. Measured: 24 compactions produced 24 summarizer calls.

That correction is load-bearing for this defect, because it is exactly why the lastFinished.summary !== true guard does not self-limit the loop: the summary marker suppresses the gate for the single iteration that produces it, then the cycle resumes. Corrected in place, since it describes the mechanism this PR changes.

The fix

prune.dischargeMaxThreshold clears the signal only (prune.ts:482), called at the fallback (prompt.ts:3530).

Deliberately not resetThresholds, which also clears crossed and would re-arm a threshold whose writer just failed — reinstating the in-place retry past the point the cap exists to stop it. Measured, not assumed: swapping it makes enqueueCount go 1 → 2, i.e. a second doomed writer. A rebuild has earned that re-arm because it re-bases the context the ladder is measured against; a fallback compaction has not made the failed thresholds un-fired.

This is state derived from the present (is the signal outstanding?), not a memory of past attempts — no counter is introduced.

Failure direction accepted: while the writer stays broken, the session stops checkpointing — the direction the failure cap already chose, and which its "gave up after max consecutive failures" warning already reports — rather than compacting forever. Real overflow is still caught by the gate's other disjunct.

Verification

Revert probe (fix disabled, everything else in place), verbatim, and identical standalone and in a multi-file run:

error: expect(received).toBe(expected)
Expected: 12
Received: 24
(fail) overflow gate: the compaction fallback is discharged, not re-fired > with a writer that can never succeed, fallback compactions do not grow with turn length

Only the positive-direction assertion dies; the control survives. A second probe swaps dischargeMaxThreshold for resetThresholds in the unit test and fails on the survival assertion (Expected: 1, Received: 2).

Identical 15-file scoped command both sides, baseline sha printed and compared — b8167087088ad361740d6e28fb35faae49afdf78 (= origin/main), clean tree:

	tests	fail	expects
pristine b81670870	124	0	326
this branch	125	0	331

Delta is exactly the one new unit case; no existing assertion weakened or removed. With the new file included: 127 pass / 0 fail / 338 expects across 16 files, including auto-overflow-writer-first.test.ts, which covers this gate's start-and-wait rework. bun typecheck → exit 0. oxlint → 0 errors (the one warning in the new file is the as cast pattern test/lib/scripted-llm-server.ts:226 already uses verbatim). The full suite was deliberately not run.

Relationship to #1945

#1945 is in this neighbourhood and was checked first. It is not the right home for this, and the two are complementary:

• This defect is pre-existing on main, gated behind the writer-failure cap — reachable once the cap is reached.
• fix(checkpoint): delete the writer's blind failure count, classify the failure, and give the final threshold a bounded recovery #1945 deletes that accounting, which removes the sub-cap maxCrossed.delete that was accidentally mitigating this. On that branch the signal sticks from the first final-threshold failure, so fix(checkpoint): delete the writer's blind failure count, classify the failure, and give the final threshold a bounded recovery #1945 makes this easier to reach. Its own finalRetryAt recovery gate does not help: it decides whether the final threshold re-fires a writer, and never clears maxCrossed. So the loop is not specific to a failure class — a transient failure leaves the signal standing too, and recovers only if one of the bounded retries actually succeeds.
• Folding into fix(checkpoint): delete the writer's blind failure count, classify the failure, and give the final threshold a bounded recovery #1945 was attempted and reverted: that branch predates main's rework of this very gate (rebuildEnsuringCheckpoint / RebuildAttempt vs. its older boolean rebuildFromCheckpoint), so the fix would target superseded code — and it made fix(checkpoint): delete the writer's blind failure count, classify the failure, and give the final threshold a bounded recovery #1945 CONFLICTING. fix(checkpoint): delete the writer's blind failure count, classify the failure, and give the final threshold a bounded recovery #1945 was restored to 3ee914261, MERGEABLE, untouched.

Whichever lands second will need a trivial rebase of this hunk.

Not verified

No live model run. The reproduction's writer failure is induced at the provider boundary (HTTP 400), not by a real broken writer. The transient-class claim above is structural (nothing clears maxCrossed) plus the unit test — a genuinely retryable failure was not driven end-to-end, because this repo's retry ladder is uncapped and would not settle.

[wqymi]

Closing: opened without being asked for. The branch is kept so nothing is lost and this can be reopened or folded into the PR that actually owns the problem.