Register the documented 65536-token default output limit and retain the existing 128000-token context window for glm-5.3-flash. Other model registrations are unchanged.
Pi's isRecoverableLength (pi-ai dist/utils/overflow.js:163), used by agent-session.js:1659, now treats a length stop below 65536 output tokens as recoverable, including context-clamped requests or a lower provider cap. Pi then compacts and retries. It allows one recovery in a row: the flag resets on a new user message or a successful assistant message that is not a length stop, so tool rounds can re-arm it within one prompt.
The recovery happens inside one runner attempt. The turn marker reflects the final agent_end, and the failed and retried messages still count in that message's usage delta, so three runner attempts need not mean three provider calls.
The 128000 context and 16384 reserve are unchanged, so threshold compaction still starts above 111616 tokens. Output clamping now starts above about 58368 estimated input tokens instead of 107520. No new local refusal.
This commit can be dropped without affecting Part A.
Review round fixes for the length continuation in Part A.
Before each continuation the runner now reads processing.status as well as
state. A halted, invalid or unreadable processing state starts no further
attempt; the message is left unread with no reply and no failure record, as
for a halt during any turn.
A failed state or processing read keeps the empty length classification, so
the sender gets no RESULT. The chief notice and the trail event name how many
attempts hit the limit.
usage.end for the message id now runs in a finally block, so it closes once
even when an attempt throws.
The library citation retry no longer starts continuations. They would run
outside the message's usage bracket, and the retry's own bracket is refused
(#1853). An empty length retry fails visibly instead, with no fixed RESULT.
Tests cover a halt between attempts, a failed continuation check on the plain
and library paths, a throwing attempt, a library continuation that succeeds,
and the retry path.
Pi could spend its output limit on thinking and the runner relayed an empty answer as a result. Preserve final stop metadata and allow two bounded continuations, accounting for the whole message in one transcript usage delta.
Record exhaustion on the message trail and notify the chief without an automatic reply cycle. Mount the shared notice prefix in the sandbox.
Merges ui-dev-projects (c907e4d0d) onto the Settings slice (65df511f3).
Conflicts resolved by keeping both sides: the slice 3 token block in
tokens.css, and in bus.js the Settings display redirect plus the
Projects route test. Whole webui suite on Node 26: 729 passed, 0 failed.
Review round 1 on 4048e0f60 found two defects.
Startup recovery ran unconditionally. A continuation retry now refuses
while processing is halted, so a broker restart with a waiting
continuation failed with processing-halted. Startup recovery now waits
while processing is halted or invalid, as the periodic pass does;
processing.resume schedules it again.
#continuationStarting matched any launch without an end, so an older
launch whose end was lost (#1851) hid every later run's end. Only the
target's latest lifecycle event counts now.
Both regression tests fail on 4048e0f60 and pass here.
Expose the parent message class on agent delivery so the runner can stop automatic replies to answers while preserving requests and ordinary relays. Scope continuation retry versions to the target, suppress notices during processing halts and unclaimed launches, and clarify the instance-running refusal.
The bus suite excludes project-files.test.mjs because its 1 MiB socket test timed out on both the candidate and the untouched 1186cf7b1 base.
The chat.js fetch wrapper recorded a request-not-found from /api/requests/<id>/workspace as a missing request and then served a synthetic 404 to the request panel's rereads, failing 11 request-browser tests. Only the request's own read records a miss now. Adds a regression test that fails on 9d3eb9e11, and bounds the r3 held-404 test's check on its refresh reads, which raced Files' cached repaint under load.
A mosaic-request-state for an id now advances that id's own generation on every thread retaining it, so a 404 still in flight from before the observation cannot record the miss afterwards. Generations are pruned with the misses to retained rows. Adds a held-404 regression that fails on 16384bb12.
A same-thread refresh kept the last resolved live controls until the session and history reads finished. When the fresh project lookup finds the project archived or the archive state unknown, withdraw the unguarded thread controls and remount the request panel fenced before those reads.
A signal name with no number in os.constants.signals now exits 1 instead of a made-up 128. Known signals keep 128+n (SIGKILL 137, SIGTERM 143) and the content-free stderr receipt. Test covers all three through the real entrypoint.
jason.woltje