Clarify that preflight compression is intentionally excluded because
the later authoritative 'Compacting context' marker from
conversation_compression is the real signal that compression
proceeded. The preflight status can fire even when compression
exits before compaction (e.g. durable guard refresh, Codex-Hermes
mode with no active thread).
PR #6184 narrowed the agent-status-to-SSE bridge to real compression
start notices, but the 'preflight compression:' pattern matches the
Hermes agent's preflight check message ("Preflight compression: ~X
tokens >= Y threshold"). This fires on every new session where the
system prompt alone exceeds the threshold, showing a "Compressing
context" card before any model output — a false positive.
Remove 'preflight compression:' from the pattern. The preflight check
is a one-time pre-model-call estimate, not a mid-conversation
compression event that the user needs to see.
The mid-stream reconnect path could render the just-sent user message below
the live streaming assistant response it triggered, because
`refreshSession()` appended `pending_user_message` at the end of
`S.messages` unconditionally while `loadSession()` correctly inserted
it before the first live assistant message. When the client went through
its reconnect/recovery path mid-turn (typical on iOS PWA screen
lock/resume) the live assistant tail rendered above the prompt.
Centralize every recovery/reconnect reinsertion through one
identity-aware helper (`_mergePendingSessionMessage`) that:
- inserts the owning user turn before its live assistant/Worklog tail,
- appends only when no live tail exists,
- deduplicates against the current tail user row,
- remains idempotent across repeated resume probes.
`refreshSession()` now calls the same helper instead of pushing at
the end. `loadSession()` continues to use the helper unchanged.
Adds `tests/test_issue6419_pending_user_reconnect_order.py` pinning:
- the helper is defined exactly once at module top level,
- it inserts before live assistant and falls back to append,
- `refreshSession()` uses the helper and does not push unconditionally,
- `loadSession()` still merges before renderMessages in the INFLIGHT
reattach branch.
Renames the inner end-marker in
`tests/test_gateway_sse_reconnect_dedupe.py` to a stable inside-loadSession
boundary (`// Phase 2a:`) since the helper is no longer nested.
Gate review found that the broadened token-prefix check caused
deepseek-r1-distill-qwen2.5-bakeneko-32b to enter the Qwen branch
(qwen2 token), return False, and prevent the DeepSeek detector from
running — a silent regression.
Fix: split into two phases:
1. Positive-only scan: any token matching qwen(\d+) with major >= 3
immediately returns True. No False path — Qwen 2.x in hybrid IDs
falls through to subsequent detectors.
2. Terminal guard (original behavior): standalone "qwen" token or
normalized.startswith("qwen") still applies the 3+ gate with
return False for 2.x, since those ARE pure Qwen models.
Add regression test for deepseek-r1-distill-qwen2.5-bakeneko-32b.
- Drop the now-subsumed `normalized.startswith("qwen")` guard
(token-prefix matching already covers the bare case)
- Add test_qwen_prefixed_alias_reasoning_detection covering:
prefixed Qwen 3+ (al-/sn-), bare Qwen 3+, and Qwen 2.x exclusion
The R5b static-pattern test pinned the old inline ('stream_end','error','cancel')
tuple; #6527 replaced it with the named SSE_RELAY_CLOSE_EVENTS frozenset. Accept
either shape and assert behavior (cancel+apperror in the resolved close set)
rather than source syntax. Intent-preserving; non-vacuous (fails on master:
SSE_RELAY_CLOSE_EVENTS absent + apperror not in old tuple).
Adversarial review of the previous commit found a leak it introduced.
`attach_terminal()` subscribes and only then are the response headers written.
A client that dropped in that instant makes `end_headers()` raise
BrokenPipeError/ConnectionResetError, and because the `try/finally` started
*after* those writes, the exception escaped `_handle_terminal_output` with the
queue still in `_subscribers`. That pins `unwatched_since` at None — only
`unsubscribe()` on an empty subscriber list resets it — so the terminal became
permanently unreapable: the exact fd/thread leak (#4633) this reaper exists to
prevent, and strictly worse than the TOCTOU the previous commit fixed, since
that at least still allowed cleanup.
The try now starts immediately after the attach, so every exit path
unsubscribes. Regression test drives a handler whose `end_headers()` raises and
asserts the subscriber list is empty, `unwatched_since` is restored, and the
terminal is reapable again once the grace elapses. It fails against the
previous ordering.
`_terminals_to_reap()` snapshots victims under `_LOCK` and then releases it,
but `subscribe()` clears `unwatched_since` and appends to `_subscribers` under
the terminal's own `_sub_lock`, which the selection pass never held. A viewer
reconnecting in that gap left object identity — all `close_terminal(expected=…)`
ever checked — still true, so the reaper killed a terminal that now had a live
viewer.
- `_claim_reap_victim()` re-establishes the *whole* selection predicate while
holding `_LOCK` and the terminal's `_sub_lock`, and removes the registry
entry in that same critical section: same object, no subscribers, and
`unwatched_since` still set and still expired. A dead process is still
reaped unconditionally — it cannot come back, and a viewer watching a corpse
must not pin the entry forever.
- `attach_terminal()` does the registry lookup and the subscribe atomically
under `_LOCK`, replacing `get_terminal()` + `term.subscribe()` in the SSE
route. It takes the same two locks in the same order as the claim, so
exactly one side wins: the viewer attaches (and the reap is refused), or the
entry is already gone (and the route answers "terminal not running" before
committing a 200, instead of streaming from a corpse). Registration is the
authority — a registered-but-`closed` terminal still attaches so the viewer
receives `terminal_closed`.
- Teardown moved to `_teardown_terminal()` and runs after both locks are
released; `killpg`+`wait` can block ~2.5s and must not pin the registry.
- Dropped the duplicated `self.last_activity = time.time()` the rebase left in
`put_output()`.
Tests: a barrier regression that pauses the reaper after selection, attaches a
viewer through the production path, resumes, and asserts the terminal stays
registered/live and still receives output; the inverse where the reap claims
first and the attach reports not-running; refreshed-grace and replaced-entry
claims; dead-with-viewer; and a probe that `_LOCK` is free during teardown.
Three of them fail against an identity-only claim.
The fd-leak fix retires a terminal whose shell EXITS, and _MAX_TERMINALS bounds
the worst case, but an interactive shell abandoned by its client (tab closed
without POSTing /api/terminal/close, browser crash, network drop) keeps running
forever — there is deliberately no PDEATHSIG. A background reaper now closes any
terminal that has had zero attached output streams for longer than a 15-minute
idle grace, plus any dead-process terminal as a belt-and-suspenders for the
reader-loop retire.
Attachment is tracked on the subscriber model from the output-broadcast change:
subscribe() clears unwatched_since, the last unsubscribe() sets it. A terminal
is born unwatched, so a spawn nobody ever attaches to is reaped too; the grace
spans a tab refresh / brief drop so a real reconnect keeps the session. Eviction
uses close_terminal(sid, expected=term) so a restart replacement is never torn
down. The reaper thread is a lazily-started daemon singleton like the spawn
supervisor.
Stacked on the terminal fd-leak + output-broadcast changes (same file).
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Pin the exact parser finalization count in the interim-assistant-to-tool regression.
Assisted-by: Hermes Agent:gpt-5.6-sol
Assisted-by: Codex:gpt-5.6-luna
Treat status-less interim assistant events as complete semantic boundaries so streaming Markdown flushes its final lookahead character before tool activity. Add behavior coverage for the real interim-to-tool Anchor pipeline.
Assisted-by: Hermes Agent:gpt-5.6-sol
Assisted-by: Codex:gpt-5.3-codex-spark
