fix(a2a): orphan sweep honors A2A_REPLY_TIMEOUT

[gaoanze888]

Summary

Fixes #106972. The A2A orphan-task sweep used a hardcoded _ORPHAN_TIMEOUT = 300 while the caller's reply deadline already honored A2A_REPLY_TIMEOUT via _reply_timeout(). A turn taking 350–600s was marked TASK_STATE_FAILED by the sweep before the configured reply deadline arrived, and once terminal the TaskStore.complete guard silently discarded the real (eventally-produced) reply — so raising A2A_REPLY_TIMEOUT had no effect, contradicting the docs that present it as the fix for "Replies time out on long tasks."

Root cause

• plugins/platforms/a2a/adapter.py: _ORPHAN_TIMEOUT = 300 (hardcoded) fed self.tasks.fail_orphans(_ORPHAN_TIMEOUT) in _watchdog_loop. Separately, _reply_timeout() reads os.getenv("A2A_REPLY_TIMEOUT", "300") — but only for the active caller's own wait deadline, never for the orphan sweep.
• plugins/platforms/a2a/protocol.py: fail_orphans marks any task older than timeout_seconds as STATE_FAILED; complete() then has if rec["state"] in TERMINAL_STATES: return None, so a later legitimate reply is silently dropped.

Fix

Extract the sweep into _sweep_orphans() and resolve the window via _reply_timeout() so the sweep window matches the configured reply deadline: a long-running turn survives long enough for its real reply to complete the task (the reply arrives → task goes STATE_COMPLETED → the sweep's not in TERMINAL_STATES guard skips it). _ORPHAN_TIMEOUT is removed; _WATCHDOG_INTERVAL is unchanged.

This is option (1) from @leyhux101-bit's report (make _ORPHAN_TIMEOUT respect A2A_REPLY_TIMEOUT). With it, the docs become accurate as written; no doc change is needed.

Tests

New TestOrphanSweepRespectsReplyTimeout in tests/plugins/test_a2a_plugin.py (4 tests):

• default window (300s) sweeps a 400s-old task → STATE_FAILED + metrics++
• regression: A2A_REPLY_TIMEOUT=600 preserves a 400s-old task (under the new 600s window; would be swept under the old hardcoded 300) — this is the A2A: A2A_REPLY_TIMEOUT does not affect the orphan-task sweep (hardcoded 300s) — late replies get silently discarded #106972 scenario
• terminal (completed) tasks are never swept, regardless of age
• invalid env value falls back to 300 (the existing _reply_timeout except branch)

Mutation-verified: reverting _sweep_orphans to the old orphan_timeout = 300 makes the regression test fail (400s-old task swept at 300 even with A2A_REPLY_TIMEOUT=600); restored, all 4 pass. Full a2a suite: 159 passed (4 new + 155 existing), 18 integration deselected. ruff check clean.

Note

Huge credit to @leyhux101-bit for the precise root-cause writeup (the split between _reply_timeout() and the hardcoded _ORPHAN_TIMEOUT is the whole story). The reporter offered to open this PR — I put it up to unblock the fix, but if you'd rather land your own implementation @leyhux101-bit I'm happy to close this in favor of yours.

The diff also carries unrelated .github/workflows changes — that's a fork/main ↔ origin/main sync artifact (my fork is behind on two workflow files and I lack a workflow-scope token to fully resync); the actual code change is only the two files listed above.

[gaoanze888]

Closing this as a duplicate — thanks to @alt-glitch's triage I see the orphan-sweep / reply-timeout mechanism already has three open PRs in flight:

• fix(a2a): honor configured reply timeout in watchdog #100312 (@rtx9jckz2g2g-prog) — the closest match, and a better design than what I landed on here: it introduces _orphan_timeout() = max(300, _reply_timeout() + _WATCHDOG_INTERVAL), which (a) preserves a five-minute floor so a misconfigured low A2A_REPLY_TIMEOUT can't make the sweep fire almost immediately, and (b) adds one watchdog interval of grace so the requester's own reply deadline is guaranteed to fire before the sweep marks the task failed. My version used int(_reply_timeout()) directly — neither a floor nor grace — so it's strictly less robust on those two edges.
• fix(a2a): reply watchdog honors the route configured timeout #90158 (@johnmwhitman) — per-route timeout + grace framing.
• fix(a2a): do not fail tasks on observation timeout #96983 (@RuniThomsen) — broader (also touches protocol.py).
• fix(a2a): honor A2A_REPLY_TIMEOUT in orphan-task sweep #106977 (@KoNit-K) — also opened today, a few minutes before this one.

Five PRs for a one-symbol fix is just noise for the maintainers, and #100312 is both first and more correct, so I'd rather not split reviewer attention. The four regression tests I added here (default-window sweep, the A2A_REPLY_TIMEOUT=600-preserves-a-400s-task regression, terminal-task skip, invalid-env fallback) are mutation-verified and available if @rtx9jckz2g2g-prog wants to lift them onto #100312 — but no obligation. Credit to @leyhux101-bit for the clean root-cause split between _reply_timeout() and the hardcoded _ORPHAN_TIMEOUT; that writeup is the whole story.