Summary

Active Antigravity (agy) runs can be killed after 120 seconds with:

[jaw:watchdog] killing main — no first progress after 120s

This can happen even while the Antigravity transcript is growing with planner/thinking activity. Telegram then receives 응답 없음, and the stored assistant result may contain Error: timed out waiting for response.

Environment

• cli-jaw: 2.1.5
• agy: 1.0.10
• backend: Gemini 3.5 Flash, reproduced with Low and Medium reasoning
• Linux

Reproduction Pattern

1. Use the agy backend.
2. Start a task that spends more than two minutes planning before its first parsed tool event or final stdout response. A multi-step maintenance task reproduces this reliably; long media-generation workflows are also susceptible.
3. Observe that the Antigravity transcript file is found and tailed.
4. At approximately 120 seconds, cli-jaw kills the process as having no first progress.

Example log sequence:

[jaw:main] Spawning: agy -p ...
[jaw:agy:transcript] tailing .../transcript.jsonl
... transcript/planner activity continues ...
[jaw:watchdog] killing main — no first progress after 120s
[jaw:main] exited code=124, text=0 chars
[tg:out] ...: 응답 없음

Root Cause

agy-transcript-watcher.js correctly notices any transcript growth:

if (delta.offset > previousOffset) {
    options.ctx.agyTranscriptActive = true;
    options.onActivity?.();
}

However, spawn.js currently handles onActivity only by scheduling quiet completion:

onActivity: () => {
    scheduleAgyQuietCompletion();
},

The watchdog is reset only when applyTranscriptTool() parses a tool row:

ctx.stallWatchdog?.markProgress();

Planner/thinking rows grow the transcript but may not parse as tool rows. A run can therefore be visibly active while firstProgressMs remains unsatisfied.

This occurs before Antigravity's own print timeout. cli-jaw passes --print-timeout using the configured absoluteHardCapMs (240 minutes with the current default hard cap), while the watchdog's default firstProgressMs is 120 seconds.

Expected Behavior

Any current-turn Antigravity transcript growth should count as process progress and prevent the no-first-progress watchdog from killing an active run.

Suggested Fix

Mark watchdog progress in the transcript watcher's onActivity callback:

onActivity: () => {
    ctx.stallWatchdog?.markProgress();
    scheduleAgyQuietCompletion();
},

It would also be useful to add a regression test covering:

• planner-only transcript growth for longer than firstProgressMs
• a later parsed tool event
• genuinely silent runs that should still be terminated
• long-running tool calls with sparse output

Notes

Increasing agentTimeout.agy.firstProgressMs is a workable local mitigation, but it only delays the false positive. Treating verified transcript growth as progress addresses the integration mismatch directly while retaining the watchdog for genuinely stalled processes.