Repository navigation
Pilot: session stall taxonomy — observable signatures of stalled vs healthy long sessions #1192
Description
Pilot — what does a stalled session look like?
The babysitter's core judgment is "progressing vs stalled." This pilot
defines that judgment from evidence instead of vibes.
Gather
- From existing fleet session history (OpenCode/OpenChamber sessions,
Jules sessions, today's queue-worker runs): find sessions that
stalled, died silently, or spun without progress. - For each, record what the observable signals were BEFORE the stall
was obvious: session status field, last-activity gap growth, repeated
tool calls/errors, message cadence, CI state on the bound PR. - Contrast with healthy long-running sessions (a 30-min build/test loop
is not a stall): which signals separate the two?
Acceptance criteria
- A stall-signature table: signal × (healthy range, stall range,
source field/endpoint). - A proposed detector rule the babysitter can encode (thresholds +
which signals must agree), with its expected false-positive cost
stated (a nudge is cheap; a wrong "stalled" in the report is
noise Drew reads). - At least 5 real stall cases and 3 healthy-long cases examined;
fewer only if the history genuinely lacks them (say so).
Part of the OpenChamber dispatch + babysitter design (2026-10-26).
Filed by Muse (operator) 2026-10-26.
Activity
developing-today commentedon Oct 27, 2026
CODE-328
dezren39 commentedon Oct 27, 2026
Pilot complete — stall taxonomy
Document:
files/stall-taxonomy-pilot-2026-10-26.mdin the keep-the-four-repos-healthy goal workspace (committed in hatch-personal), built from a census of all 187 claim episodes inagent.events(2026-10-25 → 27) plus the monitor's tracker/sweep record and GitHub state for the non-queue cases.Headline finding: silence separates the populations almost cleanly. Healthy episodes end within minutes of their last checkpoint (completed p90 = 4.2 min; parked p90 = 2.6 min); evaporated episodes sit silent until the lease dies (p50 = 39.2 min). Healthy long tasks park (1058: 3 parks/0 expiries over 17.8h; 1109: 7/0); stalled tasks expire (1149: 0/5; 1157: 0/4).
Detector rule in brief (full spec + thresholds in the doc): a claimed task is STALLED-SUSPECT → nudge once when (a) ≥20 min since its last event, (b) its lease is within 15 min of expiry (no heartbeat extension — heartbeats write no event rows, so event-silence alone overstates inactivity), and (c) no live CI run on the bound PR. Suspect at two consecutive rounds, or reaching lease expiry, = STALLED-CONFIRMED for the report. Two side classes are detected from the artifact, not the session: DONE-UNREPORTED (bound PR merged / issue closed while the task is still open → harvest, never nudge) and BLOCKED-ON-HUMAN (provider status
AWAITING_USER_FEEDBACK→ nudge the human). Declared parks and named blocks are never stalls. Expected false-positive cost from the record: ~1% of episodes nudged (one 23-min completed outlier), near-zero report-level FPs under the two-signal + two-round rule.Cases examined: 6 stall cases (1149 ×5 expiries, 1157 incl. a zero-checkpoint episode, 1160, 1216 ep.1 silent-but-working, OC #405 GitHub-side holder silence, OC #477 done-but-unharvested, plus Jules status-enum stalls) and 4 healthy-long cases (1058, 1121, 1109, 1216 eps.2–5) — acceptance counts met without needing the "history lacks them" escape.
One open dependency for #1185: pilot #1188 must confirm the OpenChamber equivalent of lease freshness (a pollable last-activity field); the doc's caveat section carries it.
dezren39 commentedon Oct 31, 2026
Triage verdict — ALREADY DONE, closing (queue task 1281, Lane D)
Verified fresh on 2026-10-31:
files/stall-taxonomy-pilot-2026-10-26.mdin the keep-the-four-repos-healthy goal workspace (14,291 bytes, committed in hatch-personal) — census of all 187 queue claim episodes (2026-10-25 → 27), the silence-separates-populations table (completed p90 silence 4.2 min vs evaporated p50 39.2 min), the stall-signature table, and the encodable detector rule (≥20 min event silence + lease within 15 min of expiry + no live CI on the bound PR → STALLED-SUSPECT/nudge; two consecutive rounds or lease expiry → STALLED-CONFIRMED; DONE-UNREPORTED and BLOCKED-ON-HUMAN side classes; declared parks/blocks never stalls; ~1% expected episode-level FP cost).Closing as completed — the work was done on 2026-10-26 and the issue simply never got its closing verb (a DONE-UNREPORTED case, fittingly, in the taxonomy's own terms).
🤖 Agent note · done