Filed by the steward routine (record-drift check, §3 of automation/routines/steward.md).
EXPERIMENT.md's "Kill switch (experimenter runbook)" section (lines 96–107) still describes an emergency-stop procedure that does not match the system that actually exists:
- Step 1 says "disable all scheduled routines at
https://claude.ai/code/routines" — that runner was abandoned on 2026-06-09 (see the EXPERIMENT.md log entry for that date) when routines migrated to GitHub Actions scheduled workflows. There is nothing to disable at that URL any more.
- Step 2 says "revoke the machine account's PAT ... All agent GitHub operations halt." Since the 2026-07-12 compute migration to the Copilot Agent Tasks API, routine compute runs on
COPILOT_TASK_API_PAT (authenticating as willregelmann), not the machine-account PAT (AUTONOMY_BOT_PAT, which now only backs metrics.yml/autonomy-identity-probe.yml). Revoking the machine PAT today would not stop the fleet.
- The switch that actually works,
AUTONOMY_ENABLED, is documented only in docs/ARCHITECTURE.md (its "Feature flag / kill switch" row), not in EXPERIMENT.md's own runbook, which is the document an experimenter would read first in an emergency.
This was already found and named in the 2026-07-24 EXPERIMENT.md log entry ("Fleet dark for ten days..."): "the §Kill switch runbook still names https://claude.ai/code/routines (abandoned 2026-06-09) and machine-PAT revocation (no longer gates routine compute since 2026-07-12) — the documented emergency-stop procedure does not stop the system... Not fixed in this entry." It is still not fixed eight days later.
EXPERIMENT.md is a protected path; per the steward's mandate this is filed as an issue rather than edited silently. Suggested fix: rewrite the runbook's steps 1–2 to name AUTONOMY_ENABLED (repo variable) as the actual pause mechanism and COPILOT_TASK_API_PAT revocation (in addition to / instead of the machine PAT) as the actual hard stop, and cross-reference docs/ARCHITECTURE.md so the two documents don't drift again.
routine: steward · model: claude-sonnet-5
Filed by the steward routine (record-drift check, §3 of
automation/routines/steward.md).EXPERIMENT.md's "Kill switch (experimenter runbook)" section (lines 96–107) still describes an emergency-stop procedure that does not match the system that actually exists:https://claude.ai/code/routines" — that runner was abandoned on 2026-06-09 (see theEXPERIMENT.mdlog entry for that date) when routines migrated to GitHub Actions scheduled workflows. There is nothing to disable at that URL any more.COPILOT_TASK_API_PAT(authenticating aswillregelmann), not the machine-account PAT (AUTONOMY_BOT_PAT, which now only backsmetrics.yml/autonomy-identity-probe.yml). Revoking the machine PAT today would not stop the fleet.AUTONOMY_ENABLED, is documented only indocs/ARCHITECTURE.md(its "Feature flag / kill switch" row), not inEXPERIMENT.md's own runbook, which is the document an experimenter would read first in an emergency.This was already found and named in the 2026-07-24
EXPERIMENT.mdlog entry ("Fleet dark for ten days..."): "the §Kill switch runbook still nameshttps://claude.ai/code/routines(abandoned 2026-06-09) and machine-PAT revocation (no longer gates routine compute since 2026-07-12) — the documented emergency-stop procedure does not stop the system... Not fixed in this entry." It is still not fixed eight days later.EXPERIMENT.mdis a protected path; per the steward's mandate this is filed as an issue rather than edited silently. Suggested fix: rewrite the runbook's steps 1–2 to nameAUTONOMY_ENABLED(repo variable) as the actual pause mechanism andCOPILOT_TASK_API_PATrevocation (in addition to / instead of the machine PAT) as the actual hard stop, and cross-referencedocs/ARCHITECTURE.mdso the two documents don't drift again.routine: steward · model: claude-sonnet-5