Skip to content

Watchdog: one visible replace per session per tick, verified next tick - #13

Open
bgmcmullen wants to merge 1 commit into
mainfrom
watchdog-one-replace-per-tick
Open

bgmcmullen wants to merge 1 commit into
mainfrom
watchdog-one-replace-per-tick

Conversation

@bgmcmullen

Copy link
Copy Markdown

Summary

Implements HypAware recommendation hyprec-023eb3ce7a3041f7, "Keep a watchdog respawn from stopping the whole fleet" (usage review, 2026-09-21 to 2026-10-11, recommendation 3).

What the recordings showed

Three fleet-wide second failure within 60 minutes holds in twelve days, each waiting on an operator to rearm: about 60 h from Sep 26, about 20 h from Oct 3, about 16 h from Oct 7. In the first two, the second failure was the watchdog's own safety replace, duplicated or issued with its output hidden behind | tail -20. The third was two gateway failures with no replace in the window.

Changes

  1. Watchdog skill, respawn rung (.claude/skills/neutral-watchdog/SKILL.md): at most one neutral safety replace --session <name>; echo "exit=$?" per session per tick, unpiped; read held from the JSON and stop if true; never re-issue in the same tick; verify recovery from a new transcript event on the next tick; log exit code and lastHold.
  2. LLP 0083 (Draft RFC): what src/safety.js and docker/safety-controller.js actually count, the three holds replayed against that rule, and the finding that the controller behaved as LLP 0072 specifies. The replace path records a failure for the replaced session whether or not it is alive, which is the accepted rule. Whether a verified replace should keep counting, or gateway failures should count per component, is left as a decision for the fleet owner. No rule or code changes.
  3. Extended-by: forward reference on LLP 0072.

Context

Neutral was stopped on 2026-10-12 in favor of HypForge. This records the recommendation's deliverables against the codebase the evidence came from.

Checks

🤖 Generated with Claude Code

Implements HypAware recommendation hyprec-023eb3ce7a3041f7, "Keep a watchdog
respawn from stopping the whole fleet."

- Amend the respawn rung of the neutral-watchdog skill: issue at most one
  `neutral safety replace` per session per tick as its own unpiped call with
  the exit code echoed, read `held` from the controller's JSON, never re-issue
  in the same tick, verify recovery on the next tick, and log exit code and
  lastHold.
- Add LLP 0083 (Draft RFC) recording what the safety controller counts, the
  three Sep 26 to Oct 7 holds replayed against that rule, and the accounting
  decision left to the fleet owner. No rule changes.
- Forward-reference LLP 0083 from LLP 0072.

Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

1 participant