PID audit — 2026-04-18

Executive verdict

PID is PARTIALLY working. Breaker + fixer + precheck stages are wired, coded, and have recent eval rows. The loop-closure stage (re-test / A4) is spec-only: regen.sh exists as a script but has zero callers, is not cron-wired, and state/log/regen-deltas.ndjson does not exist. On top of that, the breaker and fixer crontab entries are currently PAUSED (commented # PAUSED 2026-04-18: lines in the live crontab), so even the stages that work are not running autonomously right now.

Stages present vs. planned

The plan (plans/read-the-plan-nifty-wolf.md) names six A-primitives for the PID loop. Here is the receipts column next to each:

Supporting infra:

• state/bin/pid-detect.ts — wired, called from auto-regen.sh.
• state/bin/pid-drain.ts — wired, called from auto-regen.sh.
• state/lint/pid-trends.ts — wired, live output shows 15 trends detected

this run (recurring-area=5, friction-spike=7, eval-decline=3).

Observed per-verb eval means (autopilot skill)

From jq -s 'group_by(.verb)' <(jq -c 'select(.skill=="autopilot")' state/log/evals.ndjson):

[
  { "verb": "breaker-tick",  "n": 34, "mean": 1.000 },
  { "verb": "fixer-tick",    "n": 28, "mean": 0.964 },
  { "verb": "dispatch-tick", "n":  2, "mean": 0.000 },
  { "verb": "tick",          "n":  1, "mean": 1.000 }
]

Plus snappy-fix.verb=pick rows (not the autopilot cron — operator-invoked): recent mean is 1.0 across 15+ runs; rows repeatedly cite skill-silent-meeting-triage, stale-pull-blocked, gateway-drift-batch — which matches the top recurring-area signals in pid-trends.ndjson:

These are the areas where the fixer is closing rows that keep re-opening — the exact symptom A4 is supposed to catch. A4 is not wired.

Cron reality check

crontab -l | grep autopilot:

# PAUSED 2026-04-18: 0,30 * * * * /Users/robertboulos/projects/snappy-os/state/bin/autopilot/break.sh >> /Users/robertboulos/.claude/logs/snappy-os-breaker.log 2>&1
# PAUSED 2026-04-18: 15,45 * * * * /Users/robertboulos/projects/snappy-os/state/bin/autopilot/fix.sh   >> /Users/robertboulos/.claude/logs/snappy-os-fixer.log 2>&1

Both autopilot crons are commented out as of today. The 34 breaker-tick and 28 fixer-tick eval rows are from before the pause (latest fixer-tick: 2026-04-18T05:16:54Z; latest breaker-tick: 2026-04-18T06:09:52Z). Nothing about autopilot has run autonomously since ~06:09 UTC today.

No regen.sh crontab entry has ever existed (grep regen in crontab: 0 hits).

Documentation score

End-to-end docs live in four places:

1. state/wiki/pid-loop.md — quorum / cross-machine PID (phase 7).
2. state/wiki/autopilot.md — goal-file loop (the older single-cron model).
3. state/skills/snappy-autopilot-twocron.md — breaker/fixer split (the

current design).

1. state/skills/pid-regen.md — detect/drain/dispatch brief flow.

Grade against Robert's 5 criteria:

Options surface — cli.js autopilot

From node bin/cli.js autopilot:

cli.js autopilot [--status|--break|--fix|--tick]  twocron detection/remediation wrapper

Hidden levers (not in --help, but real):

• SNAPPY_REGEN_MIN_INTERVAL env (auto-regen.sh:42) — rate-limit seconds,

default 900 (15 min).

• state/engaged.json — engage/disengage autopilot / pid-loop recipes.
• pid-detect.ts --stats / pid-detect.ts --dry (per pid-regen.md:80).

Gaps in the options surface:

• No cli.js autopilot --regen <skill> passthrough for regen.sh, even

though the script exists and takes a single arg.

• No way to see --status for regen deltas (because deltas file doesn't

exist).

• No --help page describes the engagement gate or SNAPPY_REGEN_MIN_INTERVAL.

Gaps (concrete list)

1. A4 re-test gate is orphan. state/bin/autopilot/regen.sh is coded

but called by nothing. Neither bin/cli.js, state/bin/auto-regen.sh, state/bin/autopilot/fix.sh, nor crontab -l references it. Consequence: the loop cannot measure whether a rewrite improved a skill.

1. state/log/regen-deltas.ndjson does not exist. Direct consequence

of #1. The PID-loop-closure artifact is unobservable.

1. Autopilot crons are paused. Breaker + fixer entries in crontab -l

are commented with # PAUSED 2026-04-18:. Nothing will run autonomously until those lines are uncommented.

1. No canonical end-to-end doc. Four pages each cover one slice;

no page shows breaker → fixer → detect → drain → dispatch → regen → deltas as one diagram.

1. Thresholds are hardcoded, not tunable via docs. pid-detect.ts

fails >= 3, mean <= 0.75, zeros >= 2 are prose-documented in pid-regen.md but there is no "how to change these" page.

1. pid-trends.ndjson is undocumented. 230 rows on disk, 15 fresh this

run, consumed by fix.sh A4a — but no wiki page explains the schema.

Honesty section

• A1 (quota guard): wired and tested — see fixer-quota grep pattern.
• A2 (already-fixed precheck): code is wired, **but zero fixer-precheck

eval rows have ever been written** (evals.ndjson contains 0). Either the precheck condition never fired before the crons were paused, or the precheck fires silently. Worth a probe when crons resume.

• A3 (self-rewrite loader): wired. No direct receipts (would require

a git-log search for fixer: self-write pitfall commits).

• A4a (recurring detect): wired — IS_RECURRING_CODE logic present.
• A4b/c (autofix branch): wired — branch-push logic present.
• A4 proper (re-test gate): SPEC-ONLY. regen.sh exists, is

well-commented, has zero callers. regen-deltas.ndjson does not exist on disk.

One-line verdict: PID loop is three-quarters wired. The part that closes the loop (measure whether the rewrite helped) is spec-only, and the part that runs autonomously (cron) is paused.