Reflect verify

Skill stevengonsalvez/ainb-reflect-memory/.agents/skills/reflect-verify

Verify Reflect end-to-end health, especially drain, ingest/consolidate maintenance, launchd timers, statusline error surfacing, and real Claude/Codex harness smoke checks. Use when Reflect seems stuck, drain is not clearing pending reflections, ingest has gone quiet, statusline ERR needs proof, or a change needs tmux-backed validation.From its SKILL.md

Install

npx -y skills add stevengonsalvez/ainb-reflect-memory --skill reflect-verify

Assembled from the repository path, not quoted from the project. Check it against their README if it does not work.

One thing to look at

2 stars2 stars. Stars are a popularity signal and not a quality one, but at this level it is likely that nobody has read this closely except its author, and you would be relying on your own review.

SKILL.md

3.3 KB, as published. Nobody here has run it

Reflect Verify

Run Reflect validation in layers: static checks, isolated queue/drain smoke, live read-only health, then optional real Claude tmux smoke. Do not run full ingest or consolidate as part of verification unless the user explicitly asks; the watchdog only surfaces due/malfunction states.

Workflow

Check source and install drift.
- Compare repo hook files against installed harness copies and Claude plugin cache.
- Confirm launchd labels for com.reflect.drain, com.reflect.idle, and com.reflect.maintenance when running on macOS.
- Never kill tmux server or broad process names; only use exact session names created by this verification.

Run static gates.

.agents/skills/reflect-verify/scripts/verify-reflect.sh --static

Run isolated drain smoke.
```
.agents/skills/reflect-verify/scripts/verify-reflect.sh --isolated
```
This uses a temp REFLECT_STATE_DIR and a stub Claude binary. It must prove timeout/no-output entries are quarantined and removed from the queue without touching live Reflect state.
Run live read-only health.
```
.agents/skills/reflect-verify/scripts/verify-reflect.sh --live-readonly
```
Report queue depth, drain lock status, latest drain activity, unacked errors, ingest log age, and loaded launchd labels. Do not mutate queue, errors, or KB.

Run maintenance watchdog smoke when statusline surfacing is in scope.

REFLECT_STATE_DIR="$(mktemp -d)" \
  GLOBAL_LEARNINGS_PATH="$(mktemp -d)" \
  REFLECT_WATCH_FORCE_JSON_ERRORS=1 \
  REFLECT_WATCH_SKIP_LAUNCHD=1 \
  plugin/hooks/reflect-maintenance-watch.sh

Verify errors.json receives ingest_never_ran or drain malfunction entries.

Run real Claude tmux smoke only when needed.
```
.agents/skills/reflect-verify/scripts/verify-reflect.sh \
  --real-claude --workdir ~/d/git/ai-coder-rules
```
The script creates a tmux session named reflect-verify-<timestamp>, runs a tiny claude -p probe, captures a log, waits with a timeout, and kills only that exact session if cleanup is needed.

Done Criteria

Static checks pass.
Isolated drain smoke proves bad entries cannot head-of-line block the queue.
Live read-only check explains current Reflect state with exact queue depth, launchd labels, and latest drain/error timestamps.
If real Claude smoke is run, a tmux log proves claude -p produced a result or captures the exact failure mode.
Final report lists commands and pass/fail evidence, not just conclusions.

Safety Rules

Do not run full /reflect:ingest or /reflect:consolidate inside automated smoke checks.
Do not delete queue files. Use isolated temp state for destructive proofs.
Do not use tmux kill-server, pkill tmux, wildcard tmux kills, or process name kills.
Prefer read-only live checks unless the user explicitly asks to repair live state.