roast-review
Skill2 starsChanged 22 days agoMITView on GitHub
Post-roast debrief — pull a completed roast's trace + MCP logs and assess what worked / what to change (T0, FC, drop, dev%, trim, RoR, safety verdicts) against the profile targets and the operator's known-good thresholds. Use after a roast to review it, or when asked "review the roast".
1 other project uses a version of this.
Adopted by
Most recently taken up by
- aewing/skillstheir own version, first seen Oct 13, 2026
Other versions1 project uses a version of this.
Word for word · 0
None found.
Edited · 0
None found.
Wrote their own version · 1
- aewing/skills★ 1same name
How matches are found. Word for word: the exact same text. Edited: shares passages with this file (“99% same” means 99% of its five-word phrases also appear here; 10% or more counts). Their own version: similar meaning, or, weakest, the same name. We can’t tell which project had it first. Projects are counted among the 200 most-starred candidates, and mentions are links to the repository we found, up to 50 per site, so both counts are a floor.
In this repository9 other files it gives its agents.
- AGENTS.mdAGENTS.md
- CLAUDE.mdCLAUDE.md
- capture-agent-usageSkill
- add-bean-profileSkill
- captureSkill
- pre-roast-preflightSkill
- pr-preflightSkill
- register-roastSkill
- triage-prSkill
The file
We split this skill into the rules it gives an agent
2 rules. About 0 are common, 0 are distinctive, and 2 are very specific to this project.
≈0
Common · estimate
Many files like this have a rule of this kind.
≈0
Distinctive · estimate
This author's own choices. Worth a look.
2
Very specific
Tied to their project's names, files or tools. Probably skip.
- Common
- Distinctive
- Very specific
- Very specific: line 11
Default to the latest completed run.
- Very specific: line 16
If the server is down, read the trace DB directly ($ROASTPILOT_DB, default ~/roasts/roastpilot.sqlite3) — roast_runs, roast_events, …
≈ “Common” is an estimate: at least 1 in 5 of the 372,429 skill files we read have a rule of the same kind, not the same words. “Very specific” is our guess from the words: it names a file, a command, a setting or the project itself.
Files
Codifies the post-roast review. Produce an honest, data-backed debrief — not a summary — and end with the rate + fixture closeout. All temperatures Celsius.
1. Resolve the run
Default to the latest completed run; accept an explicit run id if given. The
live server (default :8000) is the easiest source; fall back to the DB.
!curl -sf -m5 http://127.0.0.1:8000/api/roasts | python3 -c "import json,sys; d=json.load(sys.stdin); r=sorted(d.get('runs',[]),key=lambda x:x.get('started_at_utc',''),reverse=True); print('\n'.join(f\"{x['id']} | {x.get('started_at_utc','')[:19]} | {x.get('outcome')} | FC={x.get('first_crack_at_utc')}\" for x in r[:5]))"
If the server is down, read the trace DB directly ($ROASTPILOT_DB, default
~/roasts/roastpilot.sqlite3) — roast_runs, roast_events, telemetry,
safety_evaluations, advisor_decisions. Remember the clock mismatch:
telemetry is on a run-relative elapsed_seconds; events are on absolute
monotonic_seconds — rebase events to the t0_detected event's monotonic.
2. Pull the data
For the chosen <RID>:
GET /api/roasts/<RID>/timeline→events(kind, source, monotonic_seconds, payload),safety_evaluations,advisor_decisions,commands.GET /api/roasts/<RID>/telemetry→points(tick, elapsed_seconds, bean_temp_c, env_temp_c, bean_ror_c_per_min, heat/fan_level_percent, development_percent).- MCP export:
~/roasts/logs/roasts/<MCP_SESSION>/{summary.json,roast.jsonl}— the MCP session id ≠ the agent run id; match by start time. Readsummary.json → first_crack_model(the model that detected FC; all-null ⇒ FC was operator-marked) and any per-window confidence (#175, once it lands).
3. Compute the milestones (rebased to T0)
- T0: bean temp at
t0_detected(payloadbean_temp_c). ⚠️ Flag if it fired near the turning-point minimum (~150) rather than the charge/decline onset (~179) — that is mcp #174, ~15 s late. - FC:
first_crackeventsource—mcp= the audio detector fired;operator= the detector MISSED and you hand-marked it (flag mcp #175; note how late vs the real crack ⇒ under-counts dev%). - Drop: the
phase_changed → coolingevent; bean temp ≈ the peak bean temp; the dev% at drop = the last non-coolingdevelopment_percent. Was it the auto-drop (anadvisor_decisionsrow withdecision.should_drop=truethat the safety guard ALLOWED) or held (guard REJECTed while the bean climbed — the #323 conflict) or your manual drop? - Control: did the trim engage (heat 100 → ~65 before FC)? heat → 0 post-FC? fan used as a brake? RoR shape — steady declining (good), a crash (RoR → ~0 within 90 s of FC = baked), or a flick (RoR re-accelerates ~90–120 s post-FC = char)?
- Safety: count ALLOW/CLAMP/REJECT/RECOVERY/FAULT/EMERGENCY_STOP; surface any non-ALLOW with its reason.
4. Assess against targets + the operator thresholds
Compare to the run's frozen profile (target_drop_temp_c, target_development_percent) AND the empirical Hottop ground truth (memories operator-hottop-roast-profile, per-origin-dtr-washed-highgrown):
- Drop: ≤195–196 = good; >197 = over-done/bitter; >200 = likely bad. Compare bean-at-drop to the profile ceiling.
- Dev %: vs the profile target; washed high-grown wants ~18 % eventual, naturals ~13 %; a late operator FC mark deflates the measured number.
- FC: detector fired vs missed; display BT ~178 at the true crack.
- RoR: the dominant lever — a flick chars more than a high DTR does.
5. Output the review
A scannable debrief:
- ✅ What worked — with the numbers.
- ❌ What to change — with the numbers + the candidate follow-up issue for each (mcp #174/#175/#176, agent #380, etc.).
- One-line verdict (clean / over-done / needs control change) and whether it met the run's intent.
6. Closeout
Prompt the operator to:
- Rate the roast in the UI (the D42 corpus label) — score + notes.
- Register the fixture:
python scripts/store_to_fixture.py "$ROASTPILOT_DB" --out-dir ~/roasts/fixtures/<slug> --run-id <RID>(gitignored working dir; real stores are never committed). - File any new issue the review surfaced.
Does it work?
Reviews from people and from AI agents that ran it: what they used it for, and whether it worked.
No reviews yet
People and AI agents can review whether this works in practice. Say what you used it for, and whether it worked.
- It worked
- Partly
- It didn't
Using an agent? It can post a review for you after it has run the file: the MCP tool registry_write, action report. How to connect one.
Your trail
Files you open will collect here, so you can find them again.
Open two or more, and you can remix the rules you like from each into one list. Kept in this browser only.
Now do this for your own.
Bring your CLAUDE.md, AGENTS.md or skills. agentleFS shows which of your rules are common, which are yours and which only fit one project, and keeps one version your agents read everywhere.

