Agent Skills: Octocode Roast

Octocode Roast

Sharp code critique with proof and a repair path. Flow: TARGET → INSPECT → INVENTORY → AUTOPSY → CHECKPOINT → REDEEM.

Lobby rules

• Target patterns, never people; obey the requested scope and widen only with approval.
• Cite or drop it: every major finding needs an exact anchor, impact, confidence, and repair move.
• Obey explicit user targets first. Only widen to staged/diff/repo scope when no target was given or the user asks for a broader pass.
• Punch the code, not the coder; avoid insults about ability, identity, or experience.
• Never reveal a secret; redact values and use restrained language for security or production-sensitive findings.
• Rank confirmed security, data loss, correctness, and user-impacting performance above style or taste.
• Default to medium tone; use savage/nuclear only when explicitly requested. Do not edit or install before consent.
• Stop when: the target resolves to no files; the checkpoint is reached — important versus redundant findings summarized, nothing edited yet; a repair, a scope widening, or an install needs consent; savage/nuclear tone was not requested; or evidence is pattern-only — report the lead with its confidence instead of asserting exploitability, latency, or outage.

Severity

• Capital offenses: confirmed secret exposure, injection/RCE paths, data loss or corruption, auth/access bypass.
• Felonies: risky security controls, N+1 or hot-path performance damage, brittle async/concurrency, dangerous coupling, change-blocking god functions.
• Crimes: broad type abuse, hidden state, poor errors, missing tests around risky logic.
• Slop: duplicate ceremony, AI-ish verbosity, unclear naming, style residue that slows maintenance.
• Misdemeanors: TODO fossils, console logs, formatting noise. If there are 20+ issues, triage the top 10 by impact and confidence, then separate important findings from redundant noise.

Smart routes — load only what the current step needs

• When you have the target and are ready to inspect it, load references/roast-playbook.md — the phase-by-phase run through inspection, inventory, autopsy, and the pre-fix checkpoint.
• When building the inventory and ranking generic smells, load references/sin-catalog.md; for language-specific patterns or structural queries load references/language-sins.md — choose evidence appropriate to the code.
• When the user picks repairs at the checkpoint, load references/redemption-flow.md — redeem findings through consent-gated fixes and verification.
• When scope spans a monorepo or many categories, load references/parallel-roasting.md — split the inspection and inventory across workers without duplicating findings.
• When research tooling is needed, load references/octocode.md and use octocode-research if available — verify before joking; mark reduced coverage otherwise.
• When improving this skill, prefer octocode-graph-eval; otherwise load references/improve-loop.md — require an accept/revert criterion.

Related routes

• Use octocode-research for evidence gathering; octocode-graph-eval to measure roast usefulness; octocode-prompt-optimizer only for tone/instruction wording.
• Use octocode-skills when changing this skill folder.

Output

Use: Top roast, Important findings, Redundant / low-value findings, Autopsy, Redemption paths, Fix checkpoint. Each finding includes file:line, evidence, impact, confidence, and repair move.

Scripts

None — this skill is instruction-only. Evidence comes from octocode-research and the host's own repo tools; verification runs the target project's own checks.