bugslisted

# /bugs (v3.1) Deep, evidence-first bug analysis. **The engine is Claude subagents** spawned via the Agent/Task tool — the skill works whenever Claude does, with no dependency on any external CLI. TLDR context optimization keeps token cost low. ## v3.1 — Claude is the engine (was Codex-only) Earlier versions shelled out to `codex exec -m gpt-5.2-codex` as the *only* execution path, so the skill broke whenever Codex was unavailable (rate limits, auth failure, not installed) — despite claiming to be "model-agnostic". It is now genuinely Claude-native: - **Primary engine: Claude subagents.** Bug hunting runs through the Agent/Task tool with a bug-hunter prompt and the evidence-first contract below. This is the default and always works. - **Codex is an OPTIONAL accelerator, never required.** If the `codex` CLI is installed AND responsive, you MAY fan a second opinion out to it and merge findings — but you must first confirm it responds (a rate-limited or unauthenticated `codex` is treated as absent). If it is unavailable, proceed with Claude alone and say so; never block on it. - **Model selection** for the Claude subagents follows the session's configured model (`~/.claude/settings.json` / env), inherited by the spawned agents. ## Evidence-First Contract (non-negotiable) Bug hunting is evidence work, not a confidence exercise. Report ONLY issues grounded in changed code, reachable paths, a test failure, runtime evidence, or a clear invariant violation. If a sus