القائمة
Hunt for bugs in NeuNorm's src/neunorm using parallel Claude auditors plus, when installed, an independent Codex review, then consolidate verified findings. Use before a release or when reviewing risky changes.
التثبيت باستخدام Codex أو Claude انسخ هذا Prompt والصقه في Codex أو Claude أو مساعد آخر ليراجع صفحة Skill ويثبّتها لك.
استنادا إلى تصنيف SOC المهني
Detection-only two-LLM-family review of the current branch's diff vs next (or a named branch/scope). Runs Claude + Codex finders through the dual-family-review engine, cross-family verifies P0/P1 findings, and reports at a human gate. Use to review a PR/branch before pushing or merging. Never edits, commits, or pushes.
Cut a NeuNorm release — pre-flight checks, promote next->qa->main, tag the version, watch the publish pipeline, and verify published artifacts. Use when shipping a new NeuNorm version to PyPI + conda.
Post-merge integration gate for NeuNorm — clean rebuild, full test run, and housekeeping (close addressed issues, update the task tracker / memory) after a PR lands on next. Use right after merging a PR.
| name | bughunt |
| description | Hunt for bugs in NeuNorm's src/neunorm using parallel Claude auditors plus, when installed, an independent Codex review, then consolidate verified findings. Use before a release or when reviewing risky changes. |
Goal: surface real, verified defects in src/neunorm (or a given scope), via a
Claude-native audit plus an automatic second opinion from the Codex CLI when it
is available.
Scope. Default to all of src/neunorm. If the user named modules or a
diff range, scope to those. List the files in scope.
Claude audit (always). Launch the bug-hunter subagent (Agent tool,
subagent_type: "bug-hunter"). For a large scope, run several in parallel —
one per subsystem (loaders, processing, tof, pipelines,
exporters+filters) — for breadth. Collect their structured findings.
Codex second opinion (only if available). Check command -v codex.
codex exec "Review <files> in src/neunorm for correctness, numerical, and edge-case bugs. Output: severity | file:line | issue. Do NOT edit files."
(write only to /tmp if it needs scratch space).Consolidate & verify. Merge Claude + Codex findings, de-duplicate by
file:line. Re-confirm each candidate against the actual code before keeping
it; drop anything you cannot ground. Diverging Claude/Codex verdicts are worth
highlighting.
Report. Verified findings grouped by severity (P0/P1/P2/nit) with
file:line and a fix suggestion, plus a short "what was checked / what each
reviewer contributed" summary.
The steps above are the lightweight, interactive default. For an exhaustive,
reproducible pass — e.g. the pre-release audit — escalate to the deterministic
bughunt-deep workflow (.claude/workflows/bughunt-deep.js): it fans out one
finder per subsystem in parallel, adds the Codex lane, dedups across all
finders, then has three independent skeptics verify every finding (majority
keeps it), returning only confirmed findings grouped by severity.
Run it with the Workflow tool (name: "bughunt-deep", optional args = a scope
string). It is multi-agent and token-heavy, so it requires explicit user
opt-in — use it for release gates, not routine checks.
Do not fix anything in this skill — fixing is a separate, explicit step.