agentleFS
armelhbobdadarmelhbobdad/oh-my-skills/.claude/skills/skf-audit-skill/SKILL.md

skf-audit-skill

Skill7 starsChanged 8 days agoApache-2.0View on GitHub

Drift detection between skill and current source code. Use when the user requests to "audit a skill" or "audit skill" for drift.

The file

We split this skill into the rules it gives an agent

8 rules. About 0 are common, 2 are distinctive, and 6 are very specific to this project.

  • ≈0

    Common · estimate

    Many files like this have a rule of this kind.

  • ≈2

    Distinctive · estimate

    This author's own choices. Worth a look.

  • 6

    Very specific

    Tied to their project's names, files or tools. Probably skip.

  • Common
  • Distinctive
  • Very specific
  1. Distinctive:

    Never fabricate findings — all data must trace to source code with file:line citations

    line 29
  2. Distinctive:

    Only load one step file at a time — never preload future steps

    line 30
  3. Very specific:

    Update stepsCompleted in output file frontmatter before loading next step

    line 32
  4. Very specific:

    Always communicate in {communication_language}

    line 33
  5. Very specific:

    If {headless_mode} is true, auto-proceed through confirmation gates with their default action and log each auto-decision.

    line 34
  6. Very specific:

    Every HARD HALT names its exit code, halt_reason and phase.

    line 35
  7. Very specific:

    Run state (the gates' decisions, the warnings, the emitter's payloads) lives in {run_dir}: step 6 deletes it, a HALT keeps it

    line 37
  8. Very specific:

    Pre-flight: uv, the helpers and the run folder: Before the first prompt, resolve {emitEnvelopeHelper} ← …

    line 91

≈ “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.

Keep what fits and drop the rest. Your version, ready for your agents.
Files
SKILL.md

Audit Skill

Overview

Detects drift between a skill and its current source code and writes a severity-graded drift report with AST-backed findings and remediation suggestions; analysis depth follows the forge tier (Quick/Forge/Forge+/Deep). Stack skills: a compose-mode stack checks its constituents' freshness by metadata hash; a code-mode stack is diffed library by library. A docs-only skill is audited by its documents' hashes; a skill with no provenance map has no baseline, so step 1 stops it.

Conventions

  • Bare paths (e.g. references/<name>.md) resolve from the skill root.
  • Module-level path exception: bare paths beginning with knowledge/ or shared/ resolve from the SKF module root ({project-root}/_bmad/skf/ installed, src/ in dev), not the skill root.
  • Sibling skills: a path that names another SKF skill's folder (skf-<name>/...) resolves from the SKF module root, and that skill must be installed with this one.
  • scripts/ holds render-drift-tables.py, which steps 3, 5 and 6 run to print their drift tables.
  • {skill-root} resolves to this skill's installed directory (where customize.toml lives, if present).
  • {project-root}-prefixed paths resolve from the project working directory.
  • {skill-name} resolves to the skill directory's basename.
  • Cross-skill data coupling: re-index.md loads extraction-patterns.md and tier-degradation-rules.md from skf-create-skill/references/.

Role

You are a skill auditor in Ferris Audit mode: a deterministic drift-detection workflow where the source code is the ground truth and every finding traces back to it.

Workflow Rules

  • Never fabricate findings — all data must trace to source code with file:line citations
  • Only load one step file at a time — never preload future steps
  • Run every subagent in the foreground, passing run_in_background: false on every Agent call (Claude Code runs a subagent in the background when the flag is left out), with parallel ones as several such calls in one message, and wait for every long-running command to exit before the next instruction; in a headless run, never end a turn while either still runs, since under claude -p a run stops when its turn ends.
  • Update stepsCompleted in output file frontmatter before loading next step
  • Always communicate in {communication_language}
  • If {headless_mode} is true, auto-proceed through confirmation gates with their default action and log each auto-decision; step 1's gates (manifest-vs-link, upstream-drift, baseline confirm) also record theirs in the run sink as they decide
  • Every HARD HALT names its exit code, halt_reason and phase; in headless mode it prints its envelope through the shared emitter, as each stage's Halt envelope paragraph shows
  • Once the [C] of references/upstream-checkout.md has bound {source_tree} (the private tree it read the upstream ref into), every step reads the source there, never at the recorded source_path, and every HALT after it first runs uv run {sourceTreeHelper} close --tree "{source_tree}" from {project-root} and goes on whatever it prints
  • Run state (the gates' decisions, the warnings, the emitter's payloads) lives in {run_dir}: step 6 deletes it, a HALT keeps it

Stages

#StepFileAuto-proceed
1Initialize & Baselinereferences/init.mdNo (confirm on a doubt: tier below the compile tier, or provenance map over 90 days old)
1bUpstream Checkout (only when upstream moved)references/upstream-checkout.mdNo (Upstream-Drift Gate)
1cConstituent Freshness (compose-mode stacks only)references/constituent-freshness.mdYes
2Re-Index Sourcereferences/re-index.mdYes
3Structural Diffreferences/structural-diff.mdYes
4Semantic Diffreferences/semantic-diff.mdYes (skip at non-Deep)
5Severity Classificationreferences/severity-classify.mdYes
5aDoc Driftreferences/doc-drift.mdYes
6Reportreferences/report.mdYes
7Workflow Health Checkreferences/health-check.mdYes

Stage 1b returns to init.md at §5b's Record for report, then §6. Step 1 §4 picks stage 1c from the provenance map: it replaces stages 2 to 4 for a compose-mode stack, which has no source tree to re-index (init.md → constituent-freshness.md → severity-classify.md). A docs-only skill skips stages 2 to 4 too (step 1 §3): init.md → doc-drift.md → severity-classify.md → report.md.

Invocation Contract

AspectDetail
Inputsskill_name [required], skill_path [optional: the skill folder, bypassing manifest/symlink resolution], tier_override [optional: Quick / Forge / Forge+ / Deep, in place of the detected tier], upstream_drift_choice [optional: C / S / X; answers the upstream-drift gate in references/upstream-checkout.md; C, auditing the upstream ref from a private tree, when unset]
Gatesstep 1: Manifest-vs-Symlink Gate [N/M/X] · Upstream-Drift Gate [C/S/X] (references/upstream-checkout.md, only when upstream moved) · Baseline Confirm Gate [C/X] (only on a doubt: the forge tier is below the compile tier, or the provenance map is more than 90 days old)
OutputsThe drift report, the JSON it was built from and the result contract, in {forge_version}/ (the audited version's folder): references/headless-contract.md names each file
HeadlessEvery gate takes its default action, a pre-supplied upstream_drift_choice answers the upstream-drift gate, tier_override sets the tier (step 1 §2), and each gate's decision lands in the envelope's headless_decisions. A caller reads the run's own SKF_AUDIT_RESULT_JSON line. A run that printed no envelope line has no result to read, whatever the process exit code: it was cut off, or could not write its result, and audit-skill-result-latest.json may then describe an earlier run, so it is never read in its place.
Exit codesreferences/headless-contract.md: each halt class's exit code and halt_reason, and the SKF_AUDIT_RESULT_JSON envelope

On Activation

  1. Load config from {project-root}/_bmad/skf/config.yaml and resolve:
    • project_name, output_folder, user_name, communication_language, document_output_language
    • skills_output_folder, forge_data_folder, sidecar_path
    • timestamp ← now, as YYYYMMDD-HHmmss, fixed for the whole run
    • run_dir ← {project-root}/_bmad-output/.skf-run/skf-audit-skill-{timestamp}, the run folder step 4 below creates
  2. Resolve {headless_mode}: true if --headless or -H was passed as an argument, or if headless_mode: true in {sidecar_path}/preferences.yaml. Default: false.
  3. Resolve workflow customization. Bind {resolveCustomizationHelper} ← {project-root}/_bmad/skf/shared/scripts/skf-resolve-customization.py, else {project-root}/src/shared/scripts/skf-resolve-customization.py (the first that exists), and run:

    uv run {resolveCustomizationHelper} --skill {skill-root} --project-root {project-root} --key workflow

    It merges the bundled {skill-root}/customize.toml with {project-root}/_bmad/custom/skf-audit-skill.toml (team overrides, committed) and .user.toml (personal overrides, gitignored). When it exits non-zero, prints no JSON or is missing, print one line, [activation/warn] customization_resolver_unavailable: <reason> (<reason>: its first stderr line, SKF helper skf-resolve-customization.py not found when the script is missing, no JSON when it printed none). If the resolver cannot run, read {skill-root}/customize.toml alone and use its bundled defaults: the {project-root}/_bmad/custom/ overrides do not apply to this run. Then keep the reason as {customization_resolver_unavailable}, which step 6 hands to the emitter as a warning.

    Bind the two scalars the stages read, each to its bundled default when the merged value is empty or absent:

    • {driftReportTemplatePath} ← workflow.drift_report_template_path, else assets/drift-report-template.md
    • {onCompleteCommand} ← workflow.on_complete, else empty (step 6 then runs no hook)

    Run workflow.activation_steps_prepend now, treat workflow.persistent_facts as standing context for the run (a file: entry loads its file or glob contents as facts: the bundled default loads every project-context.md under {project-root}; an entry prefixed ! drops each earlier entry it names and loads nothing itself, so an override's "!file:{project-root}/**/project-context.md" turns that default off), then run workflow.activation_steps_append once step 4 below has passed.

  4. Pre-flight: uv, the helpers and the run folder. Before the first prompt, resolve {emitEnvelopeHelper} ← {project-root}/_bmad/skf/shared/scripts/skf-emit-result-envelope.py, else {project-root}/src/shared/scripts/skf-emit-result-envelope.py. If neither exists, HALT (exit code 3, halt_reason: "helper-missing") and display only: "Audit Skill cannot run without skf-emit-result-envelope.py, which is not installed. Re-install SKF."

    Run uv --version, and check that skf-load-provenance.py, skf-structural-diff.py, skf-severity-classify.py and skf-detect-docs.py each exist in {project-root}/_bmad/skf/shared/scripts/ or {project-root}/src/shared/scripts/. If uv --version fails, or a helper is in neither folder, HALT (exit code 3, halt_reason: "helper-missing") at phase on-activation:helpers, naming what is missing: "Audit Skill cannot run without <uv or the script>, which is not installed. Install uv (https://docs.astral.sh/uv/) or re-install SKF, then re-run." Then create the run folder:

    mkdir -p "{project-root}/_bmad-output/.skf-run" && mkdir "{run_dir}"

    If the command fails, HALT (exit code 4, halt_reason: "write-failed") at phase on-activation:run-folder: "Cannot create the run folder {run_dir}: {the first stderr line}." With no folder to stage in, a headless HALT of this step passes its payload to the emitter directly, through python3 in place of uv run when uv is what is missing:

    uv run {emitEnvelopeHelper} emit-halt --workflow skf-audit-skill --target stderr <<'SKF_AS_HALT'
    {"phase": "<the phase>", "reason": "<the halt message>", "halt_reason": "<the halt_reason>"}
    SKF_AS_HALT
  5. Load, read the full file, and then execute references/init.md to begin the workflow.

Download now (ZIP)

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
What did you use it for? e.g. Reviewing pull requests in a Rails app

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.