Necessity Audit

3-phase necessity audit with Codex adversarial debate. Identifies over-designed elements in lifecycle specs via 6-dimension YAGNI rubric.

Non-Negotiable Rules

SKILL.md is the normative source. Reference files elaborate but do not override.

Trigger

• Keywords: necessity audit, over-design, YAGNI audit, spec necessity, 過度設計, over-engineered

When NOT to Use

Alternatives by intent

Chain recommendation

/codex-review-doc (detail) → /codex-review-spec (reasoning, planned) → /necessity-audit (necessity, this skill) → /feature-completeness (completeness, planned) → /review-spec (synthesis)

Arguments

Workflow

Phase 0 preflight → Phase A classify → Phase B Codex debate → Phase C consolidate → Redact → Emit

Phase 0: Preflight (executable)

Scratch directory — read this before running any step. Each Bash invocation is a fresh shell: a variable assigned in one step does not exist in the next. Do not write TMPDIR=$(mktemp -d) and then reference $TMPDIR later — on macOS TMPDIR is an ambient variable already pointing at the shared temp root (/var/folders/…/T/), so later steps silently read and write there, and the final rm -rf $TMPDIR would target that shared root.

Instead: run mktemp -d once, read the path it prints, and substitute that literal absolute path into every later command. The placeholder <AUDIT_TMP_DIR> below marks each substitution site. Never name the variable TMPDIR.

mktemp -d
# → e.g. /var/folders/ab/cd1234/T/tmp.XyZ123 — reuse this literal path below as <AUDIT_TMP_DIR>

Immediately claim it. The claim mints a one-time capability token and stores it in a marker inside the directory; the cleanup step requires that exact token back. This is what binds the delete to this run's directory rather than to any directory that merely looks like one — or to another concurrent audit's directory, which also carries a valid marker:

node scripts/skills/necessity-audit/cleanup.js --claim "<AUDIT_TMP_DIR>"
# → token=3f9c…  (48 hex chars) — reuse this literal token in Phase 4 as <AUDIT_TOKEN>

Read the token= line it prints and carry that literal value to the cleanup step, the same way you carry the directory path. Like <AUDIT_TMP_DIR>, it cannot be held in a shell variable — each Bash invocation is a fresh shell.

node scripts/skills/necessity-audit/preflight.js \
  --path <path> --depth <depth> \
  [--skip-preflight] [--include-feasibility] \
  --output "<AUDIT_TMP_DIR>/preflight.json"

Non-zero exit = hard block. Read <AUDIT_TMP_DIR>/preflight.json to continue.

Phase A: Claude classify (LLM)

Read target file with Read tool. Apply references/phase-a-classify.md template substituting ${TARGET_PATH}, ${DOC_KIND}, ${ACTIVE_DIMENSIONS}, ${GREENFIELD} from preflight.

Extract elements (FR / NFR / Component / Abstraction / Extensibility / Config), score each against active dimensions only (depth=brief → dims 1-3; normal/deep → dims 1-6), assign initial Keep/Review/Cut.

Write result: Write tool → <AUDIT_TMP_DIR>/phase-a.json with schema { elements: ClassifiedElement[] } (only claude.* fields populated).

Phase B: Codex debate (Skill invocation)

node scripts/skills/necessity-audit/debate-topic.js build \
  --preflight "<AUDIT_TMP_DIR>/preflight.json" \
  --output "<AUDIT_TMP_DIR>/topic.txt"

Read topic, invoke:

Skill("codex-brainstorm", <contents of <AUDIT_TMP_DIR>/topic.txt>)

Write raw response: Write tool → <AUDIT_TMP_DIR>/debate.txt.

node scripts/skills/necessity-audit/debate-topic.js parse \
  --input "<AUDIT_TMP_DIR>/debate.txt" \
  --output "<AUDIT_TMP_DIR>/debate.json"

Phase C: Consolidate (executable)

node scripts/skills/necessity-audit/consolidate.js \
  --phase-a "<AUDIT_TMP_DIR>/phase-a.json" \
  --debate "<AUDIT_TMP_DIR>/debate.json" \
  --preflight "<AUDIT_TMP_DIR>/preflight.json" \
  --overrides "<id>:<rationale>[;...]" \
  --depth <depth> \
  --output "<AUDIT_TMP_DIR>/report.json"

Applies 6 deterministic checks, under-coverage check, --override handling, gate selection.

Assemble + Redact + Emit

node scripts/skills/necessity-audit/report.js \
  --input "<AUDIT_TMP_DIR>/report.json" \
  --format markdown \
  --output "<AUDIT_TMP_DIR>/report.md"

node scripts/skills/necessity-audit/redact.js \
  --input "<AUDIT_TMP_DIR>/report.md" \
  --output "<AUDIT_TMP_DIR>/report.final.md"

Read <AUDIT_TMP_DIR>/report.final.md and emit as final user-visible message.

Cleanup. The path condition is enforced by the script, not by this paragraph. It refuses (exit 1, deleting nothing) on an unsubstituted placeholder, a relative path, a symlink, a non-directory, the filesystem root, anything that is not a direct child of this process's temp root — in particular the shared temp root itself, the path an ambient-TMPDIR mistake produces — any directory carrying no --claim marker, and, decisively, any directory whose marker was minted with a different token. That last check is what rules out a substitution naming a different, equally valid-looking scratch directory: shape is common to all of them and so is the marker, since every concurrent audit claims its own — only the token is unique to this run. Removal is idempotent, so re-running after a partial failure is safe:

node scripts/skills/necessity-audit/cleanup.js --dir "<AUDIT_TMP_DIR>" --token "<AUDIT_TOKEN>"

Before running it, re-read the path you are substituting and confirm it is the exact absolute path Phase 0 printed. The guard refuses a wrong one, but a refusal is a stalled cleanup — getting the substitution right is still your job. If the guard does refuse, do not work around it with a manual rm: fix the substitution.

Output Format + Gate Selection

Output header, sections, sentinel: see references/output-template.md (normative). Gate-selection decision table + narrative rules: see references/phase-c-consolidate.md.

Invariant: ⚠️ Need Human NEVER appears as the final gate — only as a narrative line above the ✅ Audit Clear / ⛔ Audit Revise sentinel.

Review Loop (--continue)

After user revises the spec, re-run with --continue <threadId> to reuse the Codex debate context via @skills/codex-code-review/references/codex-transport.md § Resume. See references/review-loop.md.

References

• references/dimensions.md — 6-dimension × 4-tier rubric (authoritative)
• references/phase-a-classify.md — Phase A prompt template
• references/phase-b-debate-topic.md — Phase B topic builder documentation
• references/phase-c-consolidate.md — Phase C logic
• references/output-template.md — Markdown report layout
• references/review-loop.md — --continue flow
• references/redaction-rules.md — Secret / PII patterns applied by redact.js

Verification

• Phase B used Skill("codex-brainstorm"), not a raw transport dispatch
• Report contains non-empty debate.threadId
• Report contains non-empty Debate Conclusion
• Output starts with ## Necessity Audit header
• Output ends with ✅ Audit Clear OR ⛔ Audit Revise sentinel
• Output contains NO doc-review gate sentinel (the Mergeable / Needs revision pair) anywhere — an audit must not be mistakable for a doc review
• ⚠️ Need Human never used as gate (only as narrative)
• Redaction applied before emission

Examples

Input: /necessity-audit docs/features/foo/2-tech-spec.md
Action: Phase 0 preflight → Phase A classify → Phase B debate → Phase C consolidate → report + redact → emit with sentinel

Input: /necessity-audit docs/features/foo/2-tech-spec.md --continue 019dab42-xxxx
Action: Resume per `@skills/codex-code-review/references/codex-transport.md` § Resume; re-run Phase C with updated spec; emit diff-focused report

Input: /necessity-audit docs/features/foo/1-requirements.md --depth brief --skip-preflight
Action: Only challenge dims 1-3; skip state advisory; emit [PREFLIGHT SKIPPED] banner

Input: /necessity-audit docs/features/foo/2-tech-spec.md --override FR-12:"needed for Q3 rollout"
Action: FR-12 kept with justification; final gate ✅ Audit Clear if no other Cut items remain