Idd Spec Audit

IDD Spec Audit

scripts/audit-docs.mjs catches byte-level drift — sync-pair mismatches, budgets, config-vs-instruction agreement — but not semantic drift: prose that contradicts a sibling file, leaked session context, passages that are not completable from a cold read, or wording that stalls an autonomous step. This skill runs that check as N parallel LLM read passes, adapted from mew-ton/soloscrum's define-pr-lifecycle audit model.

Scope

• Audit targets (findings may cite these): .github/instructions/**/*.md (including lite/), the issue-authoring skill bundle at its installed location, and every agent entry file present in this installation (CLAUDE.md, AGENTS.md, GEMINI.md, and .github/copilot-instructions.md). All but the first are conditional on the adopter's own setup: onboarding creates each entry file unless the operator explicitly opted out of it, the issue-authoring companion is opt-in, and .github/copilot-instructions.md is touched only if it already existed — skip an audit target that does not exist in the current installation rather than fail the run over it. Cover every present entry file, not just CLAUDE.md: onboarding requires CLAUDE.md, AGENTS.md, and GEMINI.md to agree on repository-specific guidance, so a cross-file contradiction (R2) or a restatement-scope drift (R5) can land in any of them. Audit every target that does exist regardless of whether this installation happens to regenerate it from an upstream source (for example, this source repository regenerates .github/instructions/** from idd-template/ via audit/sync-manifest.json) — this skill audits the corpus a worker session actually reads, not any upstream source, so being a regenerated target never exempts a file here.
• Reference-only inputs (read for R2/R4, never a finding target): docs/idd-concept-ownership.md (R2's closed concept-index seed) and docs/idd-autonomy-contract.md (R4's reversible/irreversible source of truth), both at docs/ in an installed repository. Both are read in full every pass.
• Out of scope as an audit target / finding source: this skill's own bundle wherever it is installed (skills/idd-spec-audit/** — the skill necessarily reads its own bundle, this SKILL.md and references/report-template.md, to run at all, but no finding ever cites a file there); any generated mirror tree in this installation, if one exists (for example, in this source repository, .claude/**, since every file there mirrors a canonical source elsewhere); and every other file under docs/** besides the two reference-only inputs above (summary docs rely on the files they cite by design, so they are not audited as if they were the primary spec).

Rule sets

Run all five rule sets on every pass; do not skip one to save time. A finding names the rule set, the file, the line or section, and a short quote of the offending text.

R1 — leaked session context

Flag prose that reads as belonging to one session's transcript rather than a durable spec: time-relative phrasing without an absolute anchor ("recently", "the issue we just fixed"), first-person session voice ("I noticed", "we decided earlier"), narration of an edit instead of a stated rule ("changed this to require X"), or a workaround described in prose with no tracking link back to the issue that motivated it.

R2 — cross-file contradictions (closed v1 concept index)

Compare every in-scope file against every other in-scope file for a direct contradiction over the same concept. Check only the concepts below — this is a closed v1 index; do not add concepts to it while auditing. Expanding the index is a spec change, not an in-audit decision — file an issue instead of widening scope mid-run. The index is finalized against IDD — Concept Ownership Matrix (docs/idd-concept-ownership.md, #1593):

• claim-marker and activation-nonce semantics;
• advisory-convergence satisfaction;
• merge-gate order (F2/F2.5/F3);
• the "ready = absence of status:* labels" definition;
• phase-digest rules;
• forced-handoff marker semantics;
• the suitability/effort footer contracts.

R3 — fresh-memory completability

Flag a passage that a worker session starting from a cold read (no prior conversation, no memory of another file) could not complete: an unresolved reference ("as described above" with no anchor), an implied prerequisite never stated as a precondition, a missing exit condition (a loop or wait with no stated end state), or a half-named cross-reference (a phase or marker name used before it is defined).

R4 — automation blockers (autonomy cross-check)

Cross-check every instruction that asks an agent to pause, confirm, or escalate against IDD Autonomy Contract (docs/idd-autonomy-contract.md, #1592)'s reversible/irreversible classification, using that table as a comparison baseline rather than re-deriving it from prose — but not as unconditionally authoritative: the contract's own derivation disclaimer states that on any disagreement with an instruction file, the instruction file wins and the contract is the one that needs correcting:

• an instruction to "confirm with the user" (or equivalent) attached to a mutation the contract classifies Reversible is a finding only once the instruction's own described undo path confirms the mutation really is reversible — its named undo path means no confirmation gate is needed there;
• the same phrase attached to a mutation the contract classifies Irreversible is expected behavior and must never be flagged;
• when the table's classification looks wrong against the instruction's actual described behavior, do not flag the instruction as defective — docs/idd-autonomy-contract.md is out of scope as a finding target, so note the suspected contract drift outside this skill's normal finding flow instead (preventive; no observed incident yet — #2782).

A mutation with no row in the contract falls back to the contract's own default (irreversible); that default governs the contract itself; do not extend R4 to independently police no-row mutations beyond the two cases above.

R5 — restatement discipline (closed v1 concept index)

Flag a passage that restates a rule defined canonically elsewhere in the corpus when the restatement's scope does not match the canonical rule's scope: broader than the canonical rule, narrower than it, or phrased as unconditional where the canonical rule is conditional (has stated exceptions, applies only under a named runtime profile, or only within a bounded phase range).

Scope is the same closed v1 concept index R2 uses — reuse the exact list in R2 above rather than introduce a second, open-ended index. A restatement of a concept outside that index is out of R5's scope; do not flag it, no matter how sloppily it is worded. This keeps R5 from treating every emphatic sentence in the corpus as a finding — the rule exists to catch a scope drift on the seven concepts already load-bearing enough to have a closed index, not to police prose style generally.

Preferred remedy: cite the canonical section instead of restating it inline. Prefer See [<section>](<path>#<anchor>) (or an equivalent plain-text pointer to the file/section) over reproducing a multi-clause rule's conditions in a second location — inline restatement of a multi-clause rule is the exact failure mode this rule set exists to catch, and the instruction bundles are already close to their byte budgets, so citing is also the cheaper fix. Note the preferred remedy in the finding so the reader does not have to re-derive it.

Execution model

• Run N parallel, independent, read-only passes over the scope above, skipping any Audit target absent from this installation (see the Scope section's conditional-target note). Default N = 3; accept a --passes N-style argument to adjust it.
• Aggregate by union, deduplicating findings that describe the same file/section/issue across passes. Annotate each surviving finding with Appeared in: K/N (how many of the N passes independently raised it) as informational context only.
• Never apply a quorum filter. A finding raised by only one pass is reported exactly like one raised by all N — sampling variance is not evidence of invalidity, and dropping low-K findings would silently discard true positives that one pass framed differently from the others.
• Read-only, always. This skill never edits an in-scope file and never opens, closes, comments on, or labels a GitHub issue. Route every finding back through the normal issue-authoring flow (see the issue-authoring skill) for a human or a later session to act on; when the issue-authoring companion is not installed (Scope's conditional-target note), route findings through this installation's normal manual issue-filing process instead.
• Write the aggregated result using references/report-template.md.

See also

• references/report-template.md for the report shape.
• IDD Autonomy Contract (docs/idd-autonomy-contract.md) — R4's closed source of truth.
• IDD — Concept Ownership Matrix (docs/idd-concept-ownership.md) — R2's concept-index seed.