dual-spec-compliance

You are running the spec-compliance skill twice — locally as Claude, and remotely via the codex CLI — and merging the results. Be terse. The user wants the union of findings, not two separate reports.

ARGUMENTS: <target> — a package directory under pkg/ (e.g. pkg/proofs/internal/meta/maurer09, pkg/base/base58). Pass it through verbatim; don't second-guess the form.

Steps

1. Preflight. Run command -v codex via Bash. If it errors, stop and tell the user to install the OpenAI Codex CLI (npm i -g @openai/codex or equivalent) and authenticate.

2. Capture the target. Treat the user's argument as an opaque string <T>. Derive a filesystem-safe slug <S> (replace /, :, .., #, spaces with _) for use in tmp filenames.

3. Build the Codex prompt. Read .claude/skills/spec-compliance/SKILL.md and append a final line ARGUMENTS: <T>. Write the combined text to /tmp/dual-spec-compliance-prompt-<S>.md.

4. Launch Codex in the background. Use Bash with run_in_background: true. Pipe the prompt via stdin — the spec-compliance skill starts with --- (YAML frontmatter), which codex's arg parser treats as a flag separator and rejects with exit 2 if passed positionally:

   cat /tmp/dual-spec-compliance-prompt-<S>.md | \
     codex exec --skip-git-repo-check --sandbox read-only --full-auto - \
     > /tmp/dual-spec-compliance-codex-<S>.md 2>&1
   

   --full-auto is required for non-interactive background use (otherwise codex blocks on approval prompts). --sandbox read-only keeps codex from touching the working tree. You'll be notified when it finishes. Do not poll.

5. In parallel, run spec-compliance yourself by following .claude/skills/spec-compliance/SKILL.md for target <T>. Resolve every <!-- pdf|paper|spec|code[…]: … --> directive, fetch URLs via WebFetch, read PDFs via Read, and produce the full report (template and all). Hold it in memory; do not print it yet.

6. When the background task completes, read /tmp/dual-spec-compliance-codex-<S>.md. If it's empty or errored, skip to step 8 and note the dual leg failed — don't suppress your own findings to fit the template.

7. Merge. Walk both reports finding-by-finding. Two findings match if they cite the same file.go:line (±a few lines) and target the same reference anchor (same paper §/figure, same spec MUST clause, same upstream function), or describe the same root cause attributed to the same reference in different words. A finding cited against KOS15 §3.2 and one cited against RFC9180 §4.1 are not matches even if they touch the same line — they're describing different obligations. Preserve each finding's original severity tag ([deviation vs. <ref>], [missing vs. <ref>], [addition vs. <ref|all references>], [question vs. <ref>]) and original wording — do not paraphrase. For agreed-by-both findings:

   • Prefer the more specific reference anchor if the two diverge in precision (KOS15 §3.2 Thm 3 cond. (1) over KOS15 §3).
   • If the kinds disagree (e.g. one says [deviation], the other says [question]), cite both: [deviation vs. KOS15 §3.2 — Claude] / [question vs. KOS15 §3.2 — Codex].
   • Keep the _Impact:_ sub-bullet from whichever side has the more concrete property-at-stake (soundness / ZK / interop / round-trip). If both have one, keep both, attributed.

8. Print the merged report in this exact order:

   ## Dual spec-compliance — `<T>`
   
   **References (union):**
   - `<short-name-A>` (paper) — <role>. Locator: `<status>` (Claude) / `<status>` (Codex)
   - `<short-name-B>` (spec) — <role>. Locator: `<status>` (Claude) / `<status>` (Codex)
   - `<short-name-C>` (code) — <role>. Locator: `<status>` (Claude) / `<status>` (Codex)
   
   _(If the two agents disagree on which references exist, list the union and note the disagreement on the affected rows.)_
   
   **In-repo spec:** <synthesised; note disagreement if any>
   **Acknowledged drift (from README):** <one-liner each, or "none">
   
   ### Reference ↔ Reference consistency
   
   - `<short-name-A>` ↔ in-repo spec: <verdict — note disagreement between Claude and Codex if any>
   - `<short-name-B>` ↔ `<short-name-C>` (where they overlap): <…>
   
   ### Findings unique to Claude
   - **[deviation vs. KOS15 §3.2]** `file.go:42` — …
     - _Impact:_ …
   
   ### Findings unique to Codex
   - **[missing vs. RFC9180 §4.1 (MUST)]** `file.go:99` — …
     - _Impact:_ …
   
   ### Findings agreed by both
   - **[deviation vs. bitcoin/base58.cpp:DecodeBase58]** `base58.go:88` — …
     - _Impact (Claude):_ …
     - _Impact (Codex):_ …
   
   ### Summary
   
   **Counts (union):** N deviation, N missing, N addition, N question.
   **Per-reference coverage:** `<short-name-A>` (paper): N findings; `<short-name-B>` (spec): N; `<short-name-C>` (code): N.
   **Highest-risk:** `path:line` (vs. `<short-name>` §X) — <property at stake>.
   
   ---
   
   ### Raw — Claude
   
   <verbatim Claude spec-compliance output>
   
   ### Raw — Codex
   
   <verbatim Codex spec-compliance output>
   

   If a bucket is empty, write _None._ rather than omitting the header — the user wants to see at a glance whether each agent found something the other missed. Order findings within each bucket as spec-compliance prescribes: deviations first, then missing, then additions, then questions; within each kind, order by reference short-name (alphabetical), then file path, then line.

Don'ts

• Don't drop a finding because the other side missed it. Bucketing exists precisely to surface those.
• Don't paraphrase severity tags or finding text in the buckets — copy them verbatim. Only the trailing Summary is synthesised.
• Don't merge findings that cite different references even if they're on the same line. A [deviation vs. KOS15] and a [deviation vs. RFC9180] on file.go:42 are two findings, not one — they're claiming the line violates two different obligations.
• Don't post the report anywhere; just print. No PR comments, no commits.
• Don't strip the raw reports — they're the audit trail for the merge.
• Don't run codex exec in the foreground; it blocks long enough that running spec-compliance yourself in parallel is the whole point.
• Don't edit code (the inner spec-compliance skill forbids it; the dual wrapper inherits that).
• Don't fetch references twice — the inner skill is responsible for resolving directives. If Claude resolved a PDF and Codex didn't (e.g. because the PDF is in a developer-local .docs/), surface the asymmetry rather than re-fetching for Codex's benefit.

Caveats

• Codex inherits cwd and shell env, so git auth and repo state carry over. PDFs in .docs/ (gitignored) and docs/papers/ are visible to both as long as the working tree has them. URLs resolved via WebFetch are Claude-only — Codex's URL fetch may be blocked by its sandbox, so expect occasional asymmetry on <!-- spec[…]: <url> --> and <!-- code[…]: <url> --> directives. Surface these as "Locator: resolved (Claude) / unresolved (Codex)" rather than discarding Codex's findings on the affected reference.
• Codex output is whatever its model decides to produce; it may not follow the spec-compliance template exactly (severity tags, reference-↔-reference section, per-reference coverage). Bucket what's there; don't try to rewrite it. If Codex omits the consistency cross-check or per-reference counts, fall back to your own.
• If the two reports disagree on a fact (e.g. one claims Anchor.PreImage is verified at Extract, the other claims it isn't), put it in unique to <agent> rather than agreed-by-both. Don't arbitrate — that's the user's job.
• Reference attribution is the highest-signal part of the merge: a finding the user can trace back to "paper §3.2 condition (1)" or "RFC 9180 §4.1 MUST" is far more actionable than a generic "deviation". Preserve those anchors verbatim from each side; don't normalise them across agents.
• If only one side resolved a reference (e.g. Claude WebFetched the wiki, Codex didn't), the agent that didn't resolve will inevitably under-report findings against that reference. Note this in the per-reference coverage line so the user knows the asymmetry is structural, not a real disagreement.
• Reductions / paper-alignment findings are easy to miss for whichever agent didn't resolve the underlying PDF/URL. If only one side cites a particular reference, treat that as a real finding under unique to <agent>, not noise.