[mlieberman85]

Stacked on #669. Review the last 3 commits: the spec change, get_judgment_requests, and the skills.

Summary

The json audit output is about 1 MB on wtf-frontend, because every control waiting on a judgment carries its own copy of its files and all of its evidence. Claude Code refuses a tool result that large, so the comply flow saved it to a file and parsed it with shell tools (#644).

• get_judgment_requests audits the checkout the way submit_judgment does. Per control, it returns only the prompt, hints, evidence digest and file keys, with each file once in a shared files map: exactly the content submit_judgment checks citations against.
  • Pages stay under 40,000 bytes, and remaining names the controls for the next call.
  • Controls with a stored PASS candidate are skipped unless include_judged is set.
  • It records nothing.
• Summary rows of a PENDING control carry pending.kind, and for a candidate its model, reasoning and number of cited excerpts. A client can tell a judgment request from a candidate without the json output.
• Skills: darnit-audit and darnit-comply judge from get_judgment_requests. The "~164K" size claims are removed.
• Spec: 5.4.
• Tests: tests/integration/test_mcp_server.py now expects 23 tools.

Refs #644.

Type of Change

• New feature (non-breaking change adding functionality)

Framework Changes Checklist

• Updated framework spec (docs/architecture/framework-design.md, 5.4)
• Ran uv run python scripts/validate_sync.py --verbose and it passes

Testing

• uv run pytest tests/ on the whole rebased stack (18bd63e): 5286 passed, 18 skipped
• e2e: see the table below
• uv run ruff check .

End-to-end

darnit-e2e: headless /darnit:darnit-audit openssf-baseline through Claude Code. Confirm tools are disallowed and the repository is named in the prompt.

darnit	darnit repo	wtf-frontend	cosign-installer
main (2026-10-20; cosign-installer on 603bdda)	387s, 44 turns	415s, 47 turns	297s
#653 only (2026-10-21)	299s, 57 turns	250s, 55 turns	185s, 39 turns
full stack, rebased on 8070949 (2026-10-23)	282s, 50 turns, 30 judgments	145s, 23 turns, 15 judgments	141s, 30 turns

• Results: per-control results match main. The only differences are controls the model judged, and the effect of operator-side answers on wtf-frontend, which needs the repository named. wtf-frontend also shows QA-03.01 and QA-07.01 passing because their platform settings changed on GitHub.
• Stored candidates skew the comparison. A control with a stored PASS candidate for unchanged evidence is skipped, so run-to-run times depend on how many are stored. darnit's HEAD moved to 8070949, which invalidated its stored candidates, so it was judged in full (30 controls).
• Denials:
  • darnit: one page of get_judgment_requests was larger than Claude Code's tool-result limit and was saved to a file. The agent's attempt to read it with Bash was denied. This is the oversized single-request follow-up.
  • wtf-frontend: the agent tried to run zizmor itself, and that was denied.

AI assistance

• No AI assistance was used
• AI assistance was used

Claude Code (Claude Opus) wrote the change, tests and this description; I reviewed them and ran the e2e.

---

Drafted with AI assistance (Claude), reviewed and filed by me. Measurements quoted were executed and verified.

[Marc-cn]

Reviewed its own three commits. Right answer to the 1 MB audit JSON: shared files map keyed by path (#n on digest collisions), per request only what submit_judgment checks citations against, 40 KB pages with remaining fed back as control_ids so later pages are subset audits, summary rows carrying pending.kind, and structured_output=False so the size counted is the size sent. One note: a PENDING llm_judgment control without an llm_consultation is dropped silently (neither in requests nor remaining); #671's 'list the controls skipped' commit looks like it covers that, so fine in the stack. Good to merge in order.