judging— agent skill

Overview

This skill dispatches the Court to adjudicate a task's verification result. The Court is not a persona; it is parameterized by the contract's check types. Inputs are the contract, the captured diff, the four-level Verify output, and a confidence score. Plan, commit messages, ledger, and prior verdicts are structurally withheld.

When to Use

Invoked by the orchestrator after verifying returns allGreen on a task, or after any criterion reports suspicious. Not invoked by user prompts directly. Called exactly once per task per loop iteration.

Process

Numbered steps the orchestrator follows:

1. Load the contract via cli/lib/contract.js and confirm it was the contract used for the task.
2. Read the worktree's captured diff at <worktreePath>/anvil/diff.patch.
3. Read the Verify result at <worktreePath>/anvil/verify/verify-result.json.
4. Build the briefing with composeBriefing. The module structurally refuses the commit message, the plan, the ledger, and any prior Court verdict. These are inputs the Court must not see.
5. Dispatch a fresh subagent as the Court. The Court returns a per-criterion status (pass | fail | suspicious) plus an action (merge | request-changes | request-clarification).
6. Validate the returned verdict with validateVerdict. Reject any out-of-schema action or per-criterion status with E_INVALID_VERDICT.
7. Return the verdict plus briefingHash. The orchestrator persists the verdict record via <worktreePath>/anvil/verify/verify-result.json alongside the Verify result; the Court verdict is NEVER carried back into the next task's briefing.

Rationalizations

Reject the following shortcuts:

• "The commit message explains the rationale; let me include it." The Court does not read narration; rationale is an adversarial vector (failure-taxonomy row 2: Convincing rationale inflation).
• "The ledger has prior verdicts that would save time; pass them in." Prior verdicts are not evidence; they are context that can be pattern-matched rather than adjudicated (failure-taxonomy row 2).
• "Three criteria passed; surely the fourth is close enough." Three of four is a fail; the Court adjudicates all criteria, not a majority (failure-taxonomy row 8: Claim-without-evidence).

Red Flags

If any of these conditions obtain, the Court dispatch is rejected:

• The caller tries to pass commit_message, plan, ledger, or prior_verdicts into composeBriefing (failure-taxonomy row 2).
• The returned verdict has an unknown action or an out-of-schema per-criterion status (structural guard; E_INVALID_VERDICT).
• The Court subagent references the task's commit history or prior verdicts (structural guard; the briefing does not expose these).

Verification

Each Court dispatch checks:

1. composeBriefing rejects every key in FORBIDDEN_COURT_INPUT_KEYS with E_COURT_INPUT_VIOLATION.
2. validateVerdict rejects every out-of-schema verdict with E_INVALID_VERDICT.
3. The judge return value includes briefingHash matching sha256 of the briefing.
4. MUTABLE_STATE is frozen and empty both before and after the call.