goal

Open a new /goal contract. The manifest you write becomes the binding success criterion for whatever work follows. Workers cannot report done until Warden verifies against this manifest and writes a sign-off proof.

What to do

1. Pick a task-id. Format: YYYYMMDD-<slug>-<4hex>. Slug is 2-5 words from the goal, kebab-case. Hex is 4 random characters to disambiguate.

2. Resolve the role and project. Based on the goal text and .claude/orchestrator/ARGENT.md routing rules, decide which role (forge / quill / scout) will execute the slice and which project overlay applies. If either is unclear, ask the operator. Do not guess.

3. Translate the goal into machine checks. For each thing the operator implied, write one done_when entry. Prefer test and command checks. Fall back to file_exists / file_contains for documentation slices. Use human_review only when the goal is genuinely subjective (Scout findings, design decisions) — and then state explicitly what the reviewer should confirm.

4. Add diff constraints when scope matters. If the operator named specific files, populate diff_constraint.touches_only. If files must not be touched (auth, billing, schema migrations are common), populate forbidden. Skip both for documentation-only slices.

5. Write the manifest to .claude/goals/<task-id>.yaml. Create the directory if it does not exist.

6. Surface the manifest back to the operator before dispatching anything. If a done_when check is fuzzy ("works correctly," "is fast enough"), say so and ask the operator to tighten it. A vague manifest is worse than no manifest — it gives Warden license to rubber-stamp.

Manifest schema

id: <task-id>
created: <ISO-8601>
operator: <operator name or '@me'>
slice: <one-line description>
role: forge | quill | scout
project: <PROJECT_*.md or 'none'>
base_ref: <git ref to diff against, default: HEAD>
goal: <imperative sentence — what 'done' looks like>
done_when:
  - name: <human-readable check name>
    type: test | command | file_exists | file_contains | diff_constraint | human_review
    # type-specific fields below
    # test:
    command: <test-runner command>
    expect_pass: true
    # command:
    command: <shell command>
    expect_exit: 0
    # file_exists:
    path: <path>
    # file_contains:
    path: <path>
    pattern: <regex or literal>
    # diff_constraint:
    touches_only: [<glob>, ...]
    forbidden: [<glob>, ...]
    # human_review:
    description: <what the reviewer is confirming>
constraints:  # optional — non-negotiable rules carried from project overlay
  - <one-liner>