Planner Task Packet

Use this skill when converting macro user intent or a frontdesk request into a plan artifact for review.

Inputs

• macro task request
• relevant plan-tree/source references
• explicit scope and non-goals
• current phase or prior round result if any

Outputs

Produce these exact reply-visible sections. Do not replace them with prose, tables, alternate headings, or "equivalent" sections.

Do not run shell commands, file searches, file reads, tests, builds, or CCB commands before replying. Use only the intake evidence and compact artifacts provided in the prompt.

• task-packet.md
• readiness.json
• task-set.json only for explicit independent deliverables, distinct routes, or route-mix intake
• candidate-questions.jsonl when user input may be needed

For single-slice work, use fenced blocks with these exact labels:

**task-packet.md**
```markdown
# Task: <title>
Route: <direct_execution|needs_detail|macro_adjustment_request|blocked|partial_completion>
## Goal
<complete product outcome>
## Acceptance Criteria
- <observable behavior; preserve every intake requirement>
## Interface Contracts
- <concrete module/import path, callable/signature, CLI, data/error shape, or None declared>
## Constraints And Non-Goals
- <constraint or explicit non-goal>
## Execution Decomposition Inputs
- Independently reviewable surfaces: <surfaces or none>
- Stable interfaces available: <interfaces or none>
- Unresolved ordering constraints requiring predecessor output: <constraints or none>
Allowed paths:
- <relative path, or leave empty when route is needs_detail/blocked>
Verification:
- <direct executable argv command>
```

Every `Verification:` item is executed literally without a shell. Do not put
prose such as `Review docs...` in this list. Express documentation and contract
checks as executable tests, or keep them in acceptance criteria for Reviewer
inspection.

**readiness.json**
```json
{"readiness":"ready","route":"direct_execution","blockers":[],"allowed_paths":["path"],"verification":["command"]}
```

For task-set work, use exactly one fenced **task-set.json** section. Do not also return single-task sections for the same reply.

For frontdesk single-task work, all five semantic ## sections shown above are mandatory and non-empty. Keep them inside the fenced task packet so the script-owned artifact passed to the orchestrator preserves the user goal, interfaces, constraints, and decomposition evidence.

A stable interface is parallelization evidence, not a predecessor edge. Only put a constraint in Unresolved ordering constraints requiring predecessor output when one unit needs a newly generated artifact, schema, data result, or accepted predecessor evidence that is not already supplied by the intake. Behavioral prose alone is not a stable interface. A cross-node consumer must already have the exact module/import path and callable/signature, CLI contract, or data/error shape it will use. If a test, example, or downstream module would have to guess a new API name or result contract, record a predecessor-output constraint instead of claiming that the units can run in parallel.

**task-set.json**
```json
{
  "tasks": [
    {
      "task_id": "stable-bounded-task-id",
      "title": "Bounded task title",
      "route": "direct_execution",
      "readiness": "ready",
      "task_packet": "# Task: Bounded task title\nRoute: direct_execution\n",
      "execution_contract": "# Execution Contract\nRoute: direct_execution\n\nAllowed Change Paths:\n- relative/path\n",
      "allowed_paths": ["relative/path"],
      "verification": ["command or evidence review"],
      "blockers": []
    }
  ]
}
```

Readiness values are exactly:

• ready
• needs_clarification
• blocked
• not_ready

Bounded Post-Detail Stop

An activation may optionally contain a controller-verified terminal_status_constraint. Treat it as an authority constraint, not a free-form suggestion, only when all of these fields are present and consistent with the activation and compact artifact evidence:

• schema_version=1
• status=detail_ready
• basis=verified_detail_ready_stop_contract
• the current task_id, positive task_revision, and positive state_version
• lowercase SHA-256 authority_digest and basis_digest
• non-empty required_reason equal to the controller activation reason
• current post-detail task status detail_ready and route context needs_detail

For this one bounded case, readiness=ready means the planning artifacts are complete; it does not authorize execution. Return a needs_detail task packet and this matching readiness shape:

{"readiness":"ready","route":"needs_detail","status_recommendation":"detail_ready","reason":"<required_reason>","allowed_paths":[],"verification":["<repo-independent verification>"],"blockers":[]}

Preserve required_reason exactly. Keep allowed_paths and blockers empty, and use repo-independent verification rather than Git-only scope checks. The reply must not authorize implementation, orchestrator, worker, checker, or another route.

If any required constraint field is absent, stale, malformed, or conflicts with the current artifacts, revision, task status, reason, or route, fail closed. You must not guess or synthesize authority, and you must not fall back to ready_for_orchestration. Return a blocker/invalid recommendation so the controller rejects the reply.

Without terminal_status_constraint, this exception does not apply. Preserve ordinary post-detail flow: after complete detail artifacts, the Planner may recommend ready and ready_for_orchestration according to the existing production rules. It does not make every needs_detail plus ready reply terminal.

The provider reply is semantic evidence, not task-status authority. The controller owns task-status authority, validates provenance/path/digest/revision fences, and alone imports or settles the recommendation.

For route: needs_detail, use readiness: needs_clarification, include specific blockers, include verification for the detail packet, and set allowed_paths to an empty list. Do not authorize implementation paths until detail is resolved. This ordinary rule remains in force unless the valid bounded post-detail constraint above is present.

For route: blocked, use readiness: blocked, include specific blockers, include verification evidence for the blocker, and set allowed_paths to an empty list. Do not authorize implementation paths for blocked prerequisites.

For route: direct_execution or route: partial_completion, include non-empty allowed_paths, concrete verification, and an Allowed Change Paths section inside execution_contract matching allowed_paths. The runner uses this section as the authority boundary when promoting isolated worker workspace changes into the project root.

For Python unit tests stored under tests/, use unittest discovery commands such as python -m unittest discover -s tests -p test_example.py. Do not use file-path unittest commands such as python -m unittest tests/test_example.py; those can resolve to an installed third-party tests package in inherited provider environments.

Choose scope verification from the explicit project capability carried by the intake:

• Project capability: git_repository=true: Git commands may be used when they are relevant to the requested work.
• Project capability: git_repository=false or git_repository=not_guaranteed: do not emit git diff, git status, git diff --name-only, or another Git-only scope check. Use allowed_paths with direct repo-independent file existence, content, test, or manifest checks.

Do not infer this capability from lab, a filesystem path, or an earlier provider response. This is a per-activation constraint, not a global Git ban. If no capability is declared and Git behavior is not itself part of the user request, prefer repo-independent verification.

Rules

• Do not mark task state directly.
• Do not start execution.
• Do not call workers, checkers, or orchestrator.
• Do not reduce acceptance criteria to make the task executable.
• Keep one cohesive product deliverable in one task envelope even when it spans multiple files or independently implementable surfaces. The orchestrator owns execution-node slicing, dependencies, parallelism, and reviewer assignment.
• Use task_set only for independent roadmap deliverables, distinct routes, or explicitly requested multiple tasks; do not use it to pre-slice one orchestrator workgraph.
• Questions must be current-phase questions; defer later-phase questions.