requirements-to-task-packet

requirements-to-task-packet

Core principle: workers need explicit task packets, not inherited ambiguity. Ask only when ambiguity changes architecture, safety, public behavior, or data loss risk; otherwise record assumptions.

Distilled from detailed reads of clarification, spec-to-implementation, subagent-driven development, and skill-authoring patterns.

Clarify or Proceed

Ask before implementation when ambiguity affects:

• security boundary, permissions, ownership, or secret handling;
• destructive operations, migrations, publishing, or public API behavior;
• architecture or data model;
• acceptance criteria or rollback expectations.

Proceed with explicit assumptions when ambiguity is local, reversible, and testable.

Task Packet Template

Objective:
Scope/paths:
Allowed edits:
Forbidden edits/non-goals:
Inputs/dependencies:
Relevant context/artifacts:
Assumptions:
Risks:
Acceptance criteria:
Verification commands:
Expected output artifacts:
Escalation conditions:

Subagent Context Rules

• Give each worker fresh, curated context; do not rely on hidden parent history.
• Include exact upstream artifact paths and summaries when needed.
• Keep implementation tasks independent or explicitly sequenced.
• Require workers to report one of: DONE, DONE_WITH_CONCERNS, NEEDS_CONTEXT, BLOCKED.
• For BLOCKED/NEEDS_CONTEXT, change context/model/scope before retrying.

Acceptance Criteria

Use observable checks:

• command output, state transition, UI/status text, artifact contents;
• regression tests or named test files;
• security properties such as containment/ownership/no secrets;
• compatibility requirements such as Windows paths or Pi CLI flags;
• rollback notes.

Spec Pairs (pi-crew v0.10.1+, ADR-6)

When the target runs pi-crew, author the SpecRecord + TaskPacket pair instead of prose-only acceptance:

1. SpecRecord (workspace state/specs/<id>.json): requirements[] with must|should|could priority + stable ids; acceptance[] entries each tied to a requirementId. Machine-checkable acceptances carry command, expectedDigest (sha-256 hex of stdout) or expectedExitCode, and idempotent: true — only idempotent commands are ever re-run.
2. PROVENANCE — critical: specs you (an agent/skill path) write are persisted generated and NEVER re-executed by the orchestrator. manual+trusted (the only specs strict mode re-runs) are minted exclusively by USER-facing import actions — a worker cannot author a command the root re-executes.
3. Wire the workflow: add specRefs: [<spec ids>] to steps held to the spec; specStrict: true in frontmatter opts the whole workflow into strict mode (requires a verifier role step — the run rejects at start otherwise).
4. Executor footer contract: workers must END results with