Spec Write Tasks

spec-write-tasks

spec-write-tasks is an optional derived layer between spec-plan and spec-work. It does not execute code. It either compiles a settled source plan into a derived task pack, validates an existing task pack, or returns a no-task-pack decision.

It is a public workflow with the unified entrypoint spec-write-tasks.

Purpose

Decide whether a derived task pack is worthwhile, compile one when it reduces execution risk or context load, and validate existing task packs before they reach spec-work.

Workflow Contract Summary

When To Use

Use when a settled local source plan is too large or dependent for direct execution, when the user explicitly asks to split a plan into tasks, or when an existing local task pack needs validation before spec-work.

When Not To Use

Do not use for unresolved product or architecture scope, small changes that do not need a derived layer, remote repositories or package names, marketplace skill lookup, or implementation prompts that should go directly to spec-work.

Inputs

A local source plan path, existing task-pack path, or clear task-splitting request that resolves to one local source plan. Inputs must stay local and source-owned; remote repos, packages, marketplace identifiers, and generic task lists are rejected.

Outputs

One of: executable derived task pack, validated existing task-pack handoff, skip, return-to-plan, draft-only, or validate-only envelope. Task packs are derived execution indexes and never replace the source plan.

Artifacts

Executable task packs live under docs/tasks/ and must carry an artifact-root-relative POSIX source_plan, canonical source_plan_hash, generated_by: spec-write-tasks, mode: derived, and a valid Task Pack Contract JSON block. Executable identity is source-plan-path+body-hash; spec_id is an optional compatibility trace rather than a required identity field.

Failure Modes

Use machine-readable reason codes: source_plan_missing, ambiguous_plan, wrong_chain, stale_hash, unverifiable_hash, invalid_contract, repo_scope_missing, or scope_gap.

Workflow

Classify input, verify plan identity and repo scope, choose one branch (compile, skip, return-to-plan, draft-only, validate-only), use bounded source orientation only when it improves task boundaries, write or validate the task pack, then return the final decision envelope.

Downstream Consumers

spec-work, high-risk spec-doc-review handoff, human reviewers, and later code-review/compound workflows.

Scenario Capability

Follows docs/contracts/workflows/scenario-capability-matrix.md (default).

Overrides: none

Core Rules

1. spec-plan is always the single source of truth.
2. A task pack may rearrange execution slices, but must not change scope, acceptance criteria, non-goals, repo ownership, or product decisions.
3. A task pack is not progress state, approval state, a lifecycle database, or a second plan.
4. context_refs are bounded reading pointers, not scope authority. Whole-plan or whole-directory refs are low-quality unless paired with narrower anchors.
5. If parent-workspace scope is present, the plan or task pack must name target_repo at top level or per task. Missing repo scope returns to spec-plan.
6. Source orientation is advisory and bounded: source plan sections, plan-declared files, nearby tests, and precise local contracts only when they improve task boundaries.
7. Scripts validate identity, freshness, structure, hashes, concrete paths, and same-wave overlap. LLM/reviewers judge semantic task quality.
8. Do not hand-edit .claude/, .codex/, or .agents/skills/ as source fixes.
9. --repo <artifact-root> selects the artifact/source resolution root only. It does not authorize or select the downstream mutation target_repo.

Input Paths

• Source plan path: read the source plan frontmatter and focused sections: Requirements, Scope Boundaries, Technical Approach, Implementation Units, Files, Test Scenarios, Verification, and Deferred to Implementation.
• Existing task-pack path: read the task pack, source plan, and Task Pack Contract; validate identity/freshness/structure before treating it as executable.
• Bare task-splitting request: resolve exactly one local source plan; if resolution is ambiguous, ask for the path.

Use Task Quality Guide for accepted/rejected input classes, task-ready source plan rules, source orientation, traceability, vertical slices, context_refs, done_signal, stop_if, and review-intent quality guidance.

Branches

• compile: plan is task-ready and a task pack reduces execution risk or context load.
• skip: direct spec-work is cheaper and safer than adding a derived layer.
• return-to-plan: scope, acceptance, architecture, verification, or repo ownership is unresolved.
• draft-only: a non-executable outline may help discussion but cannot be handed to spec-work.
• validate-only: an existing task pack is checked before execution.

Use Execution Handoff Contract for branch outputs, final envelope fields, high-risk review handoff, hash rules, and failure reason codes.

Task Pack Output

When compiling, read Task Pack Schema and write the task pack body with:

• Overview
• Source Summary
• Traceability Matrix
• Task Graph
• Execution Waves
• Task Pack Contract
• Task Cards
• Orientation Evidence
• Validation Notes
• Regeneration Rules

The Task Pack Contract fenced JSON block is the machine-readable source for validators. Human-readable sections compress context for executors and reviewers, but deterministic validation proves only identity, freshness, and structure.

Final Decision Envelope

Every run must end with the compact envelope from Execution Handoff Contract.

Required posture:

• include decision, reason_code, identity_basis, artifact_root, source_plan, task_pack, task_pack_validity, deterministic_handoff, validity_scope, semantic_posture, dispatch_authorization, validation, orientation, and next_action;
• 🔴 GATE (verification): run spec-first tasks validate <task-pack-path> --repo <artifact-root> --json before reporting deterministic_handoff or validation fields;
• run spec-first tasks hash <plan-path> --repo <artifact-root> --json when computing or comparing source_plan_hash, and copy its portable source_plan field rather than deriving identity from an absolute machine path;
• 🔴 GATE (verification): never self-report deterministic_handoff: true without CLI JSON evidence;
• 🔴 GATE (handoff): allow next_action: spec-work-task-pack only for valid deterministic handoff plus semantic posture generated-this-run or reviewed-existing;
• 🔴 GATE (verification): for semantic_posture: reviewed-existing, the envelope must carry evidence metadata or a verifiable review-outcome reference; a bare reviewed-existing claim without current evidence is not eligible for next_action: spec-work-task-pack;
• 🔴 GATE (handoff): for dispatch_authorization: authorized, the envelope must carry a bounded continuation reference or doc-review outcome reference; absent that, report dispatch_authorization: missing.

High-risk packs return next_action: review-task-pack with one reason and the copy-ready current-host invocation spec-doc-review mode:headless mutation:report-only output:json roster:full <task-pack-path>。Consumer 必须把 derived task pack 保持 byte-preserving，并从 task_pack_outcome 读取 deterministic intake、semantic review 与 terminal owner；不要仅凭 Review complete 推导 reviewed-existing。roster:full 仍只保留实际 qualified personas，但不会让 high-risk review 因 standard budget 丢失同时成立的 security、scope 或 adversarial lens。Do not auto-dispatch review unless the explicit bounded continuation conditions in Execution Handoff Contract are met.

Portability Boundary

Runtime references must be packaged skill files (references/, agents/) or optional named integration points distributed with spec-first. spec-plan, spec-work, and spec-doc-review are integration names, not required files to read from this package.

evals/ files are maintainer-only validation fixtures; the official .skill packager excludes root evals/, so runtime users must not need them. Repo-local scripts, validation reports, technical plans, project-role documents, historical snapshots, and .spec-first/audits/** are maintainer evidence, not user-runtime dependencies.

Do Not

• Do not execute implementation work.
• Do not generate task packs for small low-risk plans by default.
• Do not invent missing scope, acceptance criteria, repo targets, or product decisions.
• Do not turn review_gate into approval state or validator-owned semantic risk.
• Do not make task quality analyzer warnings a hard validator gate.
• Do not list repo-local validation reports or historical plans as runtime references.

References

• Execution Handoff Contract
• Task Pack Schema
• Task Quality Guide