task-splitting（任务分割）

Task Splitting

Turn source documents into explicit task boundaries. Keep decomposition separate from creation: draft and validate first, create agent-team tasks only after explicit user approval.

Workflow

1. Intake the source document.
2. Extract goals, constraints, deliverables, non-goals, and ambiguities.
3. Decide whether brainstorming is needed.
4. Draft tasks by function boundary.
5. Validate every task boundary.
6. Write one task file per task.
7. Ask for confirmation.
8. Create agent-team task packages from the approved files.

1. Intake the source

• Accept either raw document content or a local document path.
• Read only the files needed for decomposition.
• If the document is missing the goal, deliverable, or target output location for task files, ask before drafting.
• If the user did not specify where task files should go, ask once. If they do not care, place them in a sibling directory next to the source document named <document-stem>-tasks/.
• Treat the source document plus user clarifications as the only authority for task scope.

2. Decide whether to use brainstorming

Invoke brainstorming before drafting tasks when any of these are true:

• the document goal is ambiguous
• scope conflicts exist
• several decomposition strategies are equally reasonable
• multiple functions are tangled together
• the user wants discussion before task drafting

Do not force brainstorming for a clear, well-bounded document.

3. Decompose by function boundary

Apply these rules strictly:

• Prefer one user-visible function or one clear sub-function per task.
• Keep exactly one primary responsibility per task.
• Keep task boundaries explicit.
• Prefer one function per task, not one paragraph per task.
• Reject grouping by document section when that creates mixed scope.
• Do not merge multiple functions just because the document narrates them together.

Split a task when any of these are true:

• it contains multiple peer functions
• it contains more than one acceptance target
• it mixes build and validation responsibilities
• its dependency chain makes independent execution unclear

4. Validate every task boundary

A valid task must include all of the following:

• exactly one function or sub-function
• exactly one primary responsibility
• a clear completion condition
• an explicit in-scope section
• an explicit out-of-scope section
• explicit acceptance criteria
• manageable dependencies

If a task fails any check, split or rewrite it before writing files.

5. Write one task file per task

• One task must map to one task definition file.
• Treat the task file as the authoritative boundary definition.
• Keep naming stable and reviewable, for example 01-auth-login.md, 02-auth-session-check.md.
• Use the template in references/task-file-template.md.
• Do not create a combined summary file in place of per-task files.

6. Request confirmation before creation

• Present the drafted task set for review.
• Surface ambiguities and unresolved questions.
• Ask for explicit confirmation before creating any agent-team tasks.
• If the user asks for changes, update the task files first. Do not redefine scope during the creation phase.

7. Create agent-team tasks from approved files

When the project exposes the agent-team CLI, create one task package per approved task file.

Confirmed command surface:

agent-team task create --role <role> "<title>" --design "<task-file>"

Rules:

• Do not create tasks before approval.
• Preserve title, scope, dependencies, and acceptance exactly from the task file.
• Do not reinterpret or expand scope during creation.
• --role is required. Use the role named in the document or task file; otherwise ask the user instead of guessing.
• Report exact success and failure results.
• If creation partially fails, stop and return the created tasks plus the failures explicitly.

Output contract

Draft phase output

Return:

• a short summary of the decomposition
• the list of drafted task titles
• the file path for each task file
• any unresolved boundary questions
• an explicit approval request

Creation phase output

Return:

• each created task title
• its source task file
• the created task identifier if available
• any failure with the exact file that failed

Guardrails

• Do not implement code changes.
• Do not let the creation phase rewrite scope.
• Do not merge analysis, implementation, and testing into one task.
• Do not force brainstorming when the boundaries are already clear.