agent-task-splitter (fan-out planner)

Plan a fan-out before launching it. This skill plans; the host launches the lanes. The name stays agent-task-splitter so existing prompts and plans keep working.

Pick the mode

Mode	Use when	Artifacts
Native (default)	All lanes run inside one host session: parallel or background subagents, workflow lanes, one-shot CLI delegates, with the user present	Lane table in the conversation and one brief per lane. No .coord/ plan, packets, or checkpoint.
Harness	A round spans sessions or tools, an unattended loop runs, work must resume after an interruption, a signed human gate is needed, or AGENT_COLLAB_POLICY is configured	.coord/plan.yml (schema v2) and .ai/task_<NNN>_<slug>.md packets; see references/harness-plan.md

A configured policy always selects harness mode, and an unreadable configured policy fails closed. Native mode follows the host's own limits and permissions and adds no numeric limits of its own. Host mechanics are in native lanes.

Fan out only when

• two or more pieces are independent (none needs another's result) and each has a checkable result; or
• implementation and independent review must be separate lanes.

Do not split exploratory debugging before the cause is known, a small coherent change, or work where every lane needs the same context in tight sequence. Doing it directly is the default.

The lane table

Write it before launching anything:

Lane	Role	Executor	Model tier	May write	Must not touch	Acceptance	Return	Depends on
L1	delegated-executor	subagent in a worktree	cheap	src/a/**	everything else	pytest tests/a -q	summary + paths	none
L2	delegated-executor	CLI delegate	cheap	docs/x.md	src/**	link check passes	summary + paths	none
L3	reviewer	read-only subagent	strong	nothing	**/*	verdict with evidence	verdict + findings	L1, L2

Rules:

1. Roles are primary-agent, delegated-executor, reviewer, and synthesizer, defined in the public contract. The executor and model columns are host-adapter choices, not public plan fields.
2. The orchestrator assigns the model tier; a lane never picks its own. Mechanical, pattern-following work goes to a cheap tier. Review, verdicts, root-cause analysis, security, governance, and anything ambiguous go to a strong tier.
3. Write scopes are disjoint. Two parallel writers never own the same file. If scopes might touch, isolate the writers or run them in sequence.
4. Every lane has at least one runnable or objectively checkable acceptance criterion.
5. Every lane has a capped return: a short summary plus file paths, never raw logs or full diffs. agent-context-budget covers sizing.
6. A lane cannot approve its own output. Author and reviewer are different lanes.

references/task_splitter_heuristics.md helps when a task could fit more than one role or DAG shape.

Fix the merge step before launch

Write down, before any lane starts:

• who merges the results (normally the primary-agent) and in what order;
• the verification command run on the merged result, not only per lane;
• which agent-acceptance-gate presets the merged diff will trigger;
• commit boundaries: one reviewed commit per writing lane, or one commit after the merged result passes. Read-only lanes need no commit.

Brief per lane

# Lane <id>: <one bounded deliverable>
Repo/worktree: <absolute path>
Role: <role>. Model tier: <tier>.
Depends on: <lane ids and artifact paths>

Before editing, list the exact paths you may write and must not touch.
Stop and report if they conflict with this brief.

May read: <paths>
May write: <paths>
Must not touch: <paths>
External actions: none unless named here.
Edits required: yes | unchanged output allowed because <reason>

Acceptance: <runnable check>
Return (capped): status; short summary; files changed; checks run and
results; risks; blockers. Put long output in a file and return its path.

Drift guards

• Copy identifiers (slugs, paths, field names, version strings) verbatim into the brief. Do not paraphrase them.
• Record the pre-lane baseline (commit or file hashes) so the result can be attributed. File times and a process exit status do not prove the work was done; use the task-content evidence review.
• State whether edits are required. A lane that reports success without changing its required files has failed unless the brief allowed unchanged output.
• Missing, null, failed, declined, cancelled, and timed-out lanes stay non-success. Filter them out of synthesis but keep their records.
• Agent agreement never replaces running the checks.

Harness mode

1. Write .coord/plan.yml (schema_version 2) and one packet per non-inline task, using references/harness-plan.md.
2. Immediately before each spawn the host runs agent-collab policy evaluate --policy <policy_ref> --checkpoint <checkpoint_ref> --json and spawns only on decision=continue with spawn_allowed=true. The plan never overrides that decision.
3. Coordination files are scratch and gitignored. Promote only an explicit checkpoint snapshot, shipping artifact, or acceptance record.
4. Do not encode retries in the plan; the policy decides them.

Historical schema-less plans and provider-specific task paths are parse-only; see migration 0.4.

Output

Return the lane table, each brief (or its path), the ready lanes in launch order, and the merge-and-verify step. Do not launch lanes from this skill and do not turn a planning-only request into implementation.