plan

Navigator Planning Skill

---

navigator is one skill, plan. Behavior adapts to what's asked, dispatching to whichever agent below fits the request — there is no separate navigator/estimate or navigator/challenge skill to invoke; estimating and challenging both happen inside this one skill.

• Decompose a spec into a plan — planner reads a spec@1, produces an ordered sequence of implementation tasks (exact file paths, a failing test, minimal implementation, expected test output, a commit message, covers_criteria), and groups tasks into cohesive Subsystem Batches aligned with crate/package compilation boundaries as part of that decomposition — not a separate step. Reads the spec from disk at spec_file_path when set, and propagates spec_file_path, spec_hash, and linked_requirement into the plan@1 output. Also runs in amend mode: given an existing plan@1, a corrected spec@1, and the criterion_ids that changed, patches only the affected tasks instead of re-decomposing the whole plan.

• Estimate an existing plan — estimator assigns effort estimates (in minutes) to each task in a plan@1, identifies parallelizable tasks, and surfaces blocking dependencies.

• Challenge a draft plan — challenger adversarially reviews a plan@1 for missing tasks, wrong ordering, under-specified steps, over-sized tasks, missing error handling, and acceptance criteria orphaned from every task's covers_criteria. An amended plan (from planner's amend mode) is not exempt from this review.

---

<subagent_dispatch_matrix>

Agent	Role	Model / Effort	Delegate When
planner	Plan author	sonnet / medium	Decompose a spec into ordered, subsystem-batched implementation tasks.
estimator	Effort estimator	sonnet / medium	Assign time estimates, identify parallelizable tasks, and surface blocking dependencies.
challenger	Adversarial reviewer	sonnet / medium	Stress-test a draft plan@1 (max 2 review rounds) before any implementation begins.

</subagent_dispatch_matrix>

---

<artifact_contracts>

Consumes: spec@1 (or its spec_file_path, read from disk when set) Produces: plan@1 — carries spec_file_path, spec_hash, and linked_requirement propagated from the source spec

</artifact_contracts>

---

<tdd_requirement>

Every task in a plan@1 follows this exact sequence:

1. Write the failing test
2. Run the test — confirm it fails with the expected error
3. Write the minimal implementation to make the test pass
4. Run the test — confirm it passes
5. Commit with the conventional commit message

A task that starts with implementation before a failing test violates this requirement.

</tdd_requirement>

---

<no_placeholders_rule>

Every step must show exact code. "Implement the feature", "add appropriate error handling", and "write tests as needed" are not steps — they are abdications. If the exact code is not known, it must be derived from the spec and documented as a concrete choice, not deferred as a TBD.

</no_placeholders_rule>