feature-plan

name	feature-plan
description	Research (conditionally) and decompose an approved feature into per-task files. Second step of the full feature loop. Supports --thorough for a fuller research pass on unfamiliar or high-risk work.
disable-model-invocation	true
allowed-tools	Bash, Agent, Read

<runtime_note> Explicit-invocation only (/feature-plan). Never auto-triggers. Accepts an optional --thorough flag. </runtime_note>

Turn CONTEXT.md into a set of small, dependency-ordered task files that an executor can pick up with no further ambiguity. This is the step where ambiguity is most expensive — an unclear task produces an executor that guesses, and two executors guessing differently about the same concern produce conflicts.

Shared conventions — directory layout, file ownership, task statuses, in-flight definition, feature numbering, branch and commit policy — live in references/context-contract.md in the feature skill's directory (installed at ~/.claude/skills/feature/references/context-contract.md); follow it rather than re-deriving.

1. Setup check

if [ ! -f .context/features/NNN-slug/CONTEXT.md ]; then
  echo "NO_CONTEXT — run /feature-discuss first"
fi

Stop if CONTEXT.md doesn't exist. Don't guess at decisions that step is supposed to have already captured.

2. Spawn the planner

Dispatch the feature-planner subagent (defined in ~/.claude/agents/feature-planner.md) with a fresh context window containing: CONTEXT.md, .context/OVERVIEW.md, and whether --thorough was passed.

Do not read the codebase or research anything yourself in this conversation — that's the planner's job, in its own isolated context, so this session's context budget isn't spent on it.

Without --thorough: the planner decides for itself, in the same pass, whether it needs to investigate anything before decomposing. If the domain is already well-understood from the codebase and CONTEXT.md, it skips straight to writing task files. If not, it does the minimum research needed inline and records only what mattered in RESEARCH.md — this collapses what would otherwise be a separate researcher round-trip into the one planning pass, for the common case.

With --thorough: the planner always does a dedicated research pass first, writes a fuller RESEARCH.md, and only then decomposes. Reach for this on unfamiliar libraries, new integrations, or anything where getting the ecosystem wrong would be expensive to unwind later.

3. Review the output

The planner returns: whether it researched anything, and the list of task files it wrote under .context/features/NNN-slug/tasks/. Skim the task list yourself for one thing specifically: do any two tasks touch the same file without a dependency between them? That's a conflict waiting to happen — send it back to the planner to resolve before moving on.

4. Determine parallelization shape

Run orchestrator's Diagnose and Decompose steps against the task list that just came back:

• Independent tasks with non-overlapping file ownership → subagents, grouped into waves by the Depends on field
• Tightly sequential, small task count → may be fine as a single session
• Only escalate to an agent team if tasks genuinely need to message each other mid-task — this should be rare for a single feature

State which shape applies and why, in a sentence, before handing off.

5. Hand off

Tell the user the next command is /feature-execute.