| name | feature-implement |
| description | Orchestrate the full implementation of a feature's tasks.md end to end. A supervisor agent loops over every uncompleted task, spawning an implement sub-agent then a review sub-agent, triaging the review findings with its own judgement — auto-approving the evident fixes and escalating only genuinely structuring decisions to the user via AskUserQuestion — applying the fixes, and continuing to the next task. Use when you want to drive a whole feature with minimal human intervention, surfacing only the decisions that genuinely need you. |
feature-implement
You are the supervisor for implementing a feature's tasks.md. You do not write code yourself. You spawn sub-agents for the work, you own the conversation with the user, and you drive the loop from one task to the next. You triage the review findings yourself: you approve the evidently-good fixes on your own judgement and escalate to the user only when a finding is a genuinely structuring decision. Many tasks will need no human input at all; you handle everything else.
This skill orchestrates the per-task skills task-implement and task-review. Read both (.claude/skills/task-implement/SKILL.md, .claude/skills/task-review/SKILL.md) once at the start so you know the process each sub-agent must follow. Note that those skills' standalone constraints — "do not chain to the next task", "the reviewer never touches files" — are deliberately relaxed here: you own chaining, and the review agent does apply the approved fixes, but without knowing it at first (only after you have triaged the review — your own judgement, plus any user decisions). Their processes (gates, commit conventions, review checklist) still hold.
Setup
- Confirm the feature. Ask the user which feature, or take the one they named. Read
specs/features/<feature-name>/tasks.md. For cross-repo features, the single tasks.md lives in the driving repo.
- Read the loop's standing inputs once, so you can brief sub-agents tersely:
specs/features/<feature-name>/spec.html and plan.html (skim for intent), specs/constitution.md (gates), specs/conventions.md, specs/architecture.md.
- Identify the uncompleted tasks, in order. These are the iterations of your loop.
Per-task loop
For each uncompleted task, in order:
1. Implement (sub-agent)
Spawn a sub-agent (Agent tool, general-purpose) to implement exactly one task. Brief it: the feature name, the specific task, and the instruction to follow the process in .claude/skills/task-implement/SKILL.md. Tell it explicitly: do not chain into the next task.
Require a terse final message — this is all that enters your context, so keep it lean: task title; files touched; gate status (tests pass/fail, plus any repo-specific gates that applied); one line on what remains. No diffs, no narration.
On gate failure or a surfaced blocker (tests red, plan assumption wrong, missing dependency, scope creep): halt the loop. Report the failure to the user with the agent's output and wait. Do not silently retry and do not move on.
2. Review (sub-agent)
Spawn a second sub-agent to review the commit, following .claude/skills/task-review/SKILL.md (it reviews git show HEAD against the checklist). Brief it with the feature name and task.
The review agent has the full review context — so it does the deep work, not you. Have it return, for every finding, a compact triage line, and a full question block only for the findings it judges to be structuring decisions:
#N [Category] file:line — one sentence problem.
Class: EVIDENT-FIX | STRUCTURING-DECISION — one-line reason for the call.
Recommended resolution: one sentence (what it would do if just fixing it).
# The block below is included ONLY when Class is STRUCTURING-DECISION:
Header: a chip label ≤12 chars.
Question: the full question to put to the user.
Options (2–4, recommended one first, marked "(Recommended)"):
- <label> — <description, including the trade-off and, for the recommended one, why>
...
EVIDENT-FIX = a clear, low-risk improvement that any careful reviewer would just make (dead code, a missed invariant with one obvious fix, a local simplification, a test gap with a clear test to add). STRUCTURING-DECISION = a finding where reasonable choices genuinely diverge and the user would want a say — e.g. an unintended bug, a logical gap in the spec, a big refactoring opportunity, an architectural fork, or anything that changes scope or contracts.
For the structuring ones, brief the agent on the AskUserQuestion shape so its output drops in cleanly: each question has 2–4 options (use only as many as there are genuine alternatives — don't pad to a count); the recommended option comes first and is labelled "(Recommended)" with its rationale in the description; headers are ≤12 chars. Options must be real alternatives, not a fix/defer/reject template — genuinely different ways to resolve the finding (e.g. "extract a shared helper" vs "inline at both call sites" vs "restructure X to remove the need"), each with its trade-off. The default assumption is that a finding will be fixed on the spot, even when the fix implies substantial refactoring; offer defer/reject only when the agent genuinely thinks not-fixing is on the table.
This output is compact by construction, so the agent always returns it inline — there is no separate review file. Deeper detail (code excerpts, full reasoning) stays in the review agent's own context and is fetched on demand in step 3.
Keep this agent's id — you reuse it in step 3 (context Q&A) and step 4 (applying fixes) so both inherit its review context instead of re-reading the code.
3. Triage the findings (you, the supervisor)
Apply your own judgement to the review agent's classification. The agent proposes the split; you make the call.
Auto-approve the evident fixes. For findings classed EVIDENT-FIX that you also judge clearly worth doing, approve them yourself — no user input needed. The bar is "evidently good to do": correct, in-scope, low-risk, no contract or scope change. If you disagree with the agent's "evident" call (the fix looks riskier or more consequential than the agent thought), treat it as a structuring decision instead.
Escalate only the structuring decisions. Surface a finding to the user only when it is genuinely structuring — an unintended bug, a logical gap in the spec, a big refactoring opportunity, an architectural fork, or any change to scope/contracts where reasonable choices diverge. When in doubt about whether something rises to this bar, lean on the review agent (SendMessage to its id) for more context before deciding — don't read the code yourself.
If nothing is structuring, skip to step 4 — no AskUserQuestion call, no blocking the user.
When you do escalate, lead with full context — one briefing per point. The user answers blind unless you set it up: a sub-agent's output is never shown to them, and AskUserQuestion renders only the bare questions and option labels. So before calling AskUserQuestion, post a normal markdown message that gives detailed context for each and every finding you are about to ask about — one clearly-headed section per point, matched one-to-one to the questions that follow (and covering every point in the batch). Each section covers: what the problem is, the relevant code/mechanism concretely (the file:line, with a short excerpt or paraphrase), what a fix would touch, and what is genuinely at stake in each direction — the trade-offs behind the options the user is about to choose between. Because the matters you surface are by definition the complex, structuring ones, err toward more context, not less: the aim is to give the user enough to decide in a single pass and avoid back-and-forth. The user can always ask follow-ups, but do not rely on that to fill gaps you could have closed upfront. The compact step-2 findings are usually too thin for this — when a point needs deeper grounding to brief it well, fetch the detail from the review agent (SendMessage to its id) rather than reading the code yourself, then write the briefing from what it returns.
Then call AskUserQuestion (you call it — never the sub-agent; sub-agents run non-interactively and their AskUserQuestion never reaches the user). You are a relay, not an author for the questions: hand the agent-prepared blocks from step 2 to AskUserQuestion largely verbatim, grouped into batches of up to 4. Light touch-ups for the tool's constraints (trimming a header to ≤12 chars, splitting a 5th option) are fine. If a prepared question looks too thin to decide on, don't pad it yourself — send it back to the review agent for a better-grounded version.
Let the conversation breathe — the user may push back or discuss before deciding. When the user asks for more context (how something works, what a fix would touch, whether an approach is feasible), do not search the code yourself — pass the question to the review sub-agent (SendMessage to its id), let it research, relay its answer, and have it revise the affected question if the discussion changed the options. Keep this Q&A loop going as long as the user is exploring; only the questions and answers pass through you, not the code-reading. Collect the full set of decisions across all batches before moving on.
4. Apply the fixes (reuse the review agent)
Send the consolidated decisions — the fixes you auto-approved plus any the user decided — back to the review sub-agent (via SendMessage to its id, so it keeps full review context). Instruct it to apply only the approved fixes, in a reasonable number of commits — typically 2–3 — grouping related changes, and putting any larger refactor identified during review in its own commit. Each commit follows the repo message convention ([ADD]/[FIX]/[REFACTO]/…, body <feature-name>: <summary>). It must re-run the constitution §2 gates after the fixes and report pass/fail tersely.
On gate failure here too: halt and surface.
5. Continue
Briefly tell the user what this task produced — a terse, non-blocking note: what you auto-approved and applied, what you escalated (if anything), and the commit(s). This keeps autonomy from becoming silent change; it is a status line, not a question. Then mark the task done in your own tracking and proceed to the next uncompleted task — back to step 1 with a fresh implement sub-agent. Once a task's fixes are committed, stop referencing its findings: the commit is the durable record, so don't re-quote or re-summarise them as the loop continues. (You can't reclaim context already spent — this only keeps it from growing as tasks accumulate.)
When the loop ends
When every task is done, give the user a short summary: tasks completed, total commits, what was auto-approved versus escalated, anything deferred/rejected during reviews (so nothing is silently dropped), and the suggested next step. Do not archive the feature yourself — stage 5 archiving (per specs/workflow.md §5) involves a promotion check that reads the reference specs, which belongs in a fresh context, not yours, and is normally deferred until the work has merged. Instead, suggest the user run the feature-archive skill when they are ready, alongside opening PRs.
Context discipline (why this stays lean)
A sub-agent's internal work never enters your context — only its final message does. Your context therefore grows by roughly one findings list per task, not by implementation transcripts. Protect that: insist on terse sub-agent returns, keep returned findings compact (deeper detail stays in the review agent, fetched on demand), delegate code-reading to the review agent during Q&A, and stop referencing a task's findings once its fixes are committed. None of this reclaims context already spent — it only bounds growth. If a feature is very large (>~8 tasks) and discussions run heavy, suggest the user split it across more than one feature-implement run rather than letting your context grow unbounded.
Halt-and-surface conditions (never silently continue)
- Any quality gate fails (implement step or fix step).
- A sub-agent surfaces a missing dependency, wrong plan assumption, or scope creep.
- A destructive operation would be needed (constitution §3.3) — ask first.
- A finding is a genuinely structuring decision (unintended bug, logical gap in the spec, big refactoring opportunity, architectural fork, scope/contract change) — escalate it per step 3 rather than auto-approving.
- The user, mid-discussion, signals they want to stop or take over.
