compose:subagent

Subagent-Driven Development

Execute plan by dispatching fresh subagent per task, with two-stage review after each: spec compliance review first, then code quality review.

Why subagents: You delegate tasks to specialized agents with isolated context. By precisely crafting their instructions and context, you ensure they stay focused and succeed at their task. They should never inherit your session's context or history — you construct exactly what they need. This also preserves your own context for coordination work.

Core principle: Fresh subagent per task + two-stage review (spec then quality) = high quality, fast iteration

Continuous execution: Do not pause to check in with your human partner between tasks. Execute all tasks from the plan without stopping. The only reasons to stop are: BLOCKED status you cannot resolve, ambiguity that genuinely prevents progress, or all tasks complete. "Should I continue?" prompts and progress summaries waste their time — they asked you to execute the plan, so execute it. When you must stop for ambiguity or a blocker, use compose:ask to present the situation with structured options. If no user is available, resolve it with your best judgment and continue.

When to Use

1. Have implementation plan? — No → brainstorm first or manual execution
2. Tasks mostly independent? — No (tightly coupled) → brainstorm first or manual execution
3. Stay in this session? — Yes → compose:subagent / No → compose:execute (parallel session)

vs. Executing Plans (parallel session):

• Same session (no context switch)
• Fresh subagent per task (no context pollution)
• Two-stage review after each task: spec compliance first, then code quality
• Faster iteration (no human-in-loop between tasks)

The Process

Setup: Read plan → extract all tasks with full text → note context → create a task per plan task (task create)

Per Task:

• Dispatch implementer — create+bind task (--task <TID>, auto-starts), inject covered spec as Intent, dispatch (./implementer-prompt.md)
• Implementer asks questions? — Yes → answer, provide context, re-dispatch implementer
• Implementer implements — tests, commits, self-reviews
• Spec review phase 1 — dispatch with spec + diff only, no report
• Phase 1 flagged anything? — Yes → dispatch phase 2 (report explains flags, downgrade-only)
• Spec gate: all in-scope claims pass with evidence? — No → implementer fixes → back to spec review phase 1
• Code quality review — dispatch reviewer (./code-quality-reviewer-prompt.md)
• Quality approved? — No → implementer fixes → back to code quality review
• Mark task done — task done <TID>

More tasks remain? — Yes → next task (dispatch implementer)

Finish: Dispatch final code reviewer for entire implementation → compose:merge

Spec-Anchored Review Gate

This is how each task's spec-compliance review works. It replaces a single prose review with intent-grounded implementation and a two-phase, evidence-gated verdict.

Working directory (only if isolated). Subagents run with fresh context and start in the repo root. If compose:worktree set up an isolated worktree, the work does NOT live there — tell every subagent to cd into the worktree first (fill the implementer's Work from: line, and add the cd note the reviewer prompts describe). Otherwise there's nothing to do: the current checkout is correct and you can leave those lines out. When a worktree IS in use and you forget, a reviewer's git diff and tests run against the wrong checkout and the verdict is meaningless.

1. Create and bind a task before dispatching. Before the implementer runs, create a work-item with the task tool (task create "<plan task summary>") and capture its TID. Dispatch the implementer bound to that task by passing --task <TID> on the actor call (actor run general "<desc>" "<prompt>" --task T3). Binding auto-starts the task to in_progress with the subagent as owner, and lets the postStop hook validate the task's progress. Do NOT mark the task done here — completion is gated on review (step 4).

2. Inject intent before dispatching the implementer. Read the task's Covers: field, pull the verbatim text of those [Sn] spec sections, and paste it into the implementer prompt's ## Intent (from spec) block (see ./implementer-prompt.md). The implementer never reads the spec itself — you hand it exactly the sections its task covers, with the scope boundary intact.

3. Run the spec reviewer in two phases (see ./spec-reviewer-prompt.md). If an isolated worktree is in use, add the cd <worktree> instruction to each dispatch (see the working-directory note above):

• Phase 1: dispatch with the covered spec section text + git diff ONLY. Do NOT include the implementer's report — its claims anchor the reviewer toward confirming what was reported and away from spotting silent omissions. Phase 1 returns a structured per-claim verdict.
• Phase 2: only if phase 1 flagged anything. Re-dispatch the same reviewer with its phase-1 verdict + the implementer's report, solely to let the report explain flagged diffs. Phase 2 may downgrade a flagged item; it cannot add passes.

4. Gate on the verdict. The task is complete ONLY when the final verdict is Status: pass AND every in-scope claim is status: pass with evidence. Any fail or unverifiable in-scope claim → re-dispatch the implementer with the specific failing claims, then re-review. Loop until the gate passes. Then run the code quality review (spec compliance always precedes quality), and once that also passes, mark the bound task done with task done <TID>.

A structured pass without verifiable evidence (test name, command output, or file:line) does not satisfy the gate — treat it as fail. Prose is not evidence.

Pre-Flight Plan Review

Before dispatching Task 1, scan the entire plan once for conflicts:

• Tasks that contradict each other or the plan's Global Constraints
• Anything the plan explicitly mandates that a reviewer would flag as a defect (e.g. a test that asserts nothing, verbatim duplication of a logic block)
• Ambiguous or inconsistent interface contracts between tasks

Present everything you find to your human partner as one batched question — each finding beside the plan text that mandates it, asking which governs — before execution begins. Do not interrupt mid-plan with one finding at a time. If the scan is clean, proceed without comment.

Use compose:ask for the batched question. If no user is available, resolve contradictions conservatively (strictest interpretation) and continue.

Model Selection

Use the least powerful model that can handle each role to conserve cost and increase speed.

Mechanical implementation tasks (isolated functions, clear specs, 1-2 files): use a fast, cheap model. Most implementation tasks are mechanical when the plan is well-specified.

Integration and judgment tasks (multi-file coordination, pattern matching, debugging): use a standard model.

Architecture, design, and review tasks: use the most capable available model.

Task complexity signals:

• Touches 1-2 files with a complete spec → cheap model
• Touches multiple files with integration concerns → standard model
• Requires design judgment or broad codebase understanding → most capable model

Reviewer tier: Dispatch the spec reviewer at a model tier at least as capable as the implementer's. A reviewer weaker than the implementer shares its blind spots and rubber-stamps the same misreadings; the adversarial value of review comes from the reviewer interpreting the spec independently, which a weaker model cannot reliably do.

Handling Implementer Status

Implementer subagents report one of four statuses. Handle each appropriately:

DONE: Proceed to spec compliance review.

DONE_WITH_CONCERNS: The implementer completed the work but flagged doubts. Read the concerns before proceeding. If the concerns are about correctness or scope, address them before review. If they're observations (e.g., "this file is getting large"), note them and proceed to review.

NEEDS_CONTEXT: The implementer needs information that wasn't provided. Provide the missing context and re-dispatch.

BLOCKED: The implementer cannot complete the task. Assess the blocker:

1. If it's a context problem, provide more context and re-dispatch with the same model
2. If the task requires more reasoning, re-dispatch with a more capable model
3. If the task is too large, break it into smaller pieces
4. If the plan itself is wrong, escalate to the human

Never ignore an escalation or force the same model to retry without changes. If the implementer said it's stuck, something needs to change.

Prompt Templates

• ./implementer-prompt.md - Dispatch implementer subagent
• ./spec-reviewer-prompt.md - Dispatch spec compliance reviewer (two-phase, evidence-gated verdict)
• ./code-quality-reviewer-prompt.md - Dispatch code quality reviewer subagent

Example Workflow

You: I'm using Subagent-Driven Development to execute this plan.

[Read plan file once: plans/feature-plan.md]
[Extract all 5 tasks with full text and context]
[Create a task per plan task with `task create`, capturing each TID]

Task 1: Hook installation script

[Get Task 1 text and context (already extracted)]
[Dispatch implementation subagent with full task text + context, bound to its task: actor run general "..." "..." --task T1 — binding auto-starts T1 to in_progress]

Implementer: "Before I begin - should the hook be installed at user or system level?"

You: "User level (~/.config/compose/hooks/)"

Implementer: "Got it. Implementing now..."
[Later] Implementer:
  - Implemented install-hook command
  - Added tests, 5/5 passing
  - Self-review: Found I missed --force flag, added it
  - Committed

[Inject covered spec [Sn] sections as Intent; implementer builds]
[Dispatch spec reviewer phase 1: spec sections + diff only, no report]
Spec reviewer (phase 1):
  Status: pass
  Claims: [S1 · "install at user level"] in-scope · pass — evidence: test "installs to ~/.config" 5/5
[Phase 1 all-pass → skip phase 2; gate passes]

[Dispatch code quality reviewer]
Code reviewer: Strengths: Good test coverage, clean. Issues: None. Approved.

[Gate passed → mark Task 1 done: task done T1]

Task 2: Recovery modes

[Get Task 2 text and context (already extracted)]
[Dispatch implementation subagent with full task text + context]

Implementer: [No questions, proceeds]
Implementer:
  - Added verify/repair modes
  - 8/8 tests passing
  - Self-review: All good
  - Committed

[Dispatch spec reviewer phase 1: spec sections + diff only, no report]
Spec reviewer (phase 1):
  Status: fail
  Claims:
    - [S4 · "report every 100 items"] in-scope · fail — evidence: no progress code in diff (omission)
    - [S4 · "verify/repair modes"] in-scope · pass — evidence: test "repairs index" 8/8
  Extra work: --json flag (no covered claim requires it)
[Phase 1 flagged items → dispatch phase 2 with report]
Spec reviewer (phase 2): report gives no justification for --json; progress still missing — Status: fail
[Gate blocks: in-scope fail]

[Implementer fixes: removed --json, added progress reporting]
[Re-dispatch phase 1]
Spec reviewer (phase 1): Status: pass — all in-scope claims pass with evidence
[Gate passes]

[Dispatch code quality reviewer]
Code reviewer: Strengths: Solid. Issues (Important): Magic number (100)

[Implementer fixes]
Implementer: Extracted PROGRESS_INTERVAL constant

[Code reviewer reviews again]
Code reviewer: ✅ Approved

[Gate passed → mark Task 2 done: task done T2]

...

[After all tasks]
[Dispatch final code-reviewer]
Final reviewer: All requirements met, ready to merge

Done!

Advantages

vs. Manual execution:

• Subagents follow TDD naturally
• Fresh context per task (no confusion)
• Parallel-safe (subagents don't interfere)
• Subagent can ask questions (before AND during work)

vs. Executing Plans:

• Same session (no handoff)
• Continuous progress (no waiting)
• Review checkpoints automatic

Efficiency gains:

• No file reading overhead (controller provides full text)
• Controller curates exactly what context is needed
• Subagent gets complete information upfront
• Questions surfaced before work begins (not after)

Quality gates:

• Self-review catches issues before handoff
• Two-stage review: spec compliance, then code quality
• Review loops ensure fixes actually work
• Spec compliance prevents over/under-building
• Code quality ensures implementation is well-built

Cost:

• More subagent invocations (implementer + 2 reviewers per task)
• Controller does more prep work (extracting all tasks upfront)
• Review loops add iterations
• But catches issues early (cheaper than debugging later)

Red Flags

Never:

• Start implementation on main/master branch without explicit user consent
• Skip reviews (spec compliance OR code quality)
• Proceed with unfixed issues
• Dispatch multiple implementation subagents in parallel (conflicts)
• Make subagent read plan file (provide full text instead)
• Skip scene-setting context (subagent needs to understand where task fits)
• Ignore subagent questions (answer before letting them proceed)
• Accept "close enough" on spec compliance (spec reviewer found issues = not done)
• Skip review loops (reviewer found issues = implementer fixes = review again)
• Let implementer self-review replace actual review (both are needed)
• Start code quality review before spec compliance is ✅ (wrong order)
• Move to next task while either review has open issues
• Pass the spec gate on a status: pass that has no verifiable evidence (test/exec/file:line)
• Include the implementer's report in the phase-1 spec review context
• Mark a task complete while any in-scope claim is fail or unverifiable

If subagent asks questions:

• Answer clearly and completely
• Provide additional context if needed
• Don't rush them into implementation

If reviewer finds issues:

• Implementer (same subagent) fixes them
• Reviewer reviews again
• Repeat until approved
• Don't skip the re-review

If subagent fails task:

• Dispatch fix subagent with specific instructions
• Don't try to fix manually (context pollution)

Integration

Required workflow skills:

• compose:worktree - Ensures isolated workspace (creates one or verifies existing)
• compose:plan - Creates the plan this skill executes
• compose:review - Code review template for reviewer subagents
• compose:merge - Complete development after all tasks

Subagents should use:

• compose:tdd - Distill TDD workflow into subagent prompts (write test first, then implement)

Subagent guidance:

Subagents are leaf workers that execute specific instructions. When a subagent needs to follow a compose skill's workflow (e.g., TDD), distill the relevant guidance into concrete instructions in its prompt.

Alternative workflow:

• compose:execute - Use for parallel session instead of same-session execution