assemble skill

Assemble: Subagent-Driven Execution

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

Core Principle

FRESH SUBAGENT PER TASK + TWO-STAGE REVIEW = HIGH QUALITY.
Never skip reviews. Never reuse context between tasks.

Model Selection

Use the least powerful model that can handle each role:

• Mechanical tasks (1-2 files, clear specs): fast/cheap model
• Integration tasks (multi-file, debugging): standard model
• Architecture/review tasks: most capable model

Steps

Step 1: Read Plan and Extract Tasks

Read the implementation plan. Extract all tasks and create a todo list.

Step 2: Execute Tasks (Per-Task Loop)

For each task:

2a. Dispatch Implementer Subagent

Frontend Detection: Before dispatching, check if the task involves frontend/UI work. Match against the Frontend Detection Rules in tn:compass (keywords: page/component/UI/layout/style/React/Vue/CSS/Tailwind/form/button/animation etc.). If matched, invoke tn:craft first to get aesthetic guidance, then include that guidance in the implementer prompt.

Use ./implementer-prompt.md template. Provide full task text (never make subagent read the plan file).

One commit per task: the implementer commits its own work before returning (the review stages below compare that commit's diff). Each task = one commit, so 2b/2c reviews have a clean BASE_SHA→HEAD_SHA range per task.

Handle status:

• DONE: Proceed to spec compliance review
• DONE_WITH_CONCERNS: Read concerns, address if needed, proceed
• NEEDS_CONTEXT: Provide missing context and re-dispatch
• BLOCKED: Assess blocker, provide context or re-dispatch with better model

Never ignore an escalation or force the same model to retry without changes.

2b. Dispatch Spec Reviewer Subagent

Use ./spec-reviewer-prompt.md template. Verify implementer built what was requested (nothing more, nothing less).

• If issues → implementer fixes → re-review
• Don't skip the re-review

2c. Dispatch Code Quality Reviewer Subagent

Use ./code-quality-reviewer-prompt.md template. Only after spec compliance passes.

• If issues → implementer fixes → re-review
• Don't skip the re-review

2d. Mark Task Complete

Mark the task as complete in the todo list and in techneering/changes/<name>/tasks.md.

Step 3: Parallel Scheduling

Analyze task dependencies, build DAG:

• Independent tasks can be dispatched in parallel (max 3 concurrent)
• Dependent tasks run in topological order
• Avoid resource conflicts between parallel tasks

Red Flags

Thought	Reality
"Skip review to speed up"	Review is quality assurance, not an optional step
"Let the subagent read the plan file"	Must provide full text — don't make subagent read files
"Ignore the subagent's BLOCKED"	BLOCKED = needs your help, not "try again"
"Run code quality review before spec review"	Order is irreversible: spec first, then quality

Quick Reference

Dimension	Key point
Flow	implementer → spec review → code quality review → next task
Subagents	Independent per task, don't inherit session context
Parallel	Independent tasks can run parallel, max 3 concurrent
Templates	implementer-prompt.md / spec-reviewer-prompt.md / code-quality-reviewer-prompt.md
Path placeholder	Always use techneering/changes/<name>/

Next Step

When done:

• Normal case (all tasks complete) → invoke tn:gate to run test verification
• Issue encountered → invoke tn:diagnose for systematic debugging
• User requests stop → end the current flow

Guardrails

• Never skip reviews (spec compliance OR code quality)
• Never start code quality review before spec compliance passes
• Never move to next task while either review has open issues
• Never make subagent read plan file — provide full text instead
• Never ignore subagent questions or escalations
• If reviewer finds issues: implementer fixes → reviewer re-reviews → repeat until approved
• Follow tn:redgreen for each task and tn:diagnose if issues arise