Code Build

---
name: code-build
description: Implement an already-approved plan from swarm-report/<slug>-plan.md. Routes each task to the executing agent that owns its file scope (frontend / backend / devops / mobile), runs the tests, writes an implementation report. Use AFTER /code-plan, never before.
---

# Skill: /code-build

ORCHESTRATOR. You route work to executing agents by file scope. No code edits in the
main loop.

## Invocation
`/code-build <slug>`

## Steps
1. **Find the plan**: `swarm-report/<slug>-plan.md`. Missing → abort:
   "Run `/code-plan \"<feature>\"` first."
2. **Check Blockers** in the plan. Any unresolved HIGH → abort and name it. The user must
   resolve or explicitly waive it.
3. **Route by scope**: read the plan's `affected_files`. Match each against the Executing
   table in `AGENTS.md` (frontend / backend / devops / mobile). Group files by agent.
   No match for a file → ask the user which agent owns it.
4. **Spawn the matched exec agents** — one `Task` call each, in parallel when a feature
   spans several layers (single message, N calls, `subagent_type: general-purpose`).
   Prompt each:
   > Answer TERSE. Read `.claude/agents/<scope>.md` and follow it exactly.
   > Plan: swarm-report/<slug>-plan.md. Your scope: <the files for this agent>.
   If one agent's output changes a contract another needs (e.g. backend `api_changes`),
   pass that note into the dependent agent's prompt — or run backend first, then frontend.
5. **Verify**: read each agent's `tests_result`. Any `status: blocked` or failing test →
   do NOT claim success. Report the failures verbatim and stop. Offer `/code-build <slug>`
   retry after the fix.
6. **Write** `swarm-report/<slug>-build.md`: per-agent changed files, test commands +
   real results, cross-layer notes.
7. **Report** to user: status per scope + test results. Quote real output — no "done"
   without proof.