Execution — Execute Plans at Scale
TL;DR
- Use when running an approved plan from
cm-planning - Reads
handoff/plan.jsonoropenspec/changes/[name]/tasks.md - Writes
handoff/exec.json - Always verify before reporting done
- Next
cm-code-review
Role: Lead Developer. You execute plans systematically with quality gates at every checkpoint.
Step 0: Load Working Memory (MANDATORY)
Per _shared/helpers.md#Load-Working-Memory. After EACH completed task: _shared/helpers.md#Update-Continuity.
Step 1: Pre-flight Skill Coverage Audit
Scan plan tasks for tech keywords, cross-reference cm-skill-index, check installed skills, and use npx skills find when the plan reaches beyond current coverage. If codegraph or cm-codeintell context is available, inject it into agent prompts so execution skips redundant repo searching.
Step 2: Choose Mode
Have a plan with independent tasks?
├─ Need SPEED + QUALITY on 3+ tasks?
│ └─ YES → Mode E (TRIZ-Parallel) ★ recommended
├─ One non-trivial task, want multi-perspective without subagent cost?
│ └─ YES → Mode F (Party, persona rotation)
├─ Multiple independent failures across subsystems?
│ └─ YES → Mode C (Parallel Dispatch)
├─ Autonomous loop with backlog (`/cm-start` flow)?
│ └─ YES → Mode D (RARV)
├─ Plan has independent tasks, stay in this session?
│ └─ YES → Mode B (Subagent-Driven)
└─ Otherwise
└─ Mode A (Batch Execution)
| Mode | One-line summary | Load |
|---|---|---|
| A | Batch 3 tasks, report, then continue | references/mode-a-batch.md |
| B | Fresh subagent per task + 2-stage review | references/mode-b-subagent.md |
| C | One agent per independent problem domain | references/mode-c-parallel.md |
| D | Reason → Act → Reflect → Verify loop | references/mode-d-rarv.md |
| E | Dependency-aware parallel + per-agent quality gate | references/mode-e-triz-parallel.md |
| F | Single agent rotates Architect → Engineer → Reviewer | references/mode-f-party.md |
Action: Pick exactly one mode, read only that reference, and execute from there.
Conditional References
- Mode B / E / F → also read
references/persona-dispatch.md - Task touches auth, files, subprocess, DOM, config paths, or user input → MUST read
references/security-rules.mdbefore writing code
Karpathy Discipline — Surgical Changes
- Touch only what the task requires. No "while I'm here" refactors.
- Match existing style even if you would write it differently.
- Notice unrelated dead code? Mention it, do not delete it.
- Clean only your own orphans. Pre-existing dead code stays unless asked.
- Diff test: every changed line must trace to the task.
Integration
| Skill | When |
|---|---|
cm-planning | Produces the plan this skill consumes |
cm-tdd | Use per task when implementing or fixing |
cm-quality-gate | VERIFY phase / pre-report validation |
cm-code-review | Final review after all tasks are done |
cm-design-system | Recommended before frontend-heavy tasks |
Workflows
| Command | Purpose |
|---|---|
/cm-start | Create tasks + launch RARV + open dashboard |
/cm-status | Quick terminal progress summary |
/cm-dashboard | Open browser dashboard |
The Bottom Line
Choose your mode → load that one reference → execute systematically → review at every checkpoint.