cm-execution

Execution — Execute Plans at Scale

TL;DR

• Use when running an approved plan from cm-planning
• Modes: A=batch, B=subagent-per-task, C=parallel, D=RARV, E=TRIZ-parallel
• Reads: handoff/plan.json — Writes: handoff/exec.json
• Always: tests pass before reporting done
• Next: cm-code-review

Role: Lead Developer — You execute implementation plans systematically with quality gates at every checkpoint.

Three modes, one skill. Choose based on task structure.

Persona Dispatch (Phase 2)

In Mode B (subagent-per-task) and Mode E (TRIZ-parallel), dispatch the right persona from agents/ based on the task type:

When dispatching a subagent, pass subagent_type: <persona> (Codex agents) or load the persona file as the system prompt for the inner call.

Step 0: Load Working Memory (MANDATORY)

Per _shared/helpers.md#Load-Working-Memory

After EACH completed task: Per _shared/helpers.md#Update-Continuity

Pre-flight: Skill Coverage Audit

Before choosing execution mode, scan plan tasks for technology keywords:

1. Extract technologies/frameworks/tools from ALL task descriptions
2. Cross-reference with cm-skill-index Layer 1 triggers
3. Check installed external skills: npx skills list
4. If gap found → trigger Discovery Loop (cm-skill-mastery Part C)
   → npx skills find "{keyword}" → review → ask user → install
5. Log any installations to .cm-skills-log.json
6. Code Intelligence Context (cm-codeintell):
   → IF codegraph available: codegraph_context(task) for each task
   → IF modifying shared code: codegraph_impact(symbol, depth=2)
   → IF impact > 10 files: WARN "High impact change"
   → Inject context into agent prompts → agents skip grep/glob
7. Only proceed to Mode Selection after all gaps resolved

Mode Selection

Have a plan with independent tasks?
├── Need SPEED + QUALITY on 3+ tasks?
│   └── YES → Mode E: TRIZ-Parallel ⚡ (recommended)
├── Stay in this session?
│   ├── YES → Mode B: Subagent-Driven
│   └── NO → Mode A: Batch Execution
└── Multiple independent failures/problems?
    └── YES → Mode C: Parallel Dispatch

Mode A: Batch Execution

Process

1. Load plan from Fission-AI OpenSpec (openspec/changes/[initiative-name]/tasks.md and design.md) → review critically → raise concerns
2. Execute batch (default: 3 tasks)
   • Mark in_progress → follow steps → verify → mark complete
3. Report → show what was done + verification output
4. Continue → apply feedback → next batch
5. Complete → use cm-code-review to finish
6. Archive → After all tasks complete, manually move the OpenSpec folder to openspec/changes/archive/[date]-[name]/

Rules

• Follow plan steps exactly
• Don't skip verifications
• Between batches: report and wait
• Stop when blocked, don't guess

Mode B: Subagent-Driven Development

Process

1. Read plan from openspec/changes/[initiative-name]/tasks.md → extract ALL tasks with full text
2. Per task:
   • Dispatch implementer subagent with full task text
   • Answer subagent questions if any
   • Subagent implements, tests, commits, self-reviews
   • Dispatch spec reviewer → confirm matches spec
   • Dispatch code quality reviewer → confirm quality
   • If issues → implementer fixes → re-review → repeat
3. After all tasks → final code review → cm-code-review

Prompt Template (Implementer)

Implement [TASK_NAME]:

[Full task text from plan]

Context: [Where this fits in the project]

Rules:
- Follow TDD (cm-tdd)
- Commit when done
- Self-review before reporting
- Ask questions if unclear

Return: Summary of what you did + test results

Red Flags

• Never start on main/master without consent
• Never skip reviews (spec OR quality)
• Never dispatch parallel implementers (conflicts)
• Never accept "close enough" on spec compliance

Mode C: Parallel Dispatch

When

• 3+ test files failing with different root causes
• Multiple subsystems broken independently
• Each problem doesn't need context from others

Process

1. Group failures by independent domain
2. Create focused agent prompt per domain:
   • Specific scope (one file/subsystem)
   • Clear goal
   • Constraints (don't change other code)
   • Expected output format
3. Dispatch in parallel
4. Review + integrate → verify no conflicts → run full suite

Common Mistakes

• ❌ Too broad: "Fix all the tests"
• ✅ Specific: "Fix agent-tool-abort.test.ts"
• ❌ No context: "Fix the race condition"
• ✅ Context: Paste error messages + test names

Mode D: Autonomous RARV

Self-driving execution. Tasks flow through Reason → Act → Reflect → Verify automatically.

When

• User runs /cm-start with a goal
• cm-tasks.json exists with backlog items
• You want continuous autonomous execution

Process (RARV Cycle)

LOOP until backlog empty or user interrupts:
  1. REASON  → Read cm-tasks.json → pick highest-priority backlog task
                Update task status to "in_progress"
                Log: { phase: "REASON", message: "Selected: <title>" }

  2. ACT     → Execute using the task's assigned CM skill
                (cm-tdd, cm-debugging, cm-safe-deploy, etc.)
                Log: { phase: "ACT", message: "<what was done>" }

  3. REFLECT → Update cm-tasks.json with results
                Log: { phase: "REFLECT", message: "<outcome summary>" }

  4. VERIFY  → Run tests/checks (cm-quality-gate)
                If PASS → status = "done", completed_at = now()
                If FAIL → rarv_cycles++, log error, retry from REASON
                If rarv_cycles >= 2 → attempt Skill Discovery Fallback:
                  → npx skills find "{task keywords}"
                  → If skill found + user approves → install, reset rarv_cycles = 0, retry
                  → If NOT found → rarv_cycles >= 3 → status = "blocked"
                Log: { phase: "VERIFY", message: "✅ passed" or "❌ <error>" }

  5. NEXT    → Recalculate stats, pick next task

cm-tasks.json Update Protocol

After EVERY phase, you MUST:

1. Read current cm-tasks.json
2. Sync state from openspec/changes/[initiative-name]/tasks.md (Keep both human-readable MD and AI-executable JSON in parallel sync)
3. Find the active task by id
4. Update status, logs[], timestamps
5. Recalculate stats object:

   stats.total = tasks.length
   stats.done = tasks.filter(t => t.status === 'done').length
   stats.in_progress = tasks.filter(t => t.status === 'in_progress').length
   stats.blocked = tasks.filter(t => t.status === 'blocked').length
   stats.backlog = tasks.filter(t => t.status === 'backlog').length
   stats.rarv_cycles_total = tasks.reduce((sum, t) => sum + (t.rarv_cycles || 0), 0)
   

6. Set updated to current ISO timestamp
7. Write back to cm-tasks.json

Rules

• Max 3 retries per task before marking "blocked"
• Always log — the dashboard reads logs in real-time
• Don't batch-skip — execute one task at a time through full RARV
• Respect interrupts — if user sends a message, pause and respond

Mode E: TRIZ-Parallel ⚡

Speed AND quality. 6 TRIZ principles resolve the contradiction.

When

• 3+ tasks that can potentially run in parallel
• Speed is important but quality cannot be sacrificed
• Tasks are well-defined with clear file scope
• You need to maximize throughput without merge conflicts

TRIZ Principles Applied

Process

1. ANALYZE    → Extract file dependencies from task descriptions
2. GRAPH      → Build dependency graph, group into independent batches
3. ADAPT      → Read parallel history, compute optimal batch size
4. PRE-FLIGHT → Check conflict ledger for overlaps with running agents
5. DISPATCH   → Send batch to agents with quality contracts
6. MONITOR    → Each agent reports modified files → detect conflicts
7. VERIFY     → Each agent runs mini quality gate before reporting done
8. RECORD     → Update parallel history for future batch sizing

Rules

• Never dispatch conflicting tasks — pre-flight must pass
• Each agent must self-validate — no "trust me it works"
• Adaptive sizing is mandatory — don't hardcode batch sizes
• File scope is enforced — agents must not modify files outside their scope
• Conflict = halt — stop further dispatch until conflict is resolved

Common Mistakes

• ❌ "All tasks are independent" → Always run dependency analysis first
• ❌ "Skip pre-flight, save time" → Pre-flight prevents wasted agent work
• ❌ "Batch size 5 for everything" → Start at 2, let the system adapt
• ❌ "One task failed, continue anyway" → Fix before next batch

Mode F: Party (Persona Rotation, single agent)

One agent, three voices. Cheaper than Mode B, deeper than Mode A.

When

• Task is non-trivial but small enough for one session
• You want multi-perspective scrutiny without paying for N subagents
• Quality matters more than raw speed

Process

1. ARCHITECT  → Load agents/architect.md as voice → propose design / approach
                → Append round to .cm/handoff/party.json
2. ENGINEER   → Load agents/engineer.md as voice → implement against the design
                → Append round
3. REVIEWER   → Load agents/reviewer.md as voice → critique implementation
                → verdict: "pass" | "revise" | "block"
                → Append round
4. (optional) SECURITY → Load agents/security.md when task touches auth/files/subprocess
                → Append round
5. SYNTHESIZE → Write `final` summary; mark handoff emitted_at

Stop Conditions

• Reviewer verdict pass → done, write final
• Reviewer verdict revise → loop back to ENGINEER (max 2 revisions)
• Reviewer verdict block or 3rd revision → escalate to user, do not ship

Output Contract

Writes to .cm/handoff/party.json matching PartyHandoff from src/handoff/contracts.ts:

{
  "schema": "party@1",
  "emitted_at": "<ISO>",
  "emitted_by": "cm-execution",
  "data": {
    "topic": "<task title>",
    "rounds": [
      {"persona":"architect","output":"...","ts":"<ISO>"},
      {"persona":"engineer","output":"...","ts":"<ISO>"},
      {"persona":"reviewer","output":"...","verdict":"pass","ts":"<ISO>"}
    ],
    "final": "<one-paragraph synthesis>"
  }
}

Rules

• Never skip the reviewer round — that defeats the point of party mode
• Never edit a previous round — append-only, like a journal
• One agent the whole way — do NOT dispatch subagents inside party mode (use Mode B for that)
• Persona files are the source of truth — load them as system context, don't paraphrase

Common Mistakes

• ❌ Running architect + engineer in same voice → missed trade-offs
• ❌ Skipping security persona on auth/file-touching tasks → see Security Rules below
• ❌ Letting reviewer revise > 2× → that's a planning problem, escalate

Security Rules (Learned: March 2026)

Code that touches files, subprocesses, or the DOM MUST follow these rules. No exceptions.

Frontend — DOM Safety

Always: Escape before innerHTML, prefer textContent, validate URLs via allowlist.

Backend — Python

Always: Import safe_path, validate EVERY path from CLI/config/API against a base directory.

Backend — Express/Node

Integration

Workflows

Karpathy Discipline — Surgical Changes

Every executor (single agent or subagent fleet) must follow:

• Touch only what the task requires. No "while I'm here" formatting, renames, or refactors.
• Match existing style even if you'd write it differently. Consistency > taste.
• Notice unrelated dead code? Mention it, don't delete. Out-of-scope cleanup → spawn a separate task.
• Clean only your own orphans. If your edit makes an import/var unused, remove it. Pre-existing dead code stays unless explicitly asked.

Diff test: Every changed line should trace to the task. If a reviewer asks "why this line?", you must point at the task — not "felt cleaner".

The Bottom Line

Choose your mode. Execute systematically. Review at every checkpoint.