Megaplan

High-rigor planning with adversarial critique, explicit gate decisions, and resumable execution.

When to use it

• Multi-file, ambiguous, high-risk, or multi-stage work
• Work that benefits from critique before execution
• Tasks where resumability and execution discipline matter

Skip it for obvious, low-risk edits.

Roles

Role	Subagent	Purpose
Planner	build	plan, revise, finalize
Critics	3x deep	feasibility, architecture, risk
Reviewer	deep	final independent review
Orchestrator	you	state, gate, execution loop

Workflow

init -> plan -> critique -> gate -> revise/finalize -> execute -> review

• plan, revise, finalize delegate to build
• critique runs 3 parallel deep critics and merges flags
• gate is computed by you, then decision delegated to build
• execute is owned by the main thread
• review delegates to deep

Core rules

1. Continuity source of truth

Use .megaplan/plans/<name>/runbook.md as the continuity source of truth.

Trust runbook.md over conversational memory, especially after compaction or resume.

2. Main-thread execution

During execute, the main thread owns:

• batch selection
• implementation
• progress maintenance
• commits
• runbook updates

Delegate only fresh-eyes work:

• simplify
• pre-commit review
• specialist verification when needed

3. Fresh-eyes rule

The main thread must not approve its own diff unaided.

Before each commit:

1. delegate fresh simplify
2. delegate fresh pre-commit review
3. fix anything required
4. then commit atomically

4. Never-stop

During execution, set /never-stop using the exact ## Continue section from runbook.md.

Refresh it after meaningful steps. Clear it on complete or blocked.

Minimal runbook

Keep runbook.md minimal. It should contain only:

• Status
• Step
• Next
• Batch
• Last commit
• Watch
• Blocker
• ## Continue

Do not turn the runbook into a diary. Keep history and evidence in other artifacts.

Artifacts

Use .megaplan/plans/<name>/.

Essential artifacts:

• runbook.md — continuity control plane
• plan_v<N>.md and plan_v<N>.meta.json
• faults.json
• gate.json
• finalize.json
• review.json

Additional artifacts may be written as needed for critiques, execution evidence, and summaries.

Execution loop

When Step is execute, use this loop:

1. Find next unblocked WorkingBatch
2. Implement in main thread
3. Delegate fresh simplify
4. Delegate fresh review
5. Fix anything required
6. Commit atomically
7. Update runbook.md
8. Repeat

Stop only on:

• complete — all planned work finished, reviewed, committed
• blocked — missing requirements, auth, permissions, secrets, or irreconcilable contradiction

References

• Prompt templates: references/prompts.md
• Gate logic: references/evaluation.md