Megaplan

A planning and execution harness that helps LLMs solve complex tasks through structured phases — plan, critique, gate, revise, finalize, execute, and review. Instead of one-shot attempts, Megaplan gives any model a rigorous process with independent critique and gating.

Quick Start — Claude Code / Codex

Copy and give this to your agent:

Please install megaplan and set it up for this project:

pip install megaplan-harness
megaplan setup

Once you're done, ask me what I need megaplan for.

Quick Start — Open Models via OpenRouter

Copy and give this to your agent:

Please install megaplan with the open-model backend and set it up:

pip install megaplan-harness hermes-agent

Then create ~/.hermes/.env with:
OPENROUTER_API_KEY=<my key>

Then run: megaplan setup

Once you're done, ask me what I need megaplan for.

Get an OpenRouter key at openrouter.ai/keys. Any model on OpenRouter works — Qwen, Llama, Mistral, DeepSeek, etc.

---

How it works

plan → critique → gate → [revise → critique → gate]* → finalize → execute → review

Each phase can use a different model. The critique phase uses an independent model to review the plan and raise flags. The gate decides whether to proceed or iterate. This prevents models from rubber-stamping their own work. Planning now goes through a visible prep phase so repository investigation is observable instead of hidden inside plan.

Running manually

megaplan init --project-dir . "Fix the authentication bug in login.py"
megaplan plan --plan <name>
megaplan critique --plan <name>
megaplan gate --plan <name>
megaplan finalize --plan <name>
megaplan execute --plan <name>

Using different models per phase

Models with provider prefixes route to direct APIs. Models without a prefix go through OpenRouter:

{
  "models": {
    "prep": "zhipu:glm-5.1",
    "plan": "zhipu:glm-5.1",
    "critique": "minimax:MiniMax-M2.7-highspeed",
    "execute": "zhipu:glm-5.1",
    "review": "minimax:MiniMax-M2.7-highspeed"
  }
}

Configure direct provider keys in ~/.hermes/.env:

ZHIPU_API_KEY=...          # for zhipu: prefix
MINIMAX_API_KEY=...        # for minimax: prefix
GEMINI_API_KEY=...         # for google: prefix

Robustness levels

• light — visible prep + one critique/revise pass, no gate or review
• standard — visible prep + 4 critique checks (default)
• heavy — visible prep + 8 critique checks

Observability

megaplan status --plan <name>
megaplan watch --plan <name>

status exposes additive lifecycle fields such as active_step, last_step, notes, cost, and session summaries. watch adds the current execution-progress snapshot in the same machine-readable response.

Subagent mode (Claude Code)

Subagent mode delegates the full workflow to an autonomous Claude Code agent, returning control only at defined breakpoints. It is the default orchestration mode for Claude Code. Codex and Cursor continue to run inline.

megaplan config set orchestration.mode subagent   # default
megaplan config set orchestration.mode inline      # switch back

Configuration & Defaults

View all settings with megaplan config show. Override with megaplan config set <key> <value>. Reset with megaplan config reset.

Key	Default	Description
orchestration.mode	subagent	inline or subagent (Claude Code only)
orchestration.max_critique_concurrency	2	Max parallel critique checks
execution.worker_timeout_seconds	7200	Worker process timeout (seconds)
execution.max_execute_no_progress	3	No-progress execute attempts before escalation
execution.max_review_rework_cycles	3	Review→rework loops before force-proceeding
agents.<step>	varies	Agent for each phase (claude, codex, hermes)

megaplan config set execution.worker_timeout_seconds 3600
megaplan config set agents.critique hermes
megaplan config reset

SWE-bench Experiment

Megaplan is being tested live against Claude 4.5 Opus on SWE-bench Verified:

• Live dashboard — watch the experiment in real time
• hermes-megaplan — experiment orchestration code

Code Health

License

MIT