OpenClaw CMP

Multi-AI comparison workflow for OpenClaw.

Languages: English | 繁體中文 | 日本語

Current release: v0.3.2

Overview

OpenClaw CMP registers /cmp as a real OpenClaw command and compares answers across multiple browser-backed AI platforms.

Current targets:

• ChatGPT
• Claude
• Gemini

Current output goals:

• Best Combined Answer first
• Real synthesis instead of placeholder completion text
• Graceful degradation when one provider fails
• Concise diagnostics for debugging

Features

• Registers /cmp as a real OpenClaw command
• Sends the same question to multiple platforms
• Waits for completion and extracts validated answers
• Synthesizes a final comparison with an available model
• Continues when one provider is unavailable
• Preserves raw-answer diagnostics for debugging
• Supports English, Japanese, and Traditional Chinese output flows

Example Output

🔄 Multi-AI Answer Comparison
Question
請用三點比較 Python 和 Rust 在系統編程上的核心差異

Best Combined Answer
Python and Rust differ most in memory management, runtime model, and systems-level control...

Consensus
- Python uses garbage collection; Rust uses ownership and compile-time checks.
- Python is productive at higher abstraction levels, while Rust is better suited to low-level performance-sensitive work.

Major Differences
- ChatGPT emphasized Python as a control-layer language.
- Gemini put more weight on Rust's deterministic behavior and concurrency model.

Unique Additions
- Gemini added explicit performance-range comparisons.
- ChatGPT highlighted Python plus C/C++ extension patterns.

Repo Layout

extensions/cmp/
skills/cmp/
skills/_shared/
scripts/
tests/
docs/

Install

Install into a local OpenClaw setup:

npm run install:local
openclaw gateway restart

The local installer preserves existing ~/.openclaw/skills/cmp/config/state.json, platforms.json, ~/.openclaw/skills/cmp/logs, and ~/.openclaw/extensions/cmp/.git when present, so enabled-platform state, local platform tuning, recent diagnostics, and extension working-tree metadata survive repo updates.

Manual copy:

• extensions/cmp -> ~/.openclaw/extensions/cmp
• skills/cmp -> ~/.openclaw/skills/cmp
• skills/_shared -> ~/.openclaw/skills/_shared

Validation

Smoke test:

npm test
npm run doctor

Live local run:

npm run test:live

npm run doctor checks the installed CMP plugin, browser/evaluate config, native command settings, channel allowlists, enabled platforms, gateway status, and recent run traces.

After installing or updating CMP, restart the gateway before testing Discord slash commands:

openclaw gateway restart
npm run doctor

The doctor verifies that the latest gateway startup actually loaded cmp; this catches the case where files are installed but the running gateway still has the older plugin set.

Current Synthesis Runtime

CMP runs a two-pass synthesis pipeline. Both passes use the GitHub Copilot endpoint (free/flat-rate — no per-token cost).

Pass 1 — structure synthesis (GPT-4.1, 5000 tokens)

Produces the full five-section output: platformViews, Consensus, Major Differences, Unique Additions, and an initial directAnswer.

Pass 2 — Best Combined Answer (GPT-4.1, 8000 tokens dedicated)

A second, independent call focused entirely on Best Combined Answer. It receives every platform's full answer and rewrites the section with a dedicated 8000-token budget. Pass 2 only replaces the initial draft when it produces a longer, richer result; if it fails for any reason, the Pass 1 draft is kept.

Trace events for this pass are named rich_direct_answer_* so logs describe the role of the pass rather than implying a specific vendor.

	Model	Token budget	Fallback chain
Pass 1	github-copilot/gpt-4.1	5000	gpt-4o → gpt-5-mini
Pass 2	github-copilot/gpt-4.1	8000	gpt-4o → gpt-5-mini

Changelog

See CHANGELOG.md for version history.

Roadmap

• Long-term CMP improvements are tracked in TODO.md.

Completion Detection

Each platform uses a dedicated per-poll DOM probe rather than injected persistent JS state:

• ChatGPT / Gemini / Grok: dedicated completion check functions evaluate DOM signals on each polling cycle.
• Claude: uses the same dedicated probe pattern. Claude.ai navigates from /new to /chat/<id> on submission, which invalidates the initial tab reference. The probe handles this by evaluating fresh on every cycle — no injected state means no stale-reference failures.

CMP requires a response to be stable for multiple consecutive polls before treating it as complete, so temporarily paused streaming does not trigger premature synthesis.

Notes

• For Chinese questions, synthesis is instructed to use Traditional Chinese rather than Simplified Chinese.
• Browser-backed providers remain subject to UI drift and login/session constraints.
• All providers are subject to login/session requirements. CMP never logs in on your behalf.

License

MIT