@skillfm/openclaw-plugin

V2 Agent Forcing Stack L1 — OpenClaw adapter for SkillFM Beacon.

Injects the SkillFM behavior contract into the OpenClaw agent system prompt via the before_prompt_build hook (cache-friendly), so the agent reliably picks skillfm_* MCP tools instead of going for generic shell exec.

What it does

When loaded, this plugin registers a before_prompt_build hook that returns:

{ prependSystemContext: SKILLFM_BEHAVIOR_CONTRACT }

Per OpenClaw's plugin-sdk docs, prependSystemContext is placed above the cache boundary (<!-- OPENCLAW_CACHE_BOUNDARY -->), so the contract is reused across turns by Anthropic / OpenAI / DeepSeek / etc. prompt caching.

What it does NOT do

• Does not register MCP tools — those live in the @skillfm/local Beacon MCP server, registered with OpenClaw via mcp.servers.skillfm-beacon.
• Does not block tool calls (before_tool_call) — held back to Sprint 2 (L2 Hook Enforcement) to avoid false positives on legitimate bash.
• Does not modify systemPrompt directly — coexists with OpenClaw's own GPT5_BEHAVIOR_CONTRACT and other plugins.

Install

# OpenClaw must already be installed (peer dependency)
npm install -g @skillfm/openclaw-plugin

# Then enable in ~/.openclaw/openclaw.json:
# {
#   "plugins": {
#     "skillfm-forcing": { "enabled": true }
#   }
# }

Verify after install

# Trigger a turn that exercises agent forcing:
openclaw run "看下我家 AI 的状态"

# Then grep your OpenClaw session log for SKILLFM_BEHAVIOR_CONTRACT:
grep "SkillFM Beacon helps users" ~/.openclaw/logs/*.log

If the agent calls skillfm_checkup instead of curl localhost:45219/status, forcing is working.

Architecture

This plugin is one of four surfaces injecting the same SkillFM behavior contract (V2 PRD §2.3, single-source-of-truth):

1. L0 fallback: __unlock_skillfm_beacon__ MCP tool (in @skillfm/local) — agent-discoverable bootstrap when L1 adapters fail.
2. L1 Claude Code: server.instructions field on the Beacon MCP server (Claude Code 1/7 honors).
3. L1 Claude Code: ~/.claude/skills/skillfm/SKILL.md (Anthropic Skills format).
4. L1 OpenClaw: this plugin's before_prompt_build hook.

The contract content (SKILLFM_BEHAVIOR_CONTRACT) is duplicated across these packages but kept byte-identical via the scripts/check-protocol-parity.mjs CI guard in the SkillFM monorepo.

License

MIT — © SkillFM