opencode-skill-trigger

OpenCode plugin that whispers skill suggestions and memory reminders into user messages.

What it does

Two things, both injected as natural-language commands appended to the user's message:

1. Skill Hint — When the user's message matches keywords for an installed skill, appends a command like 用grilling审。 so the AI auto-loads that skill.
2. Memory Whisper — Every non-trivial turn, appends 先memory_search相关记忆。 so the AI searches memory before answering.

The AI sees these as user commands (not system background noise), so it executes them.

Why

cerebro's memory reminder injects into system prompt. AI treats it as background noise and ignores it. This plugin injects into the user message stream via chat.message hook + parts.push — the AI treats it as a direct command.

Install

Local dev

Add to opencode.jsonc:

{
  "plugins": [
    "/path/to/opencode-skill-trigger/plugin.js"
  ]
}

Restart opencode.

Config

Edit config.json:

{
  "skills": [
    {"skill": "grilling", "keywords": ["审", "漏洞", "stress-test"]}
  ],
  "whisper": {
    "prompt": "先memory_search相关记忆。",
    "testFlag": "[skill-trigger已生效]"
  },
  "skillHint": {
    "testFlag": "[skill-hint已生效]"
  }
}

• skills[] — keyword → skill mapping. Substring match, case-insensitive.
• whisper.prompt — memory reminder injected every turn.
• testFlag — visible marker for testing. Delete for production.

How it works

user sends "审下这个方案有没有漏洞"
    ↓
chat.message hook fires
    ↓
keyword match: grilling (审, 漏洞)
    ↓
parts.push({text: "用grilling审。 [skill-hint已生效] 先memory_search相关记忆。 [skill-trigger已生效]"})
    ↓
AI sees: "审下这个方案有没有漏洞 [injection]"
    ↓
AI loads grilling skill + calls memory_search

Files

• plugin.js — hook registration, matching, injection
• config.json — editable rules
• CONTEXT.md — glossary
• DESIGN.md — architecture
• docs/adr/0001-injection-mechanism.md — why parts.push not system prompt

License

MIT