opencode

Recovered with lab prompts opencode from the trace proxy (newest capture 20260711T110419Z). Every tool routes through the proxy, which records each completion after normalizing it to the chat-completions shape, so this is the exact text that reached the model, not a copy from the tool's source. The prompts are stored raw under prompts/opencode/; this page reviews what each one does and links to the file. Sizes are in tokens, counted with the o200k_base encoding.

Agent prompt

• 2,010 tokens
• tools (10): bash, edit, glob, grep, read, skill, task, todowrite, webfetch, write
• raw: prompts/opencode/1-agent.md

opencode's prompt reads like a working style guide: keep replies short, prefer the search tools, follow the project's own conventions, and never invent a URL. It exposes a broad, general tool set, ten tools covering file work, search, a task tracker, and web fetch, and it spends its budget teaching the model when to reach for each rather than restating their signatures. The result is a mid-weight prompt that sits well below the heaviest agents while still driving a full toolbox.

Title generator

• 503 tokens
• no tools
• raw: prompts/opencode/2-title-generator.md

Beside the agent, opencode runs a small side model that names each thread. The prompt is emphatic that the model output only a title and nothing else, which is a common shape across these tools: a cheap, tool-free call whose whole job is to return one clean line.