agent-obs-opencode

agent-obs Self-Reporting

You are an observed agent. The agent-obs MCP server is a local flight recorder for everything you do: tool calls, errors, tokens, cost, and a final A-F grade. It only works if you report into it. This skill defines the reporting protocol. Follow it in every task, in every session, without being asked.

The Protocol (four tools, three phases)

Phase 1 — Start: call start_session FIRST

• At the very beginning of every task — before Read, before Bash, before anything else — call the start_session tool on the agent-obs server.
• Pass:
  • description (required): one sentence describing the task, e.g. "Fix login redirect bug in auth middleware".
  • agentType: your client, e.g. "opencode", "claude-code", "cursor", "codex".
  • model: your model id, e.g. "deepseek/deepseek-v4-pro".
• The result contains a session id. Save it. Every subsequent log_tool_call and the final end_session needs this exact sessionId.
• One task = one session. If the user pivots to a brand-new task in the same conversation, end the current session and start a new one.

Example:

start_session({
  "description": "Refactor payment webhook handler and add tests",
  "agentType": "opencode",
  "model": "deepseek/deepseek-v4-pro"
})
→ { "id": "a1b2c3d4" }   // keep this sessionId

Phase 2 — Log: call log_tool_call after EVERY tool call

• Immediately after each tool call completes (Read, Write, Edit, Bash, Glob, Grep, Task, WebFetch, TodoWrite, any MCP tool — ALL of them), call log_tool_call with:
  • sessionId (required): the id from start_session.
  • toolName (required): the tool you just used, e.g. "Read", "Bash", "Edit".
  • status (required): "success" if the tool call worked, "error" if it failed, errored, or was rejected.
  • toolServer: where the tool lives, e.g. "builtin", "playwright", "cloudflare-bindings".
  • input: a compact object of the arguments, e.g. { "filePath": "src/auth.ts" }. Redact secrets — never log tokens, keys, or passwords.
  • outputSummary: one line describing what happened, e.g. "Read 120 lines of auth middleware" or "npm test: 42 passed".
  • durationMs: approximate wall-clock duration if known.
  • errorMessage: the error text when status is "error".
• NEVER skip log_tool_call — even for fast operations. Every tool call must be logged. A one-line Glob counts. A 2ms Read counts. If you called a tool, you log it. No batching several calls into one log entry, no "it was trivial" exceptions.
• Log failures too. An errored Bash command logged with status: "error" is exactly the signal the grading system needs.
• The only calls you do NOT log are the agent-obs tools themselves (start_session, log_tool_call, end_session, get_last_session). Logging the logger would recurse forever.