Agent control plane

The problem

I run dozens of projects with coding agents. Each session started cold: it did not know the project's status, what I had decided last week, or that another repo already solved the problem it was about to solve again.

How it works

Claude Code hooks relay session events (start, prompt, tool use, stop, end, notifications) to a local server. SessionStart returns a briefing that goes straight into the agent's context. Stop blocks once when files changed, so every piece of work ends with a report.

A worker reads each project's docs with a cheap model, never secret files, redacts what it keeps and writes topic entries into SQLite FTS5. It re-indexes when a project's git fingerprint changes. The MCP server exposes report, ask, project, portfolio and briefing, plus tools for todos and deadlines.

Decisions

Fail open. A control plane that breaks the agent when it is down gets uninstalled, so hooks time out fast and spool instead.

No LLM on the hot path. Hooks and briefings are plain lookups; the model only runs in the background worker.

Block once, then allow. Guards nudge the agent to ask or report, and a retry goes through.

Measured

Retrieval quality has an eval: 33 real queries, recall@5 0.94 and MRR 0.78. A change that drops below the gate does not ship.