Feature: auto-detect Claude Code / Codex / OpenClaw instances #65
Description
Background
When running multiple AI coding agents (Claude Code, Codex, OpenClaw) simultaneously, each instance needs to manually join the office via API. This is tedious and error-prone — agents get duplicated, forget to leave, or never join at all.
Proposal
A lightweight daemon script that:
- Auto-detects running Claude Code / Codex processes every ~10 seconds (via
pgrep) - Auto-joins new instances with a descriptive name based on their working directory (e.g.
Claude Code (trading),Codex (podcast_juicer)) - Auto-removes agents when their process exits
- Monitors OpenClaw gateway logs to toggle lobster status between idle/writing
How it works
┌─────────────┐ pgrep -x claude ┌──────────────────┐
│ Claude Code │ ◄──────────────────────► │ │
│ (trading) │ │ star-office │
├─────────────┤ pgrep -x codex │ daemon │
│ Codex │ ◄──────────────────────► │ │
│ (juicer) │ │ (every 10s) │
├─────────────┤ gateway.log tail │ │
│ OpenClaw │ ◄──────────────────────► │ │
│ (lobster) │ └───────┬──────────┘
└─────────────┘ │
▼
/join-agent
/agent-push
/leave-agent
Why not hooks?
Claude Code supports PostToolUse / Stop hooks, but:
- Each hook invocation is a separate shell process with no stable session identifier
- This causes duplicate agent registrations (we saw it balloon to #97)
- Codex has no system-level hooks at all
A polling daemon avoids these issues entirely.
Status
I have a working prototype (star-office-daemon.sh) running locally. Happy to submit a PR if there's interest, or adapt to a preferred approach (e.g. integrate into the backend as a background thread).
Related PR: #64 (fixes agent overlap when many agents are in the same area)