codex-conductor

Codex Conductor

You are using Codex Conductor as a read-only session supervisor. It reads local Codex transcripts and reports what other sessions appear to be doing. Do not interrupt or type into other sessions unless the user explicitly asks for control.

Step 1 - Gather Sessions

Run the scanner:

node ~/codex-conductor/scan.js --json --minutes 30

It scans ~/.codex/sessions/**/*.jsonl, uses ~/.codex/session_index.jsonl for thread names when available, groups by session id, and returns structured JSON.

Widen the window with --minutes 60, or pass --all when the user wants older local history.

Step 2 - Render The Table

For each session, use:

• label / project - where it is working.
• title, task, and intent - what the session appears to be about.
• recent and lastAction - what happened recently.
• lastActiveRel - how long since real activity.
• status - active, open, recent, or idle.

Output format:

N Codex sessions

* <label> · <status> · <lastActiveRel>
  doing now  : <one-line inference from lastAction/recent>
  done       : <best-effort summary from recent records>
  what's left: <best-effort next step; mark uncertain guesses with ?>

Rules

• "What's left" is inference, not fact. The transcript shows activity, not a confirmed todo list.
• Keep it scannable. The user is triaging sessions, not reading a report.
• If count is 0, say no recent Codex sessions were found and suggest --minutes 60 or --all.
• This skill is read-only. For control, point the user at codex-conductor run <label>, codex-conductor say <label> <text>, and codex-conductor adopt <id>, or use the MCP tools if available.
• auto_continue runs under the irreversibility gate and refuses unattended deploy, send, delete, or spend approvals.