disinto/exec/AGENTS.md

Executive Assistant Agent

Role: Interactive personal assistant for the executive (project founder). Communicates via Matrix in a persistent conversational loop. Unlike all other disinto agents, the exec is message-driven — it activates when the executive sends a message, not on a cron schedule.

Think of it as the human-facing interface to the entire factory. The executive talks to exec; exec talks to the factory. OpenClaw-style: proactive, personal, persistent memory, distinct character.

Trigger: Matrix messages tagged [exec] or direct messages to the exec bot. The matrix listener dispatches incoming messages into the exec tmux session. If no session exists, exec-session.sh spawns one on demand.

A daily briefing can be scheduled via cron (optional):

0 7 * * *  /path/to/disinto/exec/exec-briefing.sh

Key files:

• exec/exec-session.sh — Session manager: spawns or reattaches persistent Claude tmux session with full factory context. Handles on-demand startup when the matrix listener receives an exec-tagged message and no session exists.
• exec/exec-briefing.sh — Optional cron wrapper for daily morning briefing. Spawns a session, injects the briefing prompt, posts summary to Matrix.
• exec/CHARACTER.md — Personality definition, tone, communication style. Read by Claude at session start. The exec has a distinct voice.
• exec/PROMPT.md — System prompt template with factory context injection points.
• exec/MEMORY.md — Persistent memory across conversations. Updated by Claude at the end of each session (decisions, preferences, context learned).
• exec/journal/ — Raw conversation logs, one file per day.

Capabilities (what the exec can do for the executive):

• Status briefing: summarize agent activity, open issues, recent merges, health alerts, pending vault items
• Issue triage: discuss issues, help prioritize, answer "what should I focus on?"
• Delegate work: file issues, relabel, promote to backlog — on behalf of the executive
• Query factory state: read journals, prerequisite tree, agent logs, CI status, VISION.md progress
• Research: search the web, fetch pages, gather information
• Memory: remember decisions, preferences, project context across sessions

What the exec does NOT do:

• Write code or open PRs (that's the dev agent's job)
• Review PRs (that's the review agent's job)
• Make autonomous decisions about the codebase
• Approve vault items (the executive does that directly)

Session lifecycle:

1. Matrix message arrives tagged [exec] (or dispatched to exec)
2. Listener checks for active exec-${PROJECT_NAME} tmux session
3. If no session → spawn via exec-session.sh:
   • Loads compass from $EXEC_COMPASS (required — refuses to start without it)
   • Loads CHARACTER.md from repo (voice, relationships)
   • Loads MEMORY.md, factory state into prompt
4. Inject message into tmux session
5. Claude responds → response captured and posted back to Matrix thread
6. Session stays alive for EXEC_SESSION_TTL (default: 1h idle timeout)
7. On session end → Claude updates MEMORY.md, session logged to journal

Compass separation: The compass (identity, moral core) lives outside the repo at a path specified by EXEC_COMPASS in .env or .env.enc. This is intentional — the factory can modify CHARACTER.md (voice, relationships) via PRs, but it cannot modify the compass. The executive controls the compass directly, like a secret.

Environment variables consumed:

• EXEC_COMPASS — Required. Path to the compass file (identity, moral core). Lives outside the repo. Agent refuses to start without it.
• FORGE_TOKEN, FORGE_REPO, FORGE_API, PROJECT_NAME, PROJECT_REPO_ROOT
• PRIMARY_BRANCH
• MATRIX_TOKEN, MATRIX_ROOM_ID, MATRIX_HOMESERVER — Required (exec is Matrix-native)
• EXEC_SESSION_TTL — Idle timeout in seconds (default: 3600)
• EXEC_CHARACTER — Override character file path (default: exec/CHARACTER.md)