Fleet judge
SAGE is not an orchestrator. It does not spawn agents, rewrite trees, or own merges. It is a judge for fleets of parallel coding sessions.
Keep the human at fleet altitude
When five Claude/Grok sessions run on one desk, the scarce resource is your attention. SAGE answers:
- Who is live, idle, or stuck?
- Who claims which territory (globs / paths)?
- What should I know before merge?
So you stay above the battlefield instead of living inside one pane.
Read-only, fail-open
- No writes to your product trees as “help.”
- Missing sessions, empty stores, half-wired hooks → degrade and tell you; never hard-block a tool.
Default OFF
Judging is opt-in (sage on / sage enable). A fresh install leaves the fleet unjudged until you say so. That is intentional trust design — not a missing feature.
Universal core vs project
| Layer | What |
|---|---|
| Universal core | Board, war, doctor, session store, hooks, statusline — same on every machine |
| Project layer | Optional adapters, project-scoped enable, backlog conventions you choose |
See README § “Universal core vs your project” and CONVENTIONS.md for an example controller contract (not law).
Flywheel
- Sessions stamp liveness and optional territory.
- You (or a skill) call
board/war/ merge brief. - You reallocate work; SAGE does not reallocate for you.
Agent session protocol: skill skills/sage-fleet/SKILL.md after install.
Optional live judge pane
A dedicated Claude/Grok session can run sage judge on and publish continuous
briefs (narrative only). Workers still trust CLI contested/clear as authority;
briefs layer after facts when fresh (including a short post-exit grace window).
See recipe Live judge and skill sage-judge.
Related
- Getting started
- Safety notes
- Architecture map (agents):
agentic-sage-mind/map/index.md