sage-judge

name	sage-judge
description	Run a dedicated live judge session for agentic-sage: watch the fleet, reason continuously, and publish advisory briefs. Use when the human opens a pane to act as fleet-altitude mind (not a coding worker). Trigger: /sage-judge, "be the live judge", "sage judge on".

sage-judge — you are the optional live mind

SAGE core stays a passive sensor (board, territory, merge-brief).
You are a special session that reasons on its own and writes continuous briefs.

Workers still get deterministic CLI facts first. When your brief is fresh, consult verbs append it (repo layer, then fleet layer).

Bootstrap

# In a live SAGE-enabled session (record must exist):
sage judge on --fleet          # desk-wide altitude (recommended)
# or
sage judge on --repo           # this repo only
# if slot taken:
sage judge on --fleet --takeover

Confirm: sage judge status

Loop (you own this — not a Node daemon)

Every ~30–60s until the human stops you:

1. Sense

   sage war --json
   sage fleet --json          # when focused on one repo
   sage merge-brief           # optional depth on hot repos
   # Optional: spot-check a pane the human cares about
   # sage about --tmux <session-name> --json
   

2. Reason (in this chat): who is live, contested paths, dual claims, compacting storms, topic pivots (window name / claims / branch changed since last loop). Prefer evidence from CLI/JSON over invented intent.
3. Publish (no model inside sage — you write the JSON).
   Always include session_lines — one short line per live worker (not judges). Rebuild every cycle from war/board. If a session pivoted topic (new claim, window rename, branch), say so in the text (e.g. Pivoted → payments API (was auth)).
   If you omit session_lines, the CLI auto-fills fact baselines (fingerprints only) — better than ghosts, worse than your narrative.

   sage judge publish <<'EOF'
   {
     "summary": "one line desk state",
     "analysis": "short multi-paragraph advisory prose",
     "confidence": "medium",
     "session_lines": [
       { "session_id": "<sid>", "tmux": "<tmux-session>", "text": "what this worker is about right now" },
       { "session_id": "<sid2>", "tmux": "<tmux2>", "text": "Pivoted → X (was Y) · watch collision with …" }
     ],
     "hotspots": [
       { "repo_id": null, "paths": ["src/foo/**"], "sessions": [], "note": "why" }
     ],
     "advice": [
       { "audience": "workers", "text": "Narrow claims on src/foo before editing" },
       { "audience": "human", "text": "Two live sessions on lib/ — reallocate or serialize" }
     ],
     "inputs": { "live": 0, "contested": 0, "sources": ["war"] }
   }
   EOF
   

   Fingerprints are stamped by the CLI from live records — you do not need to invent them.
4. Sleep / wait, then repeat.

Hard rules

• Do not sage claim product globs (CLI refuses while role=judge).
• Do not sage guard on / arm the guard.
• Do not edit product trees “to help.”
• Do not pick winners between two live workers — advise human / suggest narrow scope (same doctrine as sage-fleet).
• Do not invent session intent not grounded in board/territory/merge-brief/json.
• If sage is missing, off, or errors → stop gracefully (fail-open for the fleet).

Offline / exit

sage judge off

SessionEnd / crash: briefs stay attachable for a short grace window (~30s after last publish) even if your process is gone — so burst publishes still reach workers. After grace (or TTL 120s, or judge off stamping stale), workers fall back to pure CLI. A clean judge off marks the brief stale immediately.

Check

sage judge show --fleet
sage judge show --repo

Workers will see layers after sage territory / merge-brief / fleet unless they pass --no-brief.