constellation

Multi-persona council system. Assembles 4-6 PI persona subagents to analyze any topic through diverse behavioral lenses. Use /constellation audit, /constellation discover, /constellation build, /const

SKILL.md

GitHub

Constellation Command

Assembles a council of Otto's PI personas to analyze a topic. Each persona contributes their perspective, Captain synthesizes, Guardian holds veto authority.

Commands

Command Lead Purpose
/constellation audit Guardian Alignment check across repos
/constellation discover <topic> Scholar Research/brainstorm
/constellation build <topic> Maverick Innovation planning
/constellation ship <topic> Guardian Deployment readiness
/constellation search <query> Captain Free-form, auto-routed
/constellation refresh repos — Force refresh all repo relays
/constellation refresh personas — Force refresh all persona relays

Execution Protocol

Step 1: Parse Command

Extract from the ARGUMENTS string:

  • command: audit, discover, build, ship, search, or refresh
  • query: everything after the command keyword
  • scope: if --scope X,Y is present, parse as list; default to all repos
  • depth: if --depth shallow is present, use shallow; default to deep

If command is refresh, handle it directly:

  • refresh repos: Call POST http://localhost:8100/constellation/relays/repos/refresh and display results
  • refresh personas: Display all persona relays from GET http://localhost:8100/constellation/relays/personas
  • Return after refresh — no council needed.

Step 2: Fetch Relay State

# Fetch repo relays
curl -s http://localhost:8100/constellation/relays/repos

# Fetch persona relays
curl -s http://localhost:8100/constellation/relays/personas

Check freshness. If any relay is stale (>30 min), refresh it:

curl -s -X POST http://localhost:8100/constellation/relays/repos/refresh

Step 3: Select Council

Call the State Service or determine locally based on command type:

Command Council
audit captain, guardian, analyzer, controller, operator
discover captain, scholar, strategist, venturer, individualist
build captain, maverick, venturer, artisan, persuader
ship captain, guardian, operator, controller, analyzer
search captain + 3-5 most relevant based on query analysis

If --depth shallow, use only captain + lead + 1 expanded member.

Step 4: Write Briefing

Compose a ~200 token briefing document containing:

  • The original query
  • Scope (which repos are in play)
  • Key facts from repo relays (recent commits, health status, services)
  • Key facts from persona relays (current concerns, opportunities)
  • Context from the current repo (if relevant)

Step 5: Dispatch Council (Parallel Subagents)

For each council member (excluding Captain), dispatch a subagent using the Agent tool. Use model: "sonnet" for cost efficiency.

Each subagent receives the council-prompt.md template with variables filled in:

  • Persona profile data from GET http://localhost:8100/persona/profiles/{id}
  • The briefing from Step 4
  • Their persona relay cache
  • The query

IMPORTANT: Dispatch ALL council subagents in a SINGLE message (parallel execution). Do NOT dispatch them sequentially.

Example for a discover council:

Agent 1 (Scholar — lead):
  prompt: [council-prompt.md filled with Scholar profile + briefing + query]
  model: sonnet
  description: "Scholar perspective"

Agent 2 (Strategist):
  prompt: [council-prompt.md filled with Strategist profile + briefing + query]
  model: sonnet
  description: "Strategist perspective"

Agent 3 (Venturer):
  prompt: [council-prompt.md filled with Venturer profile + briefing + query]
  model: sonnet
  description: "Venturer perspective"

Agent 4 (Individualist):
  prompt: [council-prompt.md filled with Individualist profile + briefing + query]
  model: sonnet
  description: "Individualist perspective"

Step 6: Collect Perspectives

Parse JSON responses from each subagent. Each should return:

{
  "persona": "scholar",
  "position": "...",
  "conviction": 8.4,
  "concerns": ["...", "..."],
  "opportunities": ["...", "..."],
  "recommends_action": true
}

Step 7: Guardian Gate

Scan all perspectives for safety/sustainability/alignment concerns. If any such concern has conviction ≥ 7.5, set guardian_gate to HARD_DISAGREE. Otherwise PASS.

Step 8: Captain Synthesis

As Captain, write the synthesis:

  • Summarize the consensus (what do most personas agree on?)
  • Highlight dissent (who disagrees and why?)
  • Recommend next actions
  • Set decision status: INFORMATIONAL, AWAITING_ZAC, or BLOCKED

Step 9: Persist Report

Primary: Try the state service first:

curl -s -X POST http://localhost:8100/constellation/query \
  -H "Content-Type: application/json" \
  -d '{
    "command": "...",
    "query": "...",
    "scope": [...],
    "council": {...},
    "perspectives": [...],
    "briefing": "...",
    "synthesis": "...",
    "guardian_gate": {...}
  }'

Fallback (always): Write a YAML report file to the decision log regardless of whether the HTTP call succeeds. This is the durable record.

File: ~/Desktop/Airlock/repos/airlock-coordination/reports/YYYY-MM-DD-HH-MM-{command}-{slug}.yaml
Schema: ~/Desktop/Airlock/repos/airlock-coordination/reports/schema.yaml
Index: ~/Desktop/Airlock/repos/airlock-coordination/reports/INDEX.yaml

IMPORTANT — Privacy rule: The request_summary field must be a SYNTHESIZED generalization of the user's request, NOT verbatim input. The user's messages may contain personal, sensitive, or heartfelt content. Distill the intent to 1-2 neutral sentences. Example: "Evaluate whether the current auth architecture is ready for multi-tenant production" — not the user's exact words.

The report YAML must follow schema.yaml. After writing the report file, prepend an entry to the reports list in INDEX.yaml with: date, command, request_summary, decision_status, and filename.

Set resolution.action to pending initially. If the user responds with accept/override/dismiss, update the report file.

Step 10: Render Report

Display the report in this format:

══════════════════════════════════════════════════
  CONSTELLATION REPORT
  Type: {command} | Query: "{query}"
  Date: {date} | Council: {count} personas
══════════════════════════════════════════════════

BRIEFING (Captain)
{briefing}

──────────────────────────────────────────────────

{PERSONA_NAME} ({EMOJI}) — Lead — Conviction: {N}
Position: {position}
Concerns:
  • {concern1}
  • {concern2}
Opportunities:
  • {opp1}
  • {opp2}

──────────────────────────────────────────────────

(repeat for each council member)

══════════════════════════════════════════════════
  CONSENSUS
  Avg Conviction: {N} | Agreement: {X}/{Y}
  Dissent: {persona} ({reason})

  GUARDIAN GATE: {PASS|HARD_DISAGREE}
  {if HARD_DISAGREE: reason}
══════════════════════════════════════════════════

CAPTAIN SYNTHESIS
{synthesis}

Decision Status: {status}
Override History: {history or "none"}
══════════════════════════════════════════════════

If HARD_DISAGREE: Prefix with "⚠ GUARDIAN HARD DISAGREE" and show the Guardian's reasoning prominently.

If AWAITING_ZAC: End with "What's your call? (accept / override / dismiss)"