ai-council-orchestration

AI Council Orchestration — Multi-Model, All Providers

Every stage uses the best available model for its role. Model catalog is auto-discovered at runtime — no hardcoded model names that go stale. Works with GitHub Copilot, OpenCode Zen, Nvidia NIM, Ollama, Gemini, OpenAI, Groq, OpenRouter, and official Claude.

Quick Start

council-orchestrator models                   # Step 0: discover available models
council-orchestrator init "<your objective>"   # Step 1: start council
council-orchestrator status                    # Step 2: check stage

Then enter the loop below.

Model Discovery (Step 0)

Before entering the main loop, run:

council-orchestrator models

This queries http://127.0.0.1:4001/v1/models live and writes COUNCIL_MODELS.md with:

• All available models grouped by provider
• Recommended model for each council role
• Live indicators (⚡ = connected now)

If the proxy isn't running, fall back to the embedded catalog below.

Role-to-Model Mapping (Live)

Role	Best Picks (in priority order)	Selection Strategy
Thinker (deep reasoning)	copilot/claude-opus-4.6-1m → opencode/qwen3.7-max → opencode/deepseek-v4-pro → opencode/kimi-k2.6	Strongest analytical model available
Planner (task decomposition)	copilot/claude-sonnet-4.6 → opencode/qwen3.6-plus → opencode/minimax-m2.7	Best at structured planning
Creator (code + TDD)	copilot/gpt-5.4 → opencode/deepseek-v4-flash → opencode/minimax-m2.7 → copilot/grok-code-fast-1	Best code generation available
Critic (adversarial review)	copilot/claude-sonnet-4.6 → opencode/deepseek-v4-pro → opencode/kimi-k2.6	Strong at finding flaws
Reviewer (code review)	copilot/claude-sonnet-4.6 → opencode/qwen3.6-plus → opencode/minimax-m2.5	Balanced review quality
Verifier (fast checks)	copilot/claude-haiku-4.5 → copilot/gpt-5-mini → opencode/deepseek-v4-flash-free	Fast & cheap, uses FREE tier if available

Selection rule: Pick the first model from the priority list that is currently connected (⚡ in council-orchestrator models output). If none of the top picks are available, use any connected model — don't stall.

---

Embedded Model Catalog (Fallback)

If the proxy is unreachable, use this static reference:

GitHub Copilot (connected ✅)

Model ID	Capabilities	Best For
copilot/claude-opus-4.6-1m	Vision, 15x premium	★ Thinker, Critic
copilot/claude-sonnet-4.6	Vision	★ Planner, Reviewer, Critic
copilot/claude-sonnet-4.5	Vision	Reviewer, Planner
copilot/claude-haiku-4.5	Vision, 0.33x cost	★ Verifier
copilot/gpt-5.4	Vision	★ Creator
copilot/gpt-5.2	Vision	Creator
copilot/gpt-5-mini	Vision, FREE	Verifier, Critic, fallback
copilot/grok-code-fast-1	Fast coding	Creator (fast path)

OpenCode Zen (always available ✅)

Model ID	Context	Best For
opencode/minimax-m3	128K	All-rounder
opencode/minimax-m2.7	1M ctx	★ Planner (large codebases)
opencode/minimax-m2.5	1M ctx	Large context tasks
opencode/qwen3.7-max	128K	★ Thinker, Creator
opencode/qwen3.7-plus	128K	Creator
opencode/qwen3.6-plus	131K	★ Planner, Reviewer
opencode/qwen3.5-plus	131K	All-rounder
opencode/kimi-k2.6	131K	★ Thinker, Critic
opencode/kimi-k2.5	131K	Thinker, Critic
opencode/deepseek-v4-pro	65K	★ Thinker, Critic
opencode/deepseek-v4-flash	65K	★ Creator (fast)
opencode/glm-5.1	128K	All-rounder
opencode/glm-5	128K	All-rounder
opencode/mimo-v2.5-pro	262K	Large context
opencode/mimo-v2.5	262K	Large context
opencode/mimo-v2-pro	65K	General
opencode/mimo-v2-omni	65K	General
opencode/hy3-preview	131K	Preview

OpenCode Zen — FREE tier (always available ✅)

Model ID	Best For
opencode/deepseek-v4-flash-free	★ Verifier, fallback Creator
opencode/mimo-v2.5-free	Verifier, fallback
opencode/minimax-m3-free	Verifier, fallback
opencode/nemotron-3-super-free	Verifier, fallback

Nvidia NIM (if connected)

meta/llama-3.3-70b-instruct, meta/llama-3.1-8b-instruct, nvidia/llama-3.1-nemotron-70b-instruct, nvidia/nemotron-3-ultra-550b-a55b, mistralai/mistral-7b-instruct-v0.3

Ollama (local, if running)

qwen3:8b, qwen3:14b, llama3.3:70b

Google Gemini (if connected)

gemini-2.5-pro, gemini-2.5-flash, gemini-2.0-flash, gemini-1.5-pro, gemini-1.5-flash

OpenAI (if connected)

gpt-4o, gpt-4o-mini, o3-mini, o4-mini, gpt-4.1, codex-mini-latest

Groq (if connected)

llama-3.3-70b-versatile, llama-3.1-8b-instant, deepseek-r1-distill-llama-70b, mixtral-8x7b, gemma2-9b-it

OpenRouter (if connected)

google/gemma-3-27b-it:free, meta-llama/llama-3.3-70b-instruct:free, deepseek/deepseek-r1:free, qwen/qwen3-8b:free

Claude (Anthropic, official — if not using proxy)

claude-sonnet-4-6, claude-sonnet-4-5, claude-haiku-4-5, claude-opus-4-7, claude-opus-4-6, claude-opus-4-5

---

Architecture

STEP 0: council-orchestrator models     ← discover available models (live)

LOOP:
  1. council-orchestrator status        ← check current stage
  2. Select best model for stage         ← pick from connected providers
  3. Execute stage handler               ← uses embedded patterns below
  4. council-orchestrator advance/loopback ← update state
  5. GOTO step 1                        ← UNCONDITIONAL

BREAK ONLY when:
  - __delivery_check__ says done → DELIVER
  - __maxed_out__ safety limit → REPORT

Council Structure

Agent	Role	Model Selection Strategy
Thinker	Deep reasoning, ideation	Pick strongest analytical model connected
Planner	Task decomposition, file mapping	Best at structured breakdown
Creator	Implementation + TDD	Best code generator connected
Critic	Adversarial review	Strong analysis, find flaws
Reviewer	Code review	Balanced, thorough
Verifier	Fast final verification	Fastest/cheapest connected

---

Stage 1 — THINK

Embedded: Brainstorming Pattern

Model: Strongest analytical model connected (priority: copilot/claude-opus-4.6-1m → opencode/qwen3.7-max → opencode/deepseek-v4-pro → opencode/kimi-k2.6 → any connected)

1. Model selection: Run council-orchestrator models or check COUNCIL_MODELS.md. Pick the best Thinker model from what's connected.
2. Explore context & load helper skills — read project files, docs, recent commits. Also read and load all helper skills (skills/ponytail/SKILL.md, skills/ponytail-review/SKILL.md, skills/ponytail-audit/SKILL.md, skills/ponytail-debt/SKILL.md, skills/ponytail-gain/SKILL.md, skills/ponytail-help/SKILL.md, skills/loop/SKILL.md) to integrate their rules and capabilities into the session context.
3. Clarify & decompose — break objective into independent subsystems.
4. Propose 2-3 architectures with explicit trade-offs (adhering to Ponytail rules: YAGNI, standard library/native features first, no speculative abstractions).
5. Stress-test: What assumptions could be false? What could go wrong?
6. Spawn Thinker sub-agent → THOUGHT_REPORT.md
7. Spawn Critic sub-agent → CRITIQUE_REPORT.md
8. If concerns → council-orchestrator loopback think "<reason>" → GOTO LOOP
9. If clear → council-orchestrator advance think "approved" → GOTO LOOP

Stage 2 — PLAN

Embedded: Writing Plans Pattern

Model: Best planner connected (priority: copilot/claude-sonnet-4.6 → opencode/qwen3.6-plus → opencode/minimax-m2.7 → best connected)

1. Model selection: Pick the best Planner model.
2. Map the absolute minimum file structure needed. Avoid speculative helper files or interfaces.
3. Decompose into bite-sized tasks (2-5 min each).
4. Write TASK_EXECUTION_PLAN.md with real code in every step. No placeholders or boilerplate.
5. Self-review: spec coverage? placeholders? type consistency?
6. Spawn Critic: missing criteria? dependencies correct?
7. If concerns → loopback → GOTO LOOP
8. If clear → advance → GOTO LOOP

Stage 3 — CREATE

Embedded: TDD + Subagent-Driven Development + Parallel Dispatch Patterns

Model: Best coder connected (priority: copilot/gpt-5.4 → opencode/deepseek-v4-flash → opencode/minimax-m2.7 → copilot/grok-code-fast-1 → any connected)

1. Model selection: Pick the best Creator model. Verify current level using /ponytail.
2. TDD IRON LAW: No production code without a failing test first.
3. RED → Verify RED → GREEN (minimal implementation following Ponytail ladder: YAGNI → stdlib → native → one-line → minimum code) → Verify GREEN → REFACTOR (staying green, mark simplifications with ponytail: comments).
4. Dispatch fresh sub-agents per independent task. Ensure they are instructed to follow the Ponytail ladder.
5. Two-stage review per task: spec compliance → code quality.
6. Parallel dispatch for independent domains.
7. If missing capability → write pattern as skill.
8. If done → advance → GOTO LOOP
9. If issues → loopback → GOTO LOOP

Stage 4 — REVIEW & TEST

Embedded: Code Review + Systematic Debugging + Verification Patterns

Model: Best reviewer connected (priority: copilot/claude-sonnet-4.6 → opencode/deepseek-v4-pro → opencode/qwen3.6-plus → any connected)

1. Model selection: Pick the best Reviewer model.
2. Pre-review: get SHAs, summary of what was built.
3. Spawn ALL council roles to review simultaneously. Critic must run a ponytail-review for over-engineering (tags: delete, stdlib, native, yagni, shrink) and report net lines removable. Additionally, run the /ponytail-review command (or ponytail-review skill) directly on the current git diff to harvest a concrete delete-list.
4. If flaws → Systematic Debugging (4-phase: root cause → pattern → hypothesis → fix)
5. IRON LAW: No fixes without root cause investigation.
6. Fix → re-verify → loopback review → GOTO LOOP
7. When clean → advance → GOTO LOOP

Stage 5 — VERIFY & DELIVER

Embedded: Verification Before Completion + Finishing Branch Patterns

Model: Fastest/cheapest connected (priority: copilot/claude-haiku-4.5 → copilot/gpt-5-mini → opencode/deepseek-v4-flash-free → any FREE model)

1. Model selection: Pick the cheapest available model — verification is simple checks.
2. IRON LAW: No "it works" without fresh verification output.
3. Run full test suite, build, integration.
4. Spawn completeness verifier.
5. Run the /ponytail-debt command (or ponytail-debt skill) to harvest any deferred shortcuts into PONYTAIL-DEBT.md.
6. Produce VERIFICATION_SIGN_OFF.md.
7. If verified → advance → GOTO LOOP
8. If not → loopback to appropriate stage → GOTO LOOP

Delivery Check

When stage is __delivery_check__:

• If objective satisfied → DELIVER. STOP THE LOOP.
• If not → council-orchestrator next-iteration → GOTO LOOP

Standing Directives

#	Directive	Rule
1	NEVER STOP	No user input needed. Resolve blockers autonomously.
2	GOTO LOOP step 1	After every action, immediately check status
3	TDD always	No production code without a failing test first
4	Verify before claiming	Run command, check fresh exit code & output
5	Root cause before fix	No symptom fixes without investigation
6	Safety limit: 50 iterations	Loop terminates to prevent runaway tokens
7	Auto-discover models	Refresh list with council-orchestrator models
8	Follow Ponytail rules	YAGNI → stdlib → native → one line → minimum. Mark simplifications with ponytail: comments

Activation

# Step 0 — Discover models (run once per session)
council-orchestrator models

# Step 1 — Initialize
council-orchestrator init "<full objective>"

# Step 2 — Enter loop
council-orchestrator status

The council reads COUNCIL_MODELS.md, picks the best model per role from what's actually connected, and executes each stage with its embedded pattern. The loop turns until done.