@neocage/council 1.3.0

package/CHANGELOG.md

@@ -0,0 +1,75 @@
+# Changelog
+
+All notable changes to this project will be documented in this file.
+Format follows [Keep a Changelog](https://keepachangelog.com/en/1.0.0/).
+This project uses [Semantic Versioning](https://semver.org/spec/v2.0.0.html).
+
+## [Unreleased]
+
+## [1.3.0] - 2026-09-09
+
+### Added
+- npm packaging — `package.json` (scoped `@neocage/council`, public access) and an executable `bin/council.js` wrapper around `install.sh`, so the project is installable via `npx @neocage/council`.
+- `.github/workflows/publish.yml` — publishes to npmjs.org and GitHub Packages on `v*.*.*` tag pushes; refuses to publish if the tag and `package.json` version disagree.
+
+## [1.2.0] - 2026-07-04
+
+### Added
+- **Claude Code plugin-marketplace support** (#45) — `.claude-plugin/plugin.json` + `marketplace.json`; install with `/plugin marketplace add NeoCage/Council-of-Cognitive-Excellence` then `/plugin install council@council-of-cognitive-excellence`. Root `SKILL.md` remains the single source of truth via a `skills/council/` symlink; a new Asset Resolution section resolves agents/scripts/configs from either the install.sh layout or `${CLAUDE_PLUGIN_ROOT}`.
+- **Confidence-weighted vote tally** (#44) — a member's vote weight is now base weight (1.0, or 1.5× domain seat) × confidence factor from their `STANCE:` line (`high 1.0 / med 0.75 / low 0.5`). `W_total` stays at full base weights, so a hesitant council raises its own consensus bar and escalates to the user rather than forcing a verdict (Roundtable Policy, arXiv:2509.16839; ConfMAD, arXiv:2509.14034).
+- **Per-member `reasoning_method` frontmatter** (E3 / DMAD, arXiv:2410.12853) (#44) — all 18 personas declare a distinct reasoning method (elenchic-questioning, tail-stress-testing, causal-loop-mapping, …) and Round 1 prompts direct each member to reason via their method: method diversity, not just persona diversity.
+- **Project-level `./.council.yaml` overrides** (E7) (#44) — pin `profile`/`triad`/`members`/`chairman`/`models`/`no_auto_route` per project; CLI flags always win.
+- Standard repo furniture (#42) — `.gitignore`, `SECURITY.md`, `FUNDING.yml`, bug/feature/provider issue templates, and a PR template with a three-file protocol-parity checklist.
+- **Structured stance voting + deterministic weighted tie-break.** The final round (full Round 3 / quick Round 2) now requires each member to emit a machine-parseable `STANCE: <option> | CONFIDENCE: … | DEALBREAKER: …` line, so consensus is a counted weighted tally rather than a prose impression. STEP 6 specifies the exact math: `W_option ≥ (2/3) × W_total`, with the on-domain seat carrying 1.5×. New **Vote Tally** field in the full and quick verdict templates records `option → weight` and which seat carried the weight. Mirrored in `SKILL.codex.md`.
+- Cursor CLI provider support (`cursor_cli` archetype) — sixth dispatch path alongside subagent / codex_exec / gemini_cli / ollama_run / openai_compatible_api. Auto-detected via the `cursor-agent` binary; members run headless and read-only (`cursor-agent -p --mode ask --model <id>`). Cursor is a model aggregator (GPT-5.x / Claude / Gemini / Grok through one CLI), so routing treats it as a single provider for spread and steers diversity seats to cross-family models to avoid duplicating Anthropic bias. New `configs/provider-model-slots.cursor.example.yaml`, `cursor_cli` tiers in `configs/auto-route-defaults.yaml`, and `cursor_cli` as a valid `--chairman` tag.
+
+### Changed
+- **Domain-weight seat (1.5×) is now designated at panel selection (STEP 0), before any positions exist** — previously chosen at tie-break time, which let the coordinator pick the heavyweight after seeing votes. Locking it up front removes that nudge.
+- README "Enforcement Mechanisms" rewritten to state the actual forcing function (bounded round budget + anti-recursion guards) and that genuine splits escalate to the user instead of being forced into consensus.
+- **CI drift guards are now real** (#42) — `council-simulation-checklist.sh` executes in CI (it never ran before), extended with full `SKILL.gemini.md` parity checks and a three-file STANCE/Vote-Tally/1.5×/2-3-threshold parity gate; markdownlint's variadic `--disable` was missing its trailing `--` and silently linted zero files — rule selection now lives in `.markdownlint.json` and covers all doc files.
+- `install.sh` derives the generated `gemini-extension.json` version from CHANGELOG instead of a hardcoded string; model IDs reconciled across configs/docs (`gemini-3-pro`, `gpt-5.4` generation); README/CLAUDE.md document the `--gemini` install path and marketplace install.
+
+### Fixed
+- **`SKILL.gemini.md` drift regression** (#42) — the Gemini coordinator was missing structured stance voting, the weighted 2/3 tie-break, and the Vote Tally verdict field entirely; `SKILL.codex.md` referenced a Vote Tally field its own verdict template lacked.
+- `SKILL.md` `classic` profile said "All 11 members" — the roster is 18 (#42).
+- Shell-injection surface in external dispatch templates — prompts now route through quoted heredocs instead of being inlined into double-quoted argv; API keys no longer appear in `ps`-visible curl argv (#42).
+- Tracked `.DS_Store` and a stray `.gstack` error log removed; junk excluded from release tarballs (#42).
+
+## [1.1.0] - 2026-05-21
+
+### Added
+- `SKILL.codex.md` — dedicated Codex council coordinator
+- Codex install support in `install.sh` (`--codex`, `--codex-only` flags) with reliability hardening
+- NVIDIA NIM provider support (`configs/provider-model-slots.nim.example.yaml`) — auto-detection via `NVIDIA_API_KEY`
+- Round 2 cross-examination anonymization — peer Round 1 outputs are masked behind stable `Member A/B/C` labels in full and quick modes (Choi et al., arXiv:2510.07517; Karpathy `llm-council`)
+- Anti-conformity directive — Round 2 prompts in all three modes now require members to name the specific flaw in their earlier argument before updating; defends correct prior positions against social pressure (Cui et al., Free-MAD, arXiv:2509.11035)
+- Explicit Chairman role — synthesis (STEP 7, QUICK STEP 3, DUO STEP 4) is performed by a named model selected via STEP 1.7; new `--chairman <name>` flag, `chairman_defaults:` block in `configs/auto-route-defaults.yaml`, and a hard constraint that Chairman cannot be a panel member
+- Verdict actionability sections — `Acceptable Compromises`, `Kill Criteria`, `Concrete Next Step` are now required in every verdict (full mode); quick and duo modes require subsets per the per-mode policy
+- `openai_compatible_api` provider archetype — fifth dispatch path alongside subagent / codex_exec / gemini_cli / ollama_run; routes NIM seats (and future Together / Fireworks / vLLM) via `/chat/completions` with credentials resolved from `api_key_env` at runtime
+- Session Metadata block (`schema_version: 1`) — appended to every verdict with `mode`, `panel_size`, `rounds_run`, `tools_used`, `provider_count`, `fallbacks_triggered`, plus best-effort token / duration estimates
+- CI hardening — `.gitattributes` for LF normalization, `.github/workflows/lint.yml` (shellcheck + markdownlint), `.github/workflows/release.yml` (tarball + auto release notes from CHANGELOG on `v*.*.*` tag), `CHANGELOG.md` itself
+
+### Changed
+- README updated with header image, quickstart, and open-source best practices
+
+### Fixed
+- Dead code in `scripts/council-simulation-checklist.sh` that tripped ShellCheck SC2317 under CI
+
+## [1.0.0] - 2026-03-30
+
+### Added
+- 18 council member personas: Aristotle, Socrates, Sun Tzu, Ada Lovelace, Marcus Aurelius, Machiavelli, Lao Tzu, Feynman, Torvalds, Musashi, Watts, Karpathy, Sutskever, Kahneman, Meadows, Munger, Taleb, Rams
+- 3-round structured deliberation protocol: Problem Restate Gate → Blind Analysis → Cross-Examination → Crystallization → Verdict
+- Post-round enforcement scan: dissent quota, novelty gate, agreement check (>70% triggers mandatory counterfactual), evidence labeling, anti-recursion rule
+- Quick mode (2-round), Duo mode (2-member deliberation), pre-defined triads by domain
+- Multi-provider auto-detection (`scripts/detect-providers.sh`) — routes council members across Claude, Codex (OpenAI), and Ollama
+- Execution profiles (`configs/auto-route-defaults.yaml`) and simulation checklist (`scripts/council-simulation-checklist.sh`)
+- Provider model slot template (`configs/provider-model-slots.example.yaml`)
+
+### Fixed
+- Default OpenAI model for Codex set to `gpt-5.4` (E2E test found o3/o4-mini unavailable on standard ChatGPT accounts)
+
+[Unreleased]: https://github.com/NeoCage/Council-of-Cognitive-Excellence/compare/v1.2.0...HEAD
+[1.2.0]: https://github.com/NeoCage/Council-of-Cognitive-Excellence/releases/tag/v1.2.0
+[1.1.0]: https://github.com/NeoCage/Council-of-Cognitive-Excellence/releases/tag/v1.1.0
+[1.0.0]: https://github.com/NeoCage/Council-of-Cognitive-Excellence/releases/tag/v1.0.0

package/LICENSE

@@ -0,0 +1,22 @@
+MIT License
+
+Copyright (c) 2026 nyk
+Modifications copyright (c) 2026 NeoCage
+
+Permission is hereby granted, free of charge, to any person obtaining a copy
+of this software and associated documentation files (the "Software"), to deal
+in the Software without restriction, including without limitation the rights
+to use, copy, modify, merge, publish, distribute, sublicense, and/or sell
+copies of the Software, and to permit persons to whom the Software is
+furnished to do so, subject to the following conditions:
+
+The above copyright notice and this permission notice shall be included in all
+copies or substantial portions of the Software.
+
+THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
+IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
+FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
+AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
+LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
+OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
+SOFTWARE.

package/README.md

@@ -0,0 +1,258 @@
+# Council of Cognitive Excellence
+
+<p align="center">
+  <img src="assets/header.png" alt="Council of Cognitive Excellence" width="800">
+</p>
+
+<p align="center">
+  Structured multi-perspective deliberation for decisions that deserve more than one reasoning path.
+</p>
+
+<p align="center">
+  <a href="https://github.com/NeoCage/Council-of-Cognitive-Excellence/actions/workflows/lint.yml"><img src="https://github.com/NeoCage/Council-of-Cognitive-Excellence/actions/workflows/lint.yml/badge.svg" alt="Lint"></a>
+  <a href="https://github.com/NeoCage/Council-of-Cognitive-Excellence/releases"><img src="https://img.shields.io/github/v/release/NeoCage/Council-of-Cognitive-Excellence" alt="Release"></a>
+  <a href="LICENSE"><img src="https://img.shields.io/badge/license-MIT-blue" alt="License"></a>
+</p>
+
+<p align="center">
+  Claude Code · Codex · Gemini CLI · OpenCode
+</p>
+
+The council assigns a hard question to deliberately different analytical personas, keeps
+their first positions independent, forces direct disagreement, and returns a verdict that
+preserves unresolved questions, dissent, kill criteria, and the next concrete action.
+
+## Try a council
+
+Claude Code users can install the plugin directly:
+
+```text
+/plugin marketplace add NeoCage/Council-of-Cognitive-Excellence
+/plugin install council@council-of-cognitive-excellence
+```
+
+Then convene the full council, a faster panel, or a two-member dialectic:
+
+```text
+/council Should we open-source our agent framework?
+/council --quick Should we add caching here?
+/council --duo Should we use microservices or a monolith?
+```
+
+For Codex, Gemini CLI, or OpenCode, clone once and use the matching installer flag:
+
+```bash
+git clone https://github.com/NeoCage/Council-of-Cognitive-Excellence.git
+cd council-of-cognitive-excellence
+
+./install.sh --codex-only
+./install.sh --gemini-only
+./install.sh --opencode-only
+```
+
+Restart the target client after installation. The command remains `/council` on every
+supported host.
+
+## Use it for the decision boundary
+
+Good council questions have material downside, competing values, incomplete evidence, or
+an irreversible choice. Write down the decision, constraints, evidence, reversibility, and
+deadline before adding personas.
+
+![Decision field notes](assets/decision-field-notes.png)
+
+The field notes distinguish facts from inference, assumptions, and unknowns. They also make
+it harder to use a long deliberation as decoration for a decision already made.
+
+### When to use something smaller
+
+- Use a direct answer or primary documentation for factual lookups.
+- Run an experiment when the choice is cheap and reversible.
+- Use `--quick` when cross-examination will not change the outcome.
+- Use `--duo` when one tension matters more than broad coverage.
+- Do not convene a council to manufacture support for a preferred answer.
+
+## Choose a mode
+
+| Mode | Shape | Use it when |
+|---|---|---|
+| Full | Independent analysis, cross-examination, final stance, synthesis | Stakes are high and competing frames need contact |
+| Quick | Restate, rapid analysis, final positions | The decision needs breadth but not a full adversarial round |
+| Duo | Opening positions, direct response, final statements | One polarity defines the decision |
+
+![Council mode selector](assets/mode-selector.png)
+
+```text
+/council --full What is the right pricing model?
+/council --quick --triad shipping Should we release today?
+/council --duo --members torvalds,ada Is this abstraction worth it?
+```
+
+Named triads select three relevant lenses without assembling the whole council:
+
+```text
+/council --triad strategy Where is our defensible advantage?
+/council --triad risk What could make this launch irreversible?
+/council --triad ai-product Which capability belongs in the product?
+```
+
+Available domains include `architecture`, `strategy`, `ethics`, `debugging`, `risk`,
+`shipping`, `product`, `founder`, `ai`, `ai-product`, `ai-safety`, `decision`, `systems`,
+`uncertainty`, `design`, `economics`, and `bias`. The canonical routing table lives in
+[`SKILL.md`](SKILL.md).
+
+## What the protocol protects
+
+The full protocol has a fixed round budget. Members first restate the problem, analyze
+blind, cross-examine other positions, declare a final stance, and then enter synthesis.
+Enforcement checks look for premature agreement, repeated claims, missing dissent, and
+unsupported confidence.
+
+![Verdict blueprint](assets/verdict-blueprint.png)
+
+Verdicts lead with what remains unresolved. A recommendation is paired with acceptable
+compromises, kill criteria, and one concrete next step. When the weighted tally remains
+split, the result returns that split to the user instead of turning it into prose consensus.
+
+Evidence labels separate:
+
+- `FACT`: directly supported by supplied or retrieved evidence;
+- `INFERENCE`: follows from evidence but is not directly observed;
+- `ASSUMPTION`: required for the argument and still unverified;
+- `UNKNOWN`: missing information that could change the decision.
+
+## Track the outcome
+
+A verdict is useful only if it can be checked later. Before acting, record the prediction,
+owner, review date, and evidence that would change the recommendation. At the checkpoint,
+mark the result confirmed, revised, reversed, or inconclusive instead of rewriting the
+original rationale.
+
+![Council outcome ledger](assets/outcome-ledger.png)
+
+## The 18 lenses
+
+Members are analytical instruments, not impersonation claims. Each persona has a grounding
+protocol, a method, known blind spots, and a structured response contract.
+
+![Council panel composition](assets/panel-composition.png)
+
+| Member | Primary lens | Useful counterweight |
+|---|---|---|
+| Aristotle | Categories and structure | Lao Tzu on emergence and excess structure |
+| Socrates | Assumption destruction | Feynman on reconstruction from first principles |
+| Sun Tzu | Terrain and adversarial strategy | Aurelius on internal control and moral cost |
+| Ada Lovelace | Formal systems and abstraction | Machiavelli on incentives and informal power |
+| Marcus Aurelius | Resilience and moral clarity | Sun Tzu on external competition |
+| Machiavelli | Power and incentives | Ada on formal consistency |
+| Lao Tzu | Non-action and emergence | Aristotle on explicit categories |
+| Richard Feynman | Explanation and empirical debugging | Socrates on the premise itself |
+| Linus Torvalds | Shipping and maintainability | Meadows on system-level consequences |
+| Miyamoto Musashi | Timing and decisive action | Torvalds on acting before the ideal moment |
+| Alan Watts | Reframing and false problems | Torvalds on concrete implementation |
+| Andrej Karpathy | Empirical ML behavior | Sutskever on frontier risk |
+| Ilya Sutskever | Scaling and AI safety | Karpathy on observation and iteration |
+| Daniel Kahneman | Cognitive bias | Feynman on explicit causal reasoning |
+| Donella Meadows | Feedback loops and high-impact interventions | Torvalds on local fixes |
+| Charlie Munger | Inversion and model lattices | Aristotle on single-system classification |
+| Nassim Taleb | Tail risk and fragility | Karpathy on smooth empirical trends |
+| Dieter Rams | User clarity and restraint | Ada on what can be formalized |
+
+The repository stores each member contract under [`agents/`](agents/). Custom panels can
+use `--members`, named triads, or the `classic`, `exploration-orthogonal`, and
+`execution-lean` profiles documented in [`SKILL.md`](SKILL.md).
+
+## Multi-provider routing
+
+The detection script checks which supported providers are available, then the coordinator
+distributes seats across them. Polarity pairs are separated when possible so a single model
+family does not play both sides of a disagreement.
+
+![Council provider routing](assets/provider-routing.png)
+
+| Provider path | Detection |
+|---|---|
+| Native host subagents | Available through the active supported client |
+| OpenAI | `codex` executable |
+| Google | `gemini` executable |
+| Ollama | `ollama` executable |
+| NVIDIA NIM | `NVIDIA_API_KEY` |
+| Cursor | `cursor-agent` executable or configured login |
+
+Preview routing without running a council:
+
+```text
+/council --dry-route --triad decision Should we accept this acquisition offer?
+```
+
+Use `--no-auto-route` to keep native-host defaults. Use `--models <path>` with a copy of
+[`configs/provider-model-slots.example.yaml`](configs/provider-model-slots.example.yaml)
+for an explicit seat map. Provider failure is reported before the seat falls back to the
+native host.
+
+## Installation reference
+
+```bash
+./install.sh                         # Claude Code
+./install.sh --codex                 # Claude Code + Codex
+./install.sh --codex-only            # Codex only
+./install.sh --gemini                # Claude Code + Gemini CLI
+./install.sh --gemini-only           # Gemini CLI only
+./install.sh --opencode              # Claude Code + OpenCode
+./install.sh --opencode-only         # OpenCode only
+./install.sh --copy-configs          # Include provider-routing templates
+./install.sh --dry-run               # Preview writes
+```
+
+Custom target directories are supported through `--claude-dir`, `--codex-dir`,
+`--gemini-dir`, and `--opencode-dir`. Run `./install.sh --help` for the current contract.
+
+## Verify the checkout
+
+```bash
+./scripts/council-simulation-checklist.sh
+./install.sh --dry-run --codex
+./install.sh --dry-run --gemini
+./install.sh --dry-run --opencode
+```
+
+The checklist validates persona structure, host-protocol parity, routing configuration,
+execution checkpoints, verdict fields, and installer behavior. Use the
+[demo session pack](demos/session-pack.md) to exercise full, quick, and duo modes.
+
+## Repository map
+
+```text
+SKILL.md                    canonical coordinator protocol
+SKILL.codex.md              Codex host mirror
+SKILL.gemini.md             Gemini CLI host mirror
+SKILL.opencode.md           OpenCode host mirror
+agents/                     18 grounded persona contracts
+configs/                    provider and model-routing examples
+demos/                      sample sessions and verdict template
+scripts/                    detection, conversion, and validation tools
+assets/                     brand system and README visuals
+install.sh                  multi-host installer
+```
+
+Protocol changes start in `SKILL.md` and must remain behaviorally aligned with each host
+mirror. See [CONTRIBUTING.md](CONTRIBUTING.md) for the required checks and review contract.
+Security reports belong in [SECURITY.md](SECURITY.md), not a public issue.
+
+## Project links
+
+- [Release history](CHANGELOG.md)
+- [Brand kit](assets/BRAND.md)
+- [Issues](https://github.com/NeoCage/Council-of-Cognitive-Excellence/issues)
+- [Maintainer: NeoCage](https://github.com/NeoCage)
+
+<p align="center">
+  <picture>
+    <source media="(prefers-color-scheme: dark)" srcset="assets/star-history-dark.svg">
+    <img src="assets/star-history-light.svg" alt="Council of Cognitive Excellence star history" width="600">
+  </picture>
+</p>
+
+## License
+
+[MIT](LICENSE)

package/SKILL.codex.md

@@ -0,0 +1,250 @@
+---
+name: council
+description: "Convene the Council of Cognitive Excellence in Codex when the user asks for /council, council deliberation, triads, duo debates, or multi-perspective decision analysis."
+---
+
+# /council for Codex
+
+You are the Council Coordinator. Run structured multi-persona deliberation using the council agent files.
+
+## Invocation Patterns
+
+```
+/council [problem]
+/council --quick [problem]
+/council --duo [problem]
+/council --triad [domain] [problem]
+/council --members socrates,feynman,ada [problem]
+/council --profile exploration-orthogonal [problem]
+```
+
+## Flags
+
+| Flag | Effect |
+|------|--------|
+| `--full` | Use all 18 members |
+| `--quick` | 2-round fast mode |
+| `--duo` | 2-member polarity dialectic |
+| `--triad [domain]` | Use predefined 3-member panel |
+| `--members a,b,c` | Use explicit member names |
+| `--profile [name]` | Use profile panel (`classic`, `exploration-orthogonal`, `execution-lean`) |
+
+If no panel flag is present, auto-select the best triad from problem context.
+
+## Member Roster
+
+`aristotle, socrates, sun-tzu, ada, aurelius, machiavelli, lao-tzu, feynman, torvalds, musashi, watts, karpathy, sutskever, kahneman, meadows, munger, taleb, rams`
+
+## Triads
+
+| Domain | Members |
+|--------|---------|
+| `architecture` | aristotle, ada, feynman |
+| `strategy` | sun-tzu, machiavelli, aurelius |
+| `ethics` | aurelius, socrates, lao-tzu |
+| `debugging` | feynman, socrates, ada |
+| `innovation` | ada, lao-tzu, aristotle |
+| `conflict` | socrates, machiavelli, aurelius |
+| `complexity` | lao-tzu, aristotle, ada |
+| `risk` | sun-tzu, aurelius, feynman |
+| `shipping` | torvalds, musashi, feynman |
+| `product` | torvalds, machiavelli, watts |
+| `founder` | musashi, sun-tzu, torvalds |
+| `ai` | karpathy, sutskever, ada |
+| `ai-product` | karpathy, torvalds, machiavelli |
+| `ai-safety` | sutskever, aurelius, socrates |
+| `decision` | kahneman, munger, aurelius |
+| `systems` | meadows, lao-tzu, aristotle |
+| `uncertainty` | taleb, sun-tzu, sutskever |
+| `design` | rams, torvalds, watts |
+| `economics` | munger, machiavelli, sun-tzu |
+| `bias` | kahneman, socrates, watts |
+
+## Profiles
+
+- `classic`: all 18 members
+- `exploration-orthogonal`: socrates, feynman, sun-tzu, machiavelli, ada, lao-tzu, aurelius, torvalds, karpathy, sutskever, kahneman, meadows
+- `execution-lean`: torvalds, feynman, sun-tzu, aurelius, ada
+
+## Execution Protocol
+
+### Step 1: Locate Council Assets
+
+Resolve council files in this order:
+
+1. `~/.codex/skills/council/agents/`
+2. `./agents/`
+
+If neither exists, stop and tell the user to run `./install.sh --codex`.
+
+### Step 2: Parse Request
+
+Project overrides: if `./.council.yaml` exists in the working directory, treat its keys (`profile`, `triad`, `members`, `chairman`, `models`, `no_auto_route`) as default flag values. Explicit flags always win.
+
+Extract:
+
+- Mode: `full` (default), `quick`, or `duo`
+- Problem statement
+- Panel selection via `--members`, `--triad`, `--profile`, or `--full`
+
+For `--duo` without explicit members, choose a polarity pair from keywords:
+
+- architecture/structure: `aristotle` + `lao-tzu`
+- shipping/execution: `torvalds` + `musashi`
+- strategy/competition: `sun-tzu` + `aurelius`
+- ai/ml/model: `karpathy` + `sutskever`
+- decision/bias: `kahneman` + `feynman`
+- default fallback: `socrates` + `feynman`
+
+### Step 2.5: Runtime Reliability Defaults
+
+Use these defaults unless the user requests stricter/faster behavior:
+
+- `spawn_timeout_ms`: 45000 per member
+- `round_timeout_ms`: 60000 for quick/duo, 90000 for full
+- `retry_attempts`: 2 retries after initial attempt (max 3 total attempts per seat per round)
+- `retry_backoff_sec`: 2, then 5
+- `hard_min_live_seats`: 2
+
+Track seat state per member:
+
+- `live`: normal agent responses
+- `degraded`: agent timed out/failed and is being simulated from persona file
+- `offline`: could not recover enough information for this seat
+
+### Step 3: Run Restatement Gate (Parallel)
+
+Spawn one sub-agent per selected member with `spawn_agent`, `fork_context=true`.
+
+Prompt template:
+
+```
+Read and follow this persona file exactly: {agent_file_path}
+
+Problem:
+{problem}
+
+Return only:
+1) Your restatement (one sentence)
+2) Alternative framing (one sentence)
+Maximum 50 words total.
+```
+
+Wait with `spawn_timeout_ms`. If a seat fails or times out:
+
+1. Retry spawn up to `retry_attempts` using backoff.
+2. If still failing, set seat to `degraded` and produce a `[Simulated]` restatement from that persona file.
+3. If persona file cannot be read, mark seat `offline`.
+
+If live seats drop below `hard_min_live_seats`, switch to fully simulated mode for all seats and state this explicitly.
+
+### Step 3.5: External Seats (HTTP and CLI archetypes)
+
+Some provider archetypes are dispatched outside the host runtime's `spawn_agent`. Anonymization (Step 4) and Chairman selection (Step 5) apply equally to these seats — no special-case logic.
+
+**`openai_compatible_api` (NVIDIA NIM today; Together / Fireworks / vLLM in the future)** — dispatch via HTTP:
+
+- Read `base_url` and `api_key_env` from the seat config (or detection JSON for auto-routing).
+- Resolve the API key from the env var at routing time. Never inline.
+- POST to `{base_url}/chat/completions` with an OpenAI-compatible payload (system+user messages, `temperature: 0.7`, `max_tokens: 1200`).
+- Extract `.choices[0].message.content`. If empty or non-2xx, mark the seat `degraded` and apply the standard fallback (anthropic per the agent's `model` frontmatter).
+- Per-seat timeout: 90 seconds (hosted open-weight endpoints are slower than first-party APIs).
+
+**`cursor_cli` (Cursor)** — dispatch via subprocess. Cursor is a model aggregator: one binary (`cursor-agent`) serves GPT-5.x, Claude, Gemini, and Grok families.
+
+- Run headless and read-only: `cursor-agent -p --mode ask --model {model} --output-format text "{full prompt}"`.
+- Auth is resolved by the CLI itself (prior `cursor-agent login` or `CURSOR_API_KEY`). Never inline a key. On auth error, mark the seat `degraded` and apply the standard fallback.
+- Empty stdout or non-zero exit → `degraded` + fallback. Per-seat timeout: 90 seconds.
+- Counts as a single provider for spread. Because Cursor can serve `claude-*` models, prefer cross-family models (`gpt-*`, `gemini-*`, `grok-*`) for any seat opposite a native `anthropic` seat in a polarity pair. Verify live IDs with `cursor-agent --list-models`.
+
+### Step 4: Deliberation Rounds
+
+Keep the same spawned agents for all rounds via `send_input`.
+
+**Round 2 anonymization (full and quick modes).** Before sending Round 2 prompts in full or quick mode, build a stable label mapping `Member A` → first panel member, `Member B` → second, …, rewrite each Round 1 output's header to its label, strip in-body self-attribution, and instruct each agent that identities are masked and they must reference peers by label only. Retain the mapping privately in coordinator state and restore it for Round 3, tie-breaking, and the verdict. Duo mode is exempt (only two members; identity cannot be masked by elimination). Rationale: Choi et al. (arXiv:2510.07517) and Karpathy `llm-council` — identity labels in peer-review prompts drive conformity/self-bias.
+
+**Anti-conformity directive (Round 2, all modes).** When sending Round 2 prompts, include this paragraph verbatim before the per-mode instructions:
+
+> Anti-conformity directive. If your Round 1 position was correct, defend it. Do not update merely because peers disagree, because consensus is forming, or because a position is repeated by multiple members. Update only when presented with sound, validity-aligned reasoning that exposes a specific flaw in your earlier argument. Naming that flaw is required when you update; if you cannot name it, you should not update.
+
+Rationale: Choi et al. (arXiv:2510.07517), Free-MAD (arXiv:2509.11035), controlled-study arXiv:2511.07784 — generic "be critical" instructions underperform; the load-bearing piece is the "name-the-flaw" requirement that converts disposition into verifiable behavior.
+
+Round 1 prompts must instruct each member to reason via the reasoning_method field in their frontmatter (DMAD, arXiv:2410.12853) — method diversity, not just persona diversity.
+
+Full mode:
+
+1. Round 1: Independent analysis, blind-first, max 300 words/member.
+2. Round 2: Cross-examination with **anonymized** peer outputs + anti-conformity directive, max 220 words/member, each member engages at least 2 peers by Member-X label.
+3. Round 3: Final position, max 100 words/member. Real names restored.
+
+Quick mode:
+
+1. Round 1: Restate + rapid analysis, max 200 words/member.
+2. Round 2: Final position with **anonymized** peer outputs + anti-conformity directive, max 75 words/member. Real names restored in the verdict.
+
+Duo mode:
+
+1. Round 1: Opening position, max 250 words/member.
+2. Round 2: Direct response to counterpart with anti-conformity directive, max 180 words/member. (No anonymization — see rationale above.)
+3. Round 3: Final statement, max 60 words/member.
+
+(Round word caps above are deliberately tighter than `SKILL.md`'s Claude-host caps — this host runs a compressed protocol. The caps in this file are canonical when coordinating from Codex.)
+
+Structured stance & weighted tie-breaking (full + quick modes):
+
+1. **Designate the domain-weight seat at panel selection** (before any analysis): the single member whose domain most directly matches the problem carries **1.5×** weight; all others **1.0×**. Lock it up front — selecting it after seeing positions would let the coordinator nudge the outcome. If the match is ambiguous, designate none and tie-break on equal weights.
+2. The final round (full Round 3 / quick Round 2) MUST end each member's output with a structured stance line: `STANCE: <short option label> | CONFIDENCE: high|med|low | DEALBREAKER: yes|no`. Members reuse the same label where they agree; `STANCE: abstain` if backing no option. Re-prompt for a missing/unparseable line — never infer stance from prose.
+3. Tally **confidence-weighted** votes per canonical option (Roundtable Policy arXiv:2509.16839; ConfMAD arXiv:2509.14034): vote weight = base weight (1.0, or 1.5 for the domain seat) × confidence factor (`high → 1.0`, `med → 0.75`, `low → 0.5`). Consensus iff `W_option ≥ (2/3) × W_total`, where `W_total` sums **base** weights including abstainers' (abstention and low confidence both raise the bar — a hesitant council escalates instead of forcing a verdict). Highest option clearing the bar wins; `DEALBREAKER: yes` dissent goes in the Minority Report regardless.
+4. No option clears 2/3 → genuine split: do NOT force consensus and do NOT add a round (the spent round budget is the forcing function). Present each option with its weighted tally to the user. Record the tally (`option → weight`, marking the 1.5× seat) in the verdict's Vote Tally field. Duo mode issues no tally — it is dialectic, not decision-issuing.
+
+Round execution reliability policy:
+
+1. Send prompts to all `live` seats in parallel.
+2. Wait using `round_timeout_ms`.
+3. For each missing response, retry `send_input` up to `retry_attempts` with a stricter prompt: "Respond now in <= {word_limit} words."
+4. If still missing, move seat to `degraded` and generate `[Simulated]` output from persona instructions plus prior round context.
+5. Carry `degraded` seats forward for remaining rounds unless the seat recovers.
+6. If live seats drop below `hard_min_live_seats`, complete remaining rounds in fully simulated mode and mark confidence lower.
+
+### Step 5: Synthesis Output (CHAIRMAN)
+
+Synthesis is performed by an explicit **Chairman** — a model that did NOT deliberate in Rounds 1–3. The Chairman is selected before Round 1 using this algorithm (first match wins):
+
+1. **Explicit override**: `--chairman <name>` was passed (provider tag — `anthropic`, `openai`, `google`, `ollama`, `nvidia_nim`, `cursor_cli` — or a model alias).
+2. **Auto-select**: highest-tier model among available providers, **preferring one not on the panel** when possible. Tie-breaker: provider listed first by the host runtime.
+3. **Single-provider fallback**: use that provider's highest tier and note the overlap in the verdict.
+
+The Chairman is dispatched as a single call with the full audit transcript (Round 2 de-anonymized using the mapping retained in coordinator state — see Step 4 anonymization). Constraint: Chairman MUST NOT be a deliberating member in the same session.
+
+Return a verdict with this order, produced by the Chairman:
+
+1. `Selected Panel` (members + mode)
+2. `Chairman` (name, provider, model, selection rationale)
+3. `Acceptable Compromises` — what this verdict gives up, named explicitly (required in full; optional in quick; encouraged in duo)
+4. `Kill Criteria` — observable conditions that would falsify the verdict; format `"If <X> by <date>, invalidated → <Y>"` (required in full and quick; encouraged in duo)
+5. `Concrete Next Step` — exactly one action with an artifact-producing verb (required in all modes)
+6. `Unresolved Questions`
+7. `Key Agreements`
+8. `Key Disagreements`
+9. `Vote Tally` — the weighted stance tally from Step 4: one line per option `<option> — <weight> (<backers>)`, marking the 1.5× domain-weight seat, with the 2/3 threshold and whether it was cleared (full + quick modes; duo issues no tally)
+10. `Decision Options` (2-4 options with tradeoffs)
+11. `Recommended Next Steps` (additional actions beyond Concrete Next Step; ordered)
+12. `Confidence` (high/medium/low + why)
+13. `Execution Reliability` (live/degraded/offline seat counts and any timeout caveats)
+
+Always preserve dissent. Never flatten disagreements into fake consensus. Sections 3-5 are non-negotiable in full mode — they make the verdict operational (observable, falsifiable, actionable) instead of advisory prose.
+
+**Chairman fallback**: if the Chairman call fails or times out, the coordinator synthesizes the verdict directly and annotates `Chairman: <name> (FAILED — synthesized by coordinator fallback)`.
+
+### Step 6: Fallback Behavior
+
+If `spawn_agent` is unavailable or too many seats fail, run a local simulated council:
+
+- Read each selected persona file.
+- Produce clearly labeled `[Simulated]` outputs per member.
+- Keep the same round structure.
+- Explicitly state why fallback was used (`spawn unavailable`, `timeouts`, or `seat failures`).
+
+### Step 7: Session Metadata (issue #7, Phase 1)
+
+After the verdict is emitted, append a `Session Metadata` block with `schema_version: 1` containing: `mode`, `panel_size`, `rounds_run`, `tools_used`, `provider_count`, `fallbacks_triggered`, and best-effort `input_tokens_estimate` / `output_tokens_estimate` / `duration_seconds` (write `~unknown` if not available from the host runtime). Block is delimited by `---` so it's grep-able and redirectable.