the-high-council 0.6.0 → 0.7.6

package/.env.example

@@ -3,7 +3,6 @@
 ANTHROPIC_API_KEY=
 OPENAI_API_KEY=
 GOOGLE_API_KEY=
-XAI_API_KEY=
 MISTRAL_API_KEY=
 DEEPSEEK_API_KEY=
 GROQ_API_KEY=
@@ -14,6 +13,6 @@ ZAI_API_KEY=
 
 # Default per-run spend ceiling in USD. Overridden per run by --max-usd.
 # A run stops cleanly before any stage that could take it past this, and
-# resumes with a higher ceiling. Unset means $5. Use --max-usd none to run
+# resumes with a higher ceiling. Unset means $7. Use --max-usd none to run
 # without a ceiling.
-MAX_USD_PER_RUN=5
+MAX_USD_PER_RUN=7

package/README.md

@@ -1,24 +1,29 @@
 # The High Council
 
-[![Test](https://github.com/muad-yasin/the-high-council-mcp/actions/workflows/test.yml/badge.svg)](https://github.com/muad-yasin/the-high-council-mcp/actions/workflows/test.yml)
+> **Repository status (2026-09-22): going private until the open-source release.** Development
+> continues in a private repository, and the public open-source release is planned for later this
+> year. Until then, the published npm package `the-high-council` (0.7.6) stays installable as it
+> is, and the `npx github:...` lines below only work for someone with access to this repository.
+> The licence of everything already released stays MIT.
 
 See it work right now - no keys, no setup, no cost:
 
-    npx github:muad-yasin/the-high-council-mcp council demo
+    npx the-high-council council demo
 
 Add it to Claude Code as an MCP server the same way:
 
-    claude mcp add council -- npx -y github:muad-yasin/the-high-council-mcp council --mcp
+    claude mcp add council -- npx -y the-high-council council --mcp
 
-*(An npm package is coming - not yet published. Both lines above work today, straight from this
-GitHub repo, no npm account needed.)*
+*(With access to this repository, `npx github:muad-yasin/the-high-council-mcp council demo` runs
+the current source instead of the published package.)*
 
 A planning harness that runs one idea past several AI models from different labs, makes them
 argue about it on the record, and stops at a checkable result.
 
 Runs as an **MCP server** (so an agent like Claude Code can drive it) or as a **CLI**. Bring your
-own API keys. Nothing is resold, nothing is hosted for you, and no key ever leaves your machine.
-Run `npx github:muad-yasin/the-high-council-mcp council doctor` to see which of your own keys are
+own API keys. Nothing is resold, nothing is hosted for you, and your keys go only to the
+providers you choose, never to us.
+Run `npx the-high-council council doctor` to see which of your own keys are
 set, which shipped chains you can already run with them, and what each would cost - before
 spending anything.
 
@@ -76,6 +81,15 @@ anything on top of a run, read that instead. Do not parse the markdown.
     "replies": [ { "id": "DEEPSEEK-1", "action": "keep|amend|withdraw",
                    "text": "...", "replaced_by": "GLM-1" } ]
   },
+  // Only on chains with `alternatives: { enabled: true }`: one whole architecture per lab, written
+  // blind before the skeleton, and the same post/reply board the proposals get.
+  "alternatives": {
+    "items":   [ { "id": "GLM-ALT", "lab": "glm", "model": "...", "name": "...", "shape": "...",
+                   "key_tradeoffs": "...", "bad_at": "...", "amended": true, "withdrawn": false } ],
+    "posts":   [ { "by": "qwen", "on": "GLM-ALT", "stance": "object|support|merge", "text": "..." } ],
+    "replies": [ { "id": "GLM-ALT", "action": "keep|amend|withdraw", "text": "..." } ],
+    "dropouts": [ { "lab": "...", "model": "...", "stage": "alternatives", "reason": "..." } ]
+  },
   "signoff": [ { "provider": "qwen", "model": "...", "signedOff": false,
                  "objections": [ { "criterion": "...", "problem": "...", "fix": "..." } ] } ],
   "lastCritique": { "meets": false, "failures": [ { "criterion": "...", "lab": "qwen" } ] },
@@ -136,10 +150,63 @@ Fill in `.env` with only the keys your chosen chain needs. Supported providers:
 
 `ANTHROPIC_API_KEY` · `OPENAI_API_KEY` · `GOOGLE_API_KEY` · `MISTRAL_API_KEY` ·
 `DEEPSEEK_API_KEY` · `GROQ_API_KEY` · `TOGETHER_API_KEY` · `COHERE_API_KEY` ·
-`OPENROUTER_API_KEY` · `ZAI_API_KEY` · `XAI_API_KEY`
+`OPENROUTER_API_KEY` · `ZAI_API_KEY`
+
+xAI/Grok and Kimi/Moonshot are never seated, and neither is a router id (`openrouter/auto`, the
+Pareto Code router) that could route to them: chain-lint refuses such a chain and the run itself
+refuses such a seat, with no override. There is no `xai` provider.
 
 The CLI refuses to start if any seat in the chosen chain is missing its key, rather than failing
-halfway through a paid run.
+halfway through a paid run. `ollama` (below) is the one exception - it needs no key at all.
+
+### Standalone binaries (build them yourself)
+
+There are no published binary downloads yet. From a clone you can build a single-file `council`
+that needs no Node install to run: a Linux binary, a Linux AppImage, and a Windows `.exe`, all
+built on Linux.
+
+```bash
+npm run build:bin                 # dist/the-high-council-linux, dist/the-high-council-win.exe
+npm run build:appimage            # dist/the-high-council-x86_64.AppImage (downloads a pinned appimagetool)
+npm run smoke:bin -- dist/the-high-council-linux   # demo + an MCP session, offline, $0
+```
+
+Stated plainly: the `.exe` is tested under Wine on Linux, not on a real Windows machine, and none
+of the binaries are code-signed, so Windows SmartScreen will warn. Put `.env` in the directory you
+run the binary from.
+
+### Local models (Ollama, LM Studio, ...)
+
+A seat can run on a model you're hosting yourself instead of a paid API. Set its provider to
+`ollama`:
+
+```json
+{ "provider": "ollama", "model": "llama3.1" }
+```
+
+This talks to Ollama's own OpenAI-compatibility endpoint (`http://localhost:11434/v1` by
+default) - no `OLLAMA_API_KEY` needed; Ollama's docs describe a key as "required, but unused."
+If your local runner listens somewhere else (LM Studio, a different port, a remote box), add
+`baseUrl` to that seat:
+
+```json
+{ "provider": "ollama", "model": "mistral", "baseUrl": "http://localhost:1234/v1" }
+```
+
+A local seat prices at $0 in every estimate and run report - `dry_run` shows it as priced, not
+"unpriced," and it can never trip the spend cap. See `chains/local-ollama.json` for a full
+example chain with every seat local. As with every other seat in this project: BYOK is really
+BYO-compute here, and no efficacy claim is made about local models versus hosted ones - nothing
+has been measured.
+
+### Final security-review gate
+
+`"security_review": { "enabled": true }` adds one read-only reviewer as the last stage of a chain,
+after the build and everything else. It returns typed findings, and any critical or high finding
+fails the run (exit 7). A reviewer with no usable verdict fails it too (exit 8); that never counts
+as a pass. The default reviewer is `anthropic/claude-fable-5-1` (your key, priced, capped);
+`seats.security_reviewer` can be any seat, including a local `ollama` one as an offline option.
+Full details: [docs/security-review-gate.md](docs/security-review-gate.md).
 
 ## Quick start (CLI)
 
@@ -163,6 +230,34 @@ them stay on your machine - if you fork this repo, you will not accidentally pub
 
 `verify` is the cheap default: two labs, a hard two-round cap. Start there.
 
+### Exit codes
+
+Each outcome has its own code, so a script or CI job can branch on it:
+
+| Code | Meaning |
+|---|---|
+| 0 | Finished: a deliverable and `report.json` were written. |
+| 1 | The chain failed lint, or an input is missing (no task file, a resume from the wrong directory). |
+| 2 | Usage error: a flag is missing its value or has an invalid one. |
+| 3 | Paused at an external seat: answer `NEEDS-<stage>.md`, then `--resume`. |
+| 4 | Stopped at the per-run spend cap before the next stage was paid for; resume with a higher `--max-usd`. |
+| 5 | A provider key the chain needs is missing (degradable: set it and run again). |
+| 6 | Fatal: the chain file cannot be parsed, or the audit key cannot be loaded. |
+| 7 | Finished, but the final security review blocked the deliverable. |
+| 8 | Finished, but the final security review could not judge it (never a pass). |
+| 9 | Artifact gate: the task names files whose content it never includes (`BLOCKED-ARTIFACTS.md`). |
+| 10 | Preflight: a seat objected to the task description itself (`STOPPED-preflight.md`). |
+| 11 | PII gate (`--pii-gate hard-stop`): the input matches a PII shape or one of the key formats listed in `src/secret-patterns.js`. |
+| 12 | `policy.json` refused the chain. |
+| 13 | Another process is already running this run folder. |
+| 14 | The task file changed since the run started and `AMENDMENTS.md` does not cover it. |
+| 15 | `--resume` of a run that already finished (it has a `report.json`); nothing is run or spent. |
+| 16 | The run stopped at an unexpected error (a seat that could not be reached, a crash); `STOPPED-error.md` says what, and `--resume` continues from the completed stages. |
+| 17 | A draft (build, revise, dispute, final edit or handoff) was cut off at its token cap, and its one larger-cap retry was cut off too; the fragment is never graded or shipped (`STOPPED-truncated.md`). |
+
+Codes 9-17 were split out or added on 2026-09-23. Before that, several of these outcomes shared a code (2, 5,
+6 or 1).
+
 ### Telling the council what your build session can actually use
 
 One optional convention, worth the four lines it costs. If your task file ends with a section
@@ -187,7 +282,7 @@ before building it."*
 
 ### The spend cap
 
-Every run has a per-run ceiling in USD. It defaults to **$5**, and is checked *before* each paid
+Every run has a per-run ceiling in USD. It defaults to **$7**, and is checked *before* each paid
 stage - a stage that could take the run past the ceiling is never called, so the cap holds rather
 than reporting the overspend afterwards.
 
@@ -206,6 +301,14 @@ with a higher ceiling - completed stages replay from disk and cost nothing the s
 node src/cli.js --resume runs/<id> --max-usd 10
 ```
 
+A stage replays only if it would be asked the same thing again: the task text, the chain config
+and the stage's own prompt must all match what it answered. If one changed (you edited the chain,
+or an earlier stage came out differently this time), the stage runs again and its old files move
+into `superseded/` inside the run folder. What the old answer cost still counts toward the ceiling,
+`--spend` and `report.json`'s `totals.usd` (shown separately as `totals.supersededUsd`), so a resume
+cannot spend past the cap. An answer you wrote for an external stage is never re-used against a
+changed prompt: it is set aside the same way and the run asks you again.
+
 The cap governs one run. To see what you have spent across all of them:
 
 ```bash
@@ -238,12 +341,16 @@ node src/cli.js --resume runs/<id>
 
 ## Quick start (MCP)
 
-Register the server with your MCP client. No clone needed - straight from GitHub:
+Register the server with your MCP client. No clone needed - from the published npm package:
 
 ```bash
-claude mcp add council -- npx -y github:muad-yasin/the-high-council-mcp council --mcp
+claude mcp add council -- npx -y the-high-council council --mcp
 ```
 
+With access to this repository (private until the open-source release), the same line with
+`npx -y github:muad-yasin/the-high-council-mcp` runs the current source instead. The clone and
+plugin-marketplace routes below need that access too.
+
 Or from a clone of this repo:
 
 ```bash
@@ -253,7 +360,25 @@ claude mcp add high-council -- node /absolute/path/to/the-high-council-mcp/src/m
 **The path must be absolute.** The client starts the server from its own working directory, not
 from this one, so a relative path resolves somewhere unexpected and the server never starts.
 
-An npm package is coming - not yet published, no date. Once it lands:
+Or as a Claude Code plugin, which registers the same server and also installs the skills in
+[`skills/`](skills/README.md):
+
+```bash
+claude plugin marketplace add muad-yasin/the-high-council-mcp
+claude plugin install the-high-council@the-high-council
+```
+
+Claude Code copies the plugin into its plugin cache and runs `npm ci` there itself, so the first
+install takes a little longer. Start a new session (or run `/reload-plugins`) and the server shows
+up in `claude mcp list` as `plugin:the-high-council:high-council`. Remove it with
+`claude plugin uninstall the-high-council@the-high-council`. If you add the marketplace from a
+local clone instead (`claude plugin marketplace add /path/to/clone`), the server runs from the
+clone itself, so run `npm install` in it first.
+
+Registering with a client other than Claude Code (DeepSeek Harness, OpenHands, Cline, goose,
+Continue): [docs/mcp-clients.md](docs/mcp-clients.md) has the real config for each.
+
+An npm package is also live:
 
 ```bash
 claude mcp add high-council -- npx -y the-high-council --mcp
@@ -280,6 +405,7 @@ Then drive it with these tools:
 | `run_status` | Stage reached, panel verdicts, scoreboard, cost, budget remaining |
 | `spend_report` | What every run has cost across a window of days, plus `session_cost_today` (calendar-day, per-model breakdown), derived from disk |
 | `verdict_stats` | How the debate mechanism itself is doing, per chain and per lab: sign-off rate, rounds, objections, dropouts, cost/wall time, largest prompt files - derived from disk |
+| `metrics_report` | Descriptive telemetry only (never an evaluation or baseline): amendment rate, withdrawal rate, objection-follow-through rate, tool-call usage, consensus-induced-regression count - derived from disk. See [`docs/paid-comparison-run.md`](docs/paid-comparison-run.md) for what a *real* comparison against a single model would need - not built, not this metric |
 | `external_prompt` | The prompt a paused external stage is waiting on - **read it with this, not by opening the file**; the prompts are routinely tens of thousands of tokens and most clients silently truncate a file read |
 | `prepare_stage_prompt` | Write a self-contained bundle for a paused stage, so a driving session doing other work at the same time can dispatch it to a fresh subagent instead of authoring it inline - see [`docs/dispatch-pattern.md`](docs/dispatch-pattern.md) |
 | `submit_stage` | Answer a paused stage and resume. Optional `claimed_by` records who's answering; a second submission for an already-answered stage no longer errors - it's kept as `<stage>.late.md` with a warning instead of being rejected |
@@ -299,7 +425,7 @@ for anyone with a subscription, since the project's cost story depends on it.
 
 ## Chains
 
-31 chain configs live in `chains/`. Each one is plain JSON - the seat roster, which models fill
+48 chain configs live in `chains/`. Each one is plain JSON - the seat roster, which models fill
 which seat, the round cap, and whether proposals/debate/handoff stages run. They are meant to be
 copied and edited.
 
@@ -309,7 +435,8 @@ A few worth knowing:
 - **`cheap`** - small models throughout. For testing the harness itself, not for real work.
 - **`plan-debate`** - five labs propose blind, debate each other's proposals anonymised, then a
   blind panel grades the integrated draft.
-- **`plan-auto`** - the full open-scope chain, every seat on a real API, runs unattended end to end.
+- **`plan-auto`** - the full open-scope chain with every seat on a real API, so no stage waits for
+  an external session. It still pauses once for your answers to its questions.
 - **`plan-unanimous`** - every critic must independently sign off on the *same* draft. Three rounds
   here means up to fifteen critic calls, not five.
 
@@ -361,8 +488,9 @@ Read a chain's `description` field before running it; they say what they cost yo
   review round can't run), or a seat naming a provider this codebase doesn't know. The same check
   also runs automatically, fail-loud, the moment you try to actually run a broken chain - before
   any paid call.
-- **`council doctor --scan-artifacts`** - scans your `tasks/` and `chains/` for API-key-shaped
-  strings before you commit or share them. Reports the file and line, never the matched text.
+- **`council doctor --scan-artifacts`** - scans your `tasks/` and `chains/` for the key formats
+  listed in `src/secret-patterns.js` before you commit or share them. A format not on that list is
+  not caught. Reports the file and line, never the matched text.
 - **`council --forecast-cost --chain <name> [--days N]`** - a realistic-case USD range for a
   chain, built from its own historical runs on your machine and repriced at today's rates -
   distinct from `--dry-run`'s worst-case estimate from a chain's declared token assumptions.
@@ -370,7 +498,7 @@ Read a chain's `description` field before running it; they say what they cost yo
   verdict as one self-contained HTML file. No external assets, no server, `<details>` sections.
 - **`council replay --run <folder> [--json]`** - a numbered, step-by-step transcript of a run's
   reasoning in your terminal, with a `--json` mode for scripting.
-- **A structured error catalog** (`COUNCIL-E001`-`COUNCIL-E004`, see [TROUBLESHOOTING.md](TROUBLESHOOTING.md))
+- **A structured error catalog** (`COUNCIL-E001`-`COUNCIL-E005`, see [TROUBLESHOOTING.md](TROUBLESHOOTING.md))
   for the hard-fail paths this project actually has - a missing key, an unpriced model, a
   malformed chain file, an unreadable stage reply - each with a plain-words cause, a concrete fix
   naming the real file, and a doc pointer. Degradable conditions exit 5; fatal ones exit 6.
@@ -417,6 +545,24 @@ claiming 29 chains and 12 tools when there were 30 and 11.
 
 No claim is made here that any of this produces better plans. It has not been measured.
 
+## Skills
+
+`skills/` carries a small set of Claude Code skills shipped alongside this harness, not part of
+it - usable in any project, with or without the harness, and needing no API key:
+
+- **Core agentic work:** `task-scoping`, `research-and-sourcing`, `verification-and-critique`,
+  `context-and-handoff`, `tool-and-action-discipline` - the work around the work, for any agent
+  task, including setups where several models propose, critique, and hand off to each other.
+- **Building software:** `backend-developer`, `frontend-developer`, `ux-design`, `visual-craft`.
+- **Writing:** `good-news-writing`.
+
+The full catalog, with when to use each and what it guards against, is
+[`skills/README.md`](skills/README.md); a shareable presentation page is `docs/skills.html`. Both
+are checked by the test suite against the real `skills/` folder. Same MIT/BYOK ethos as the
+harness: the skills don't claim to make results better - nothing has been measured - they describe
+what each rule guards against and why. Copy the folders you want into your project's
+`.claude/skills/`, or point Claude Code at this repo's `skills/` directory.
+
 ## Known limits, stated plainly
 
 - The spend cap is enforced against a **worst case**, not a prediction: the whole prompt billed as
@@ -425,7 +571,20 @@ No claim is made here that any of this produces better plans. It has not been me
   left. That is the intended trade - resume it with a higher ceiling.
 - A seat whose model has no entry in `src/pricing.json` is **unpriced, and therefore uncapped**.
   It contributes $0 to the running total no matter what it really costs. Check `dry_run` output
-  for `unpriced` before trusting a ceiling.
+  for `unpriced` before trusting a ceiling. `ollama` seats are the one deliberate exception: they
+  price at an explicit $0 for any model name (they are genuinely free to run), so they never show
+  up in that `unpriced` list and never need a pricing.json entry.
+- A call that fails **after** it reached the provider - the connection drops while a 200 reply is
+  being read, a 200 carries a body that is not JSON, or the request times out waiting for headers -
+  is not retried, because the generation was probably already billed. Its real usage cannot be
+  read, so it is counted toward the per-run cap at its worst-case projection. It leaves no
+  `<label>.usage.json`, so `council --spend` does not see it.
+- Some hosted reasoning models stop at a provider-side reasoning ceiling well below the seat's own
+  cap, before writing any answer (seen 2026-09-23: `deepseek/deepseek-v4.1-flash` via OpenRouter, 6 of
+  28 calls stopped at ~4.2k tokens, all reasoning, under a 36k cap). A bigger cap cannot fix that, so
+  such a reply is recorded as `REASONING_EXHAUSTED` and is not retried: the seat abstains, which
+  blocks unanimity and never counts as consent. Whether to change a chain's reasoning settings for it
+  is a roster decision, not something the harness does for you.
 - Prices in `src/pricing.json` are hand-maintained list prices, last verified 2026-09-06. They are
   estimates, not invoices. Your provider's bill is the real number.
 - Chains with many labs and high round caps get expensive quickly. `plan-unanimous` at three rounds
@@ -436,6 +595,10 @@ No claim is made here that any of this produces better plans. It has not been me
   produce sign-offs that mean nothing - the quality of the run depends heavily on the quality of
   the criteria stage.
 - Some chains pause for human input by design. They are not stuck.
+- The audit log (`audit.jsonl`) detects an edited line or a cut-off end, but **with no HMAC key,
+  which is the default, it does not detect a full rewrite**: anyone who can edit the file can
+  recompute every hash. Set `AUDIT_HMAC_KEY_FILE` to a key kept outside the run folder if the log
+  has to stand up to that.
 
 ## Privacy
 
@@ -447,16 +610,21 @@ collect nothing. Stated plainly, not as a claim about quality:
   your own keys (`ANTHROPIC_API_KEY`, `OPENROUTER_API_KEY`, and so on) - and only when you start
   a real run with a non-mock, non-external seat. `council doctor` and `council demo` make zero
   network calls; both are implemented to read only local files and environment variable names.
-- The GitHub Pages site (`docs/`) ships no JavaScript and loads no third-party resource - see
+- The landing page (`docs/`, moving from GitHub Pages to sower-industries.de while this
+  repository is private) ships no JavaScript and loads no third-party resource - see
   `test/landing-page.test.js`, which fails the build if either ever changes.
 - Everything a run produces - the task, the debate, the deliverable, the cost - is written to a
   folder on your own disk (`runs/<id>/`) and nowhere else. Deleting that folder deletes the
   record. See `src/spend.js`'s own documentation for why spend accounting is derived from those
   folders rather than kept in a separate log.
+- If you opt in to the audit export (`"audit": true`, and on automatically whenever a `policy.json`
+  exists), each run also writes `audit.jsonl` into its own folder: signed, hash-chained lines (each carries the previous line's hash) that
+  `verifyAuditLog` in `src/audit.js` checks offline. Seat, cost and timing only, never task or
+  prompt text. Format: `docs/audit-schema.md`.
 
 ## Accessibility
 
-The GitHub Pages site (`docs/`) uses only native HTML elements for interaction - `<a href>`
+The landing page (`docs/`) uses only native HTML elements for interaction - `<a href>`
 links and `<details>/<summary>` for the expandable demo sections - specifically because those
 are keyboard-operable by the browser itself, with no custom JavaScript to get wrong. Tested by
 hand, 2026-09-13, on the live site:
@@ -478,8 +646,8 @@ hand, 2026-09-13, on the live site:
 
 ## Reporting a bug
 
-Open an issue: [github.com/muad-yasin/the-high-council-mcp/issues](https://github.com/muad-yasin/the-high-council-mcp/issues).
-Or write to [contact@sower-industries.de](mailto:contact@sower-industries.de).
+Write to [contact@sower-industries.de](mailto:contact@sower-industries.de). (GitHub issues open
+again with the public release; while the repository is private, email is the only route.)
 
 ## If this makes you money
 

package/TROUBLESHOOTING.md

@@ -43,3 +43,16 @@ lab actually returned. The run log also says why it was unreadable: truncated
 (hit its token cap), malformed JSON, or the provider itself returned an error
 mid-generation - each needs a different fix (a higher `maxTokens`, a smaller
 prompt, or just retrying).
+
+## COUNCIL-E005 - Policy refusal (fatal)
+
+Either `policy.json` itself isn't valid JSON, or the chain you're about to run
+violates one of its limits (an unlisted provider or region, over the per-run or
+per-month spend cap, a missing required tag, or an unpriced seat when
+`refuse_unpriced_seats` is set). A policy file is a locked, operator-set set of
+limits checked before any provider is called - a run that would violate it never
+starts, so nothing was spent.
+
+**Fix:** if the file itself won't parse, fix its JSON syntax. Otherwise, either
+change the chain so it no longer violates the listed limit(s), or change
+`policy.json` if the limit itself was wrong. See [docs/policy.md](docs/policy.md).

package/chains/cheap-7-v2.json

@@ -0,0 +1,98 @@
+{
+  "name": "cheap-7-v2",
+  "description": "Seven independent labs vote, none of them the author's, all on current-generation models through one OpenRouter key: OpenAI GPT-5.6 Luna, Google Gemini 3.8 Flash, DeepSeek v4.1 Flash, Z.ai GLM 5.3 Flash, Qwen 3.7 Plus, MiniMax M3, Mistral Medium 3.5. Blind proposals, debate, unanimous signoff, with a round cap of 7 and a stall rule: if the same objections from the same labs survive two rounds running, the revise loop stops rather than buying more rounds of the same argument, and the unresolved dissent is recorded at the top of the deliverable. A run that ends that way reports no consensus - the dispute stage records disagreement, it never resolves it. Anthropic Sonnet 5 writes the skeleton, the draft, every revision and the handoff and holds NO panel seat; Meta Llama 4 Maverick writes the criteria and also holds no panel seat, so neither the lab that drafts nor the lab that sets the bar votes on whether the bar was cleared. This is the revision of cheap-7 after its first real run, where one lab held the criteria, skeleton, builder, reviser and handoff seats AND a critic seat, voted on its own draft, and a claim it wrote survived to the deliverable. Anthropic is deliberately absent from the panel: plans from this chain are read by a Claude session before anything is built, so that review happens downstream rather than inside the vote. Kimi K3 is deliberately absent too - a seat that may route through another lab is not an independent seat. Every seat gets a 36000-token ceiling and reasoning is left on: a debate is not squeezed to a short answer, and a reply cut off at the ceiling is retried once, then the seat is re-asked alone - it is never counted as agreement. The cap is a ceiling, not a price; only generated tokens are billed. Prices are from src/pricing.json, refreshed against OpenRouter's live model list on 2026-09-20 - price a real task with `npm run dry` before running.",
+  "maxRounds": 7,
+  "signoff": "unanimous",
+  "dispute": {
+    "enabled": true,
+    "stall_rounds": 2
+  },
+  "proposals": {
+    "parts": 3,
+    "maxTokens": 4000
+  },
+  "debate": true,
+  "canary": { "enabled": true, "sampleRate": 0.1 },
+  "handoff": true,
+  "estimate": {
+    "promptTokens": 10000,
+    "draftTokens": 10000,
+    "critiqueTokens": 5000
+  },
+  "seats": {
+    "criteria": {
+      "provider": "openrouter",
+      "model": "meta-llama/llama-4-maverick",
+      "maxTokens": 36000,
+      "lab": "llama4-maverick"
+    },
+    "skeleton": {
+      "provider": "openrouter",
+      "model": "anthropic/claude-sonnet-5",
+      "maxTokens": 36000,
+      "lab": "sonnet5"
+    },
+    "builder": {
+      "provider": "openrouter",
+      "model": "anthropic/claude-sonnet-5",
+      "maxTokens": 36000,
+      "lab": "sonnet5"
+    },
+    "reviser": {
+      "provider": "openrouter",
+      "model": "anthropic/claude-sonnet-5",
+      "maxTokens": 36000,
+      "lab": "sonnet5"
+    },
+    "handoff": {
+      "provider": "openrouter",
+      "model": "anthropic/claude-sonnet-5",
+      "maxTokens": 36000,
+      "lab": "sonnet5"
+    },
+    "critics": [
+      {
+        "provider": "openrouter",
+        "model": "openai/gpt-5.6-luna",
+        "maxTokens": 36000,
+        "lab": "gpt5.6-luna"
+      },
+      {
+        "provider": "openrouter",
+        "model": "google/gemini-3.8-flash",
+        "maxTokens": 36000,
+        "lab": "gemini3.8-flash"
+      },
+      {
+        "provider": "openrouter",
+        "model": "deepseek/deepseek-v4.1-flash",
+        "maxTokens": 36000,
+        "lab": "deepseek-v4.1-flash"
+      },
+      {
+        "provider": "openrouter",
+        "model": "z-ai/glm-5.3-flash",
+        "maxTokens": 36000,
+        "lab": "glm5.3-flash"
+      },
+      {
+        "provider": "openrouter",
+        "model": "qwen/qwen3.7-plus",
+        "maxTokens": 36000,
+        "lab": "qwen3.7-plus"
+      },
+      {
+        "provider": "openrouter",
+        "model": "minimax/minimax-m3",
+        "maxTokens": 36000,
+        "lab": "minimax-m3"
+      },
+      {
+        "provider": "openrouter",
+        "model": "mistralai/mistral-medium-3-5",
+        "maxTokens": 36000,
+        "lab": "mistral-medium3.5"
+      }
+    ]
+  }
+}

package/chains/cheap-7.json

@@ -0,0 +1,93 @@
+{
+  "name": "cheap-7",
+  "description": "SUPERSEDED by cheap-7-v2, and deliberately left failing chain-lint's self-review check, so it will not run - kept unmodified as the record of what the 2026-09-20 run actually used. Sonnet 5 holds the criteria, skeleton, builder, reviser and handoff seats AND a critic seat here, so it voted on its own draft under unanimous signoff; that is the defect cheap-7-v2 fixes. Do not add \"selfReview\": \"allowed\" to make this runnable again - use cheap-7-v2. Original description follows. Seven labs' small/cheap models, one seat each, all through one OpenRouter key: Anthropic Sonnet 5, OpenAI GPT-5 mini, Google Gemini 3.7 Flash, DeepSeek v3.2, Z.ai GLM 5.3 Flash, Qwen 3.5 Plus, MiniMax M3. Blind proposals, debate, unanimous signoff. Every seat gets a 36000-token ceiling and reasoning is left on: a debate is not squeezed to a short answer, and a reply cut off at the ceiling is retried once, then the seat is re-asked alone - it is never counted as agreement. The cap is a ceiling, not a price; only generated tokens are billed. Sonnet 5 builds and revises. For small questions where every model should still get a say; the frontier equivalent is reserved for questions worth its price. Prices are from src/pricing.json - price a real task with `npm run dry` before running.",
+  "maxRounds": 3,
+  "signoff": "unanimous",
+  "proposals": {
+    "parts": 3,
+    "maxTokens": 4000
+  },
+  "debate": true,
+  "handoff": true,
+  "estimate": {
+    "promptTokens": 10000,
+    "draftTokens": 10000,
+    "critiqueTokens": 5000
+  },
+  "seats": {
+    "criteria": {
+      "provider": "openrouter",
+      "model": "anthropic/claude-sonnet-5",
+      "maxTokens": 36000,
+      "lab": "sonnet5"
+    },
+    "skeleton": {
+      "provider": "openrouter",
+      "model": "anthropic/claude-sonnet-5",
+      "maxTokens": 36000,
+      "lab": "sonnet5"
+    },
+    "builder": {
+      "provider": "openrouter",
+      "model": "anthropic/claude-sonnet-5",
+      "maxTokens": 36000,
+      "lab": "sonnet5"
+    },
+    "reviser": {
+      "provider": "openrouter",
+      "model": "anthropic/claude-sonnet-5",
+      "maxTokens": 36000,
+      "lab": "sonnet5"
+    },
+    "handoff": {
+      "provider": "openrouter",
+      "model": "anthropic/claude-sonnet-5",
+      "maxTokens": 36000,
+      "lab": "sonnet5"
+    },
+    "critics": [
+      {
+        "provider": "openrouter",
+        "model": "anthropic/claude-sonnet-5",
+        "maxTokens": 36000,
+        "lab": "sonnet5"
+      },
+      {
+        "provider": "openrouter",
+        "model": "openai/gpt-5-mini",
+        "maxTokens": 36000,
+        "lab": "gpt5-mini"
+      },
+      {
+        "provider": "openrouter",
+        "model": "google/gemini-3.7-flash",
+        "maxTokens": 36000,
+        "lab": "gemini3.7-flash"
+      },
+      {
+        "provider": "openrouter",
+        "model": "deepseek/deepseek-v3.2",
+        "maxTokens": 36000,
+        "lab": "deepseek-v3.2"
+      },
+      {
+        "provider": "openrouter",
+        "model": "z-ai/glm-5.3-flash",
+        "maxTokens": 36000,
+        "lab": "glm5.3-flash"
+      },
+      {
+        "provider": "openrouter",
+        "model": "qwen/qwen3.5-plus-20260420",
+        "maxTokens": 36000,
+        "lab": "qwen3.5-plus"
+      },
+      {
+        "provider": "openrouter",
+        "model": "minimax/minimax-m3",
+        "maxTokens": 36000,
+        "lab": "minimax-m3"
+      }
+    ]
+  }
+}

package/chains/coder-gate-v1.json

@@ -0,0 +1,56 @@
+{
+  "name": "coder-gate-v1",
+  "description": "MLLM Coder v1 (relay/runs/2026-09-14T16-14-10-757Z/deliverable.md). The governed unit is one diff against one pre-named file/hunk, not prose. The `propose` seat is external: THCMCP never writes code itself, it hands the change-request task file to an outside coding session and gets one fenced diff back (docs/coder-gate-diff-format.md). The `builder` seat is also external, held to the same session, and must return the proposal's diff verbatim as this chain's draft - there is no second, independent code-writing step; `builder` exists only because every chain needs one to reach a draft, not because a second author writes anything. `verify` (config.verify.enabled: true) wires the existing run_tests mechanism, but `verify.tools` ships empty here by design, not oversight: runVerification runs before any change-request/diff exists, and this repo has no `{{...}}` templating to inject a per-run target file/test command into a static chain config - a filled-in entry as specced would be inert. No test command runs automatically today (see test/coder-gate-verify.test.js's IN-3 fixture for how the mechanism works when a real, per-run `verify.tools` entry IS supplied downstream). No write-capable tool is used anywhere - applying the diff to a real working tree stays a human/external action. Built from existing stage types (proposal, critic, unanimous signoff); `debate: true` is set for forward compatibility but is currently unreachable - with one proposer and `proposals.parts: 1`, at most one proposal ever exists, and `runChain`'s debate stage only runs when `proposals.length > 1`, so v1 has no live debate today. No change to src/chain.js or src/tools.js.",
+  "maxRounds": 2,
+  "signoff": "unanimous",
+  "proposals": {
+    "parts": 1,
+    "maxTokens": 4000
+  },
+  "estimate": {
+    "promptTokens": 3000,
+    "draftTokens": 3000,
+    "critiqueTokens": 1000
+  },
+  "seats": {
+    "criteria": {
+      "provider": "anthropic",
+      "model": "claude-sonnet-5",
+      "maxTokens": 2000
+    },
+    "skeleton": {
+      "provider": "anthropic",
+      "model": "claude-sonnet-5",
+      "maxTokens": 1000
+    },
+    "proposers": [
+      {
+        "provider": "external",
+        "model": "claude-code-session",
+        "lab": "external-proposer"
+      }
+    ],
+    "builder": {
+      "provider": "external",
+      "model": "claude-code-session"
+    },
+    "critics": [
+      {
+        "provider": "anthropic",
+        "model": "claude-sonnet-5",
+        "maxTokens": 8000
+      },
+      {
+        "provider": "openai",
+        "model": "gpt-5",
+        "maxTokens": 8000
+      }
+    ]
+  },
+  "debate": true,
+  "handoff": false,
+  "verify": {
+    "enabled": true,
+    "tools": []
+  }
+}