amicus 2.0.0 → 2.2.0

package/.claude-plugin/plugin.json

@@ -1,6 +1,6 @@
 {
   "name": "amicus",
-  "version": "2.0.0",
+  "version": "2.2.0",
   "description": "Multi-model LLM Council + parallel AI window for Claude Code. Run structured council reviews across Gemini, GPT, DeepSeek and more — or fork a conversation to any model and fold the results back.",
   "author": {
     "name": "Christian Wagner"

package/CHANGELOG.md

@@ -5,6 +5,114 @@ All notable changes to Amicus are documented here. Format follows
 
 ## [Unreleased]
 
+## [2.2.0] - 2026-07-14
+
+### Added
+
+- **Optional council elements** (second-opinion skill): four new opt-in behaviors, presented
+  together with Claude-in-the-council as a single numbered Stage-0 menu — all default OFF, enabled
+  only when the user names them, and the launch confirmation must enumerate what's on:
+  - **Critic seat** — one bench member swaps to a four-pass adversarial brief (adversarial pass,
+    edge-case hunt, consistency check, executability test), launched as a concurrent solo beside
+    the fanout wave (`role: "critic"`). Its findings enter the same anonymized bundle and are
+    peer-adjudicated like any other seat's — the council disciplines the critic.
+  - **Expert lenses** — each seat reviews through a distinct expert perspective (panel domain
+    scoped with the user: business/technical/customer/financial/custom). Lens runs always tally
+    `--no-ledger` and the report discloses the weakened cross-review anonymity.
+  - **Debate mode** — a new Stage 2.5 rebuttal round: Contested/Disputed findings go back to their
+    raisers to DEFEND / AMEND / WITHDRAW, disputing judges re-vote, then the final ledger-recorded
+    tally. Exactly one round; withdrawn findings are auto-denied and listed in the report.
+  - **Chair verdict scale** — the chair closes with 3–5 hard questions and a final parseable
+    `VERDICT: Ship it | Fix these first | Fundamental rethink` line, surfaced at the top of the
+    report.
+- New `skills/second-opinion/SEAT-BRIEFS.md` — briefing boilerplate for the elements plus a
+  standard anti-sycophancy clause now required in **every** Stage-1 briefing. Critic and lens
+  methodologies adapted from the `/critic` and `/debate` agents in John Renaldi's product-kit
+  (MIT), with deliberate deviations documented (no findings quota; verdict moved to the chair).
+- Zero engine changes: free-form `runStats[].role` labels, tally-input re-assembly, and the
+  existing `--no-ledger` flag cover all four elements.
+
+### Changed
+
+- `COUNCIL-DESIGN.md` gains §12 documenting the elements, their caveats (critic
+  self-identification in cross-review; lens anonymity/ledger trade-offs), and parallel panels as
+  future work. The `/council` command accepts pre-requested elements in its arguments.
+- MODEL-NOTES seed: fold-back of the v2.2.0 verification council — claim-class dedup
+  adjudication limit, minimax debut (strong critic seat), qwen-coder debut.
+
+### Fixed
+
+- Resurrected both Windows e2e integration suites, silently dead since before the rebrand:
+  Node's `spawn()` cannot execute `node_modules/.bin` `.cmd` shims on Windows, so the Electron
+  toolbar suite (electron shim) and the OpenCode server test helper ENOENT'd without a visible
+  error. The helper now uses the shared `ensureNodeModulesBinInPath()` (which adds the platform
+  `opencode.exe` dirs to PATH) and the toolbar suite spawns the real binary via
+  `require('electron')`. Also refreshed two stale pins the dead suites never caught: the wave
+  document's `schemaVersion` (now pinned to the shared `SCHEMA_VERSION` constant instead of a
+  literal `1`) and the pre-rebrand `Sidecar` toolbar brand assertion (now `Amicus`).
+
+## [2.1.0] - 2026-07-04
+
+### Added
+
+- `--json` on `resume`, `continue`, and `abort`. `amicus resume <id> --no-ui --json` and
+  `amicus continue <id> --prompt "..." --no-ui --json` emit the same versioned run document as
+  `start --json` (a `continue` run's document carries the new continuation task id, not the old
+  one). `amicus abort <id|--all> --json` emits a new `type: 'abort'` document
+  (`{ schemaVersion, type, ok, scope, taskId, aborted, count }`) covering single-session, wave, and
+  `--all` aborts, success and failure alike — stdout carries exactly one parseable document either
+  way, and non-`--json` human output is unchanged (byte-identical pinned messages still hold).
+- Did-you-mean suggestions for unknown CLI commands: a near-miss typo like `amicus contnue` now
+  prints `Unknown command: contnue` followed by `Did you mean: continue` on stderr (still exits 1).
+  Suggestions are capped at 3 and only shown within edit-distance 2 of a known command; unrelated
+  garbage input gets no suggestion.
+
+### Changed
+
+- Agent-facing polling guidance now recommends `amicus_wait` first across every headless-flow
+  reminder, tool description, and guide section (MCP system-reminders, `amicus_start`/`amicus_resume`/
+  `amicus_continue`/`amicus_fanout` descriptions, `amicus_guide`'s headless workflow, and the
+  `second-opinion`/`sidecar` skill docs) — one blocking call replaces the sleep+status poll loop.
+  `sleep 25` + `amicus_status` polling remains documented as the explicit fallback for clients
+  without the `amicus_wait` tool; it is never presented as the only mechanism.
+
+### Fixed
+
+- `amicus doctor`'s MCP registration check no longer false-negatives on a healthy Claude Code
+  registration. The check's only signal was `discoverClaudeCodeMcps()`, which always strips every
+  `amicus`/`sidecar`-shaped entry as its own recursive-spawn guard — so the check could never see
+  its own registration and warned "not registered in Claude Code" even when one existed. The check
+  now reads the same config sources directly (unstripped) to answer "is amicus registered?".
+
+### CI / Security
+
+- `council-review.yml`: both fanout legs (review wave and synthesis) now request
+  `--summary-length normal` instead of `verbose`. `--summary-length` only shapes the prompt (there is
+  no engine-side output-token cap), so `verbose` was asking every model in the wave — on a paid CI
+  key — for maximally long output on every PR.
+- `council-review.yml`: the model-to-model handoff from the review wave into the synthesis leg is
+  now neutralized. The synthesis briefing previously concatenated raw model review text
+  (`reviews.md`) straight into another model's prompt with no sanitization; it now runs the same
+  neutralization (byte-identical sed rules, duplicated into the synthesis step's own shell) used on
+  the human-facing PR comment, and wraps the reviews in an explicit untrusted-data block before
+  handing them to the synthesis model. The comment path itself is unchanged.
+- `ci.yml`: the `quality` job now runs [actionlint](https://github.com/rhysd/actionlint) (pinned to
+  v1.7.7) over `.github/workflows/`, which also shellchecks every `run:` block via ubuntu-latest's
+  preinstalled shellcheck. Verified locally with the actionlint + shellcheck Windows binaries before
+  landing; both are clean against all 5 workflows (0 findings), so no suppression config was needed.
+
+### Documentation
+
+- Corrected the `--agent` default docs in `skills/sidecar/SKILL.md`: the flag defaults to `Chat`
+  only in interactive mode — headless (`--no-ui`) runs default to `Build`, since `chat` stalls
+  without user interaction. The file previously claimed an unqualified "defaults to Chat" in
+  several spots while also correctly documenting the headless-`Build` default elsewhere,
+  contradicting itself; `docs/usage.md` was already correct and unchanged.
+- Corrected `commands/council.md`'s description of the council pipeline order: `amicus council
+  validate` runs per-leg during Stage 1 (independent reviews), and `amicus council tally` runs
+  after Stage 2 (cross-review) and before Stage 3 (chair synthesis) — not, as previously worded,
+  both after all three review waves.
+
 ## [2.0.0] - 2026-07-03
 
 Amicus's first major release: the **`sidecar*` shim removal** (#19). v1.x carried a full

package/README.md

@@ -53,6 +53,10 @@ Claude is the orchestrator. The council and chat skills run *on top of* the engi
 
 ![What one install delivers: council skill, chat skill, CLI + MCP, live catalog](./docs/what-is-amicus.png)
 
+The council skill in one picture — independent review, before cross-review or the verdict, with Claude Opus seated among the models it's judging:
+
+![The Amicus council mid-ritual: five models — Gemini 3 Pro, Llama 4, Grok 4, Claude Opus — reading the same material independently, chaired by GPT-5](./docs/council.png)
+
 ---
 
 ## Quick start
@@ -209,11 +213,19 @@ Everything you need before your first run, and what's optional.
   findings-by-tier, cost — no chair prose). This is the default artifact handed to the user.
 - For an **editable source**, the accepted edits land in `<stem>-reviewed.<ext>` next to the original.
 
-**Claude in the council** (default off): you can add Claude's own fresh review to the bundle so the bench ranks and adjudicates it — Claude is *judged* but never votes or chairs, so the verdict stays independent.
+**Optional council elements** (v2.2.0, all default off): five opt-in behaviors, offered once as a menu at launch — nothing turns on unless you name it, and the confirmation lists exactly what's on.
+
+- **Critic seat** — one reviewer swaps to a four-pass adversarial brief (adversarial pass, edge-case hunt, consistency check, executability test). Its findings enter the same anonymized bundle as everyone else's, so the bench disciplines the critic: manufactured negativity lands Disputed and dies in the tally.
+- **Expert lenses** — each reviewer takes a distinct expert perspective; you pick the panel domain (business, technical, customer, financial, or custom). Lens runs never feed the reliability ledger, and the report discloses the weakened cross-review anonymity.
+- **Debate mode** — after cross-review, every Contested or Disputed finding goes back to its raiser to **defend, amend, or withdraw**, and the disputing judges re-vote. Exactly one rebuttal round, then the final tally.
+- **Chair verdict scale** — the chair closes with 3–5 hard questions and one parseable line: `VERDICT: Ship it | Fix these first | Fundamental rethink`.
+- **Claude in the council** — Claude adds its own fresh review to the bundle so the bench ranks and adjudicates it. Claude is *judged* but never votes or chairs, so the verdict stays independent.
+
+The critic and lens methodologies are adapted from the `/critic` and `/debate` agents in [John Renaldi's product-kit](https://github.com/jrenaldi79/plugin-marketplace) (MIT); the briefing boilerplate lives in [`skills/second-opinion/SEAT-BRIEFS.md`](./skills/second-opinion/SEAT-BRIEFS.md).
 
-**Cost is disclosed up front.** Before any model launches, you see the run shape — for example:
+**Cost is disclosed up front.** Before any model launches, you see the run shape — including any enabled optional elements — for example:
 
-> This run uses 3 council models across 2 fanout waves + 1 chair call (~7 model runs).
+> This run uses 3 council models across 2 fanout waves + 1 chair call, with critic seat + debate mode ON (~7 base runs + up to 6 rebuttal calls).
 
 Then the council waits for your confirmation.
 
@@ -311,7 +323,7 @@ $ amicus status demo123 --json
   "taskId": "demo123",
   "status": "complete",
   "elapsed": "5m 0s",
-  "version": "2.0.0",
+  "version": "2.2.0",
   "model": "google/gemini-2.5-flash",
   "phase": "terminal"
 }

package/bin/amicus.js

@@ -11,12 +11,12 @@
 const { loadCredentials } = require('../src/utils/env-loader');
 loadCredentials();
 
-const { parseArgs, getUsage } = require('../src/cli');
-const { validateTaskId } = require('../src/utils/validators');
-const { resolveModelFromArgs, validateFallbackModel } = require('../src/utils/start-helpers');
+const { parseArgs, getUsage, getCommandNames } = require('../src/cli');
 const { handleSetup, handleAbort, handleUpdate, handleMcp, handleKey } = require('../src/cli-handlers');
 const { handleStart, handleFanout, handleRead } = require('../src/cli-handlers-run');
+const { handleResume, handleContinue } = require('../src/cli-handlers-resume-continue');
 const { isOneShotCommand, armExitWatchdog } = require('../src/utils/lifecycle');
+const { suggestCommand } = require('../src/utils/input-validators');
 const { logger } = require('../src/utils/logger');
 
 const VERSION = require('../package.json').version;
@@ -124,7 +124,7 @@ async function main() {
         await handleKey(args);
         break;
       case 'abort':
-        await handleAbort(args);
+        exitCode = await handleAbort(args);
         break;
       case 'mcp':
         await handleMcp();
@@ -132,10 +132,16 @@ async function main() {
       case 'update':
         await handleUpdate();
         break;
-      default:
+      default: {
         console.error(`Unknown command: ${command}`);
+        // suggestCommand honors a cap-3 contract (up to 3 candidates, closest
+        // first) — print all of them, not just the closest, matching the
+        // join precedent in src/cli-handlers.js.
+        const candidates = suggestCommand(command, getCommandNames());
+        if (candidates.length > 0) { console.error(`Did you mean: ${candidates.join(', ')}`); }
         console.log(getUsage());
         process.exit(1);
+      }
     }
   } catch (err) {
     console.error(`Error: ${err.message}`);
@@ -166,94 +172,6 @@ async function handleList(args) {
   });
 }
 
-/**
- * Handle 'sidecar resume' command
- * Spec Reference: §4.3
- */
-async function handleResume(args) {
-  const taskId = args._[1];
-
-  if (!taskId) {
-    console.error('Error: task_id is required for resume');
-    console.error('Usage: amicus resume <task_id>');
-    process.exit(1);
-  }
-
-  const taskIdCheck = validateTaskId(taskId);
-  if (!taskIdCheck.valid) {
-    console.error(taskIdCheck.error);
-    process.exit(1);
-  }
-
-  const { resumeAmicus } = require('../src/index');
-
-  return await resumeAmicus({
-    taskId,
-    project: args.cwd,
-    headless: args['no-ui'],
-    timeout: args.timeout
-  });
-}
-
-/**
- * Handle 'sidecar continue' command
- * Spec Reference: §4.4
- */
-async function handleContinue(args) {
-  const taskId = args._[1];
-
-  if (!taskId) {
-    console.error('Error: task_id is required for continue');
-    console.error('Usage: amicus continue <task_id> --prompt "..."');
-    process.exit(1);
-  }
-
-  const taskIdCheck = validateTaskId(taskId);
-  if (!taskIdCheck.valid) {
-    console.error(taskIdCheck.error);
-    process.exit(1);
-  }
-
-  // BL-1: accept --prompt-file (XOR --prompt) so the MCP handler can pass a long
-  // follow-up prompt via file, dodging the ~32KB Windows command-line cap.
-  if (args['prompt-file'] !== undefined) {
-    const { resolvePromptSource } = require('../src/utils/prompt-source');
-    const promptRes = resolvePromptSource(args);
-    if (promptRes.error) {
-      console.error(promptRes.error);
-      process.exit(1);
-    }
-    args.prompt = promptRes.prompt;
-    delete args['prompt-file'];
-  }
-
-  if (!args.prompt && !args.briefing) {
-    console.error('Error: --prompt is required for continue');
-    process.exit(1);
-  }
-
-  // F5: an explicitly passed --model gets the same resolution+validation as start.
-  if (args.model !== undefined) {
-    const { model, alias } = resolveModelFromArgs(args);
-    args.model = model;
-    args.model = await validateFallbackModel(args, alias);
-  }
-
-  const { continueAmicus } = require('../src/index');
-
-  return await continueAmicus({
-    taskId,
-    newTaskId: args['task-id'],
-    briefing: args.prompt || args.briefing,
-    model: args.model,
-    project: args.cwd,
-    contextTurns: args['context-turns'],
-    contextMaxTokens: args['context-max-tokens'],
-    headless: args['no-ui'],
-    timeout: args.timeout
-  });
-}
-
 // Run main
 main().catch(err => {
   console.error(`Fatal error: ${err.message}`);

package/commands/council.md

@@ -1,6 +1,6 @@
 ---
 description: Run a structured multi-model LLM council review of the given material — wraps the second-opinion skill (independent reviews → anonymous cross-review → non-Claude chair verdict → accept/deny decisions).
-argument-hint: [material, path, or URL] [analysis request + criteria]
+argument-hint: [material, path, or URL] [analysis request + criteria] [optional elements — e.g. "with a critic seat", "debate mode", "expert lenses", "chair verdict scale"]
 disable-model-invocation: true
 ---
 
@@ -17,8 +17,16 @@ the **analysis request**, and the **criteria**. If any of the three is missing o
 ambiguous, ask for it before launching any model (the skill's Stage 0 covers this —
 don't re-ask for what is already present).
 
-Then follow the second-opinion skill end to end: Stage 0 intake/prep and run-folder
-setup, council selection with a cost estimate and explicit user confirmation, the
-three review waves, `amicus council validate` on each leg's findings block,
-`amicus council tally`, the accept/deny decision pass, and `amicus council verdict`
-to write the final `verdict.json`.
+Then follow the second-opinion skill end to end, in pipeline order: Stage 0
+intake/prep and run-folder setup, then council selection with a cost estimate
+and explicit user confirmation; Stage 1 independent reviews, running
+`amicus council validate` on each leg's findings block as it lands; Stage 2
+cross-review, followed by `amicus council tally` once cross-review settles;
+Stage 3 council-chair synthesis; Stage 4 the accept/deny decision pass; and
+Stage 5, which runs `amicus council verdict` to write the final `verdict.json`.
+
+The user may also name **optional council elements** in the arguments (critic seat,
+expert lenses, debate mode, verdict scale, Claude in the council). All elements
+default OFF; the skill's Stage-0 menu is the single opt-in point. If the user named
+elements here, carry them into Stage 0 as pre-requested — confirm them back by name
+instead of re-asking — and never enable an element the user did not explicitly name.

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "amicus",
-  "version": "2.0.0",
+  "version": "2.2.0",
   "mcpName": "io.github.BourbonDog/amicus",
   "description": "Multi-model LLM Council + parallel AI window for Claude Code. Run structured council reviews across Gemini, GPT, DeepSeek and more — or fork a conversation to any model and fold the results back.",
   "keywords": [

package/skills/second-opinion/COUNCIL-DESIGN.md

@@ -1,8 +1,8 @@
 # Second Opinion v3 — "LLM Council" Design
 
 _Status: implemented (v3). v2 (2026-06-03) added the council mechanics; v3 (2026-06-10) swapped the
-transport onto the Amicus fanout/JSON engine primitives. v2 history lives in git
-(`V2-COUNCIL-DESIGN.md`, deleted at v3)._
+transport onto the Amicus fanout/JSON engine primitives; v3.1 (2026-07-14) added the optional
+council elements (§12). v2 history lives in git (`V2-COUNCIL-DESIGN.md`, deleted at v3)._
 _Design for `SKILL.md` and `MODEL-NOTES.md` of the `second-opinion` skill._
 
 ## 1. Intent
@@ -224,6 +224,70 @@ The append-only `council-ledger.jsonl` (consumed via `amicus council stats`) is
   four qualitative tiers, if tiers prove too coarse in practice.
 
 ## 11. Implementation surface
-- `SKILL.md` — the Stage 0–6 council flow on the v3 transport (WS-3: findings contract, tally assembly recipe, `amicus council tally/stats`, `verdict.json`, ledger auto-append).
+- `SKILL.md` — the Stage 0–6 council flow on the v3 transport (WS-3: findings contract, tally assembly recipe, `amicus council tally/stats`, `verdict.json`, ledger auto-append). v3.1 adds the Stage-0 optional-elements menu, the Stage-1 critic/lens launch blocks, Stage 2.5, and the chair verdict-scale addendum.
+- `SEAT-BRIEFS.md` — v3.1 briefing boilerplate for the optional elements + the standard anti-sycophancy clause.
 - `MODEL-NOTES.md` — qualitative per-model quirks, structural-conformance notes, cost guardrail, Stage-2 briefing tips. Quantitative reliability data now generated by `amicus council stats` (ledger). Engine workarounds that F1/F2/F4 made obsolete were pruned at v3.
-- `src/council/` — the deterministic helpers (`findings.js`, `tally.js`, `verdict.js`, `ledger.js`). No other files.
+- `src/council/` — the deterministic helpers (`findings.js`, `tally.js`, `verdict.js`, `ledger.js`). No other files. **v3.1 required zero engine changes**: `runStats[].role` labels are free-form, the rebuttal round is a tally-input re-assembly, and `--no-ledger` already existed.
+
+## 12. Optional council elements (v3.1)
+
+The `/critic` and `/debate` behaviors from John Renaldi's product-kit plugin (MIT — the
+elite-advisor and expert-debate-facilitator agents), grafted onto the council. Both source
+agents are single-model theater: one model playing an adversarial reviewer or a panel of
+disagreeing experts, with no external check on manufactured negativity or fake disagreement.
+The council supplies exactly that check — independent models and deterministic peer-validated
+tiers — so the methodologies port over as *optional elements* whose output is disciplined by
+Stage-2 adjudication.
+
+**Opt-in contract.** All elements default OFF. They are presented once, as a single numbered
+menu at Stage 0 (after bench/chair selection, before launch confirmation); the user must name
+an element to enable it — silence means off, and the launch confirmation enumerates enabled
+elements by name. "Claude in the council" (§5.4, semantics unchanged) is presented as one item
+of the same menu so there is exactly one opt-in point per run. Elements compose freely except
+critic + lenses on one seat (the critic takes one seat; lenses take the rest).
+
+### 12.1 Critic seat
+One bench member swaps its standard brief for an adversarial four-pass brief (adversarial
+pass, edge-case hunt, consistency check, executability test — `SEAT-BRIEFS.md`). Launches as
+a concurrent solo run (the red-team variant pattern); same findings contract; `role: "critic"`.
+Judges are never told the seat exists. Two deliberate deviations from the source agent: the
+minimum-10-findings quota is dropped (quotas force invented findings, which waste the bench's
+adjudication capital — the anti-padding rule replaces it), and the hard questions +
+Ship/Fix/Rethink verdict moved to the chair (§12.4). Known cost: the critic model can
+self-identify its review in the Stage-2 bundle by shape, weakening self-bias wash-out for that
+one seat — disclosed in the report. A single-seat brief variant is judged acceptable to record
+to the ledger (precedent: the red-team variant records normally).
+
+### 12.2 Expert lenses
+Each seat gets a distinct expert-perspective brief (panel domain scoped with the user:
+business/technical/specialty/customer/financial/custom). All legs launch as concurrent solos
+(no shared-prompt wave); lens↔model assignment is random and private; `role: "lens:<slug>"`.
+Two standing caveats, both disclosed: (a) **anonymity leak** — every judge can spot its own
+lens-flavored review, so self-bias wash-out is weakened bench-wide; (b) **ledger pollution** —
+lens reviews are not comparable to standard reviews (street-cred and confirm-rate assume a
+shared task), so lens-run tallies always pass `--no-ledger`. The bench-wide task change is what
+separates lenses (no ledger) from the single-seat critic (ledger ok).
+
+### 12.3 Debate mode (Stage 2.5 rebuttal round)
+The debate agent's challenge/stress-test rounds, grounded in findings instead of persona
+prose: the Stage-2 tally runs `--no-ledger` (provisional); every Contested/Disputed finding
+goes back to its raiser (defense mini-wave — solo per raiser: DEFEND / AMEND / WITHDRAW), the
+disputing judges re-vote on defended/amended findings (one shared fanout), and the final tally
+runs with the ledger. Exactly one round — the source agent's open-ended rounds become a fixed
+convergence step, or cost is unbounded. Withdrawn findings stay in `findings[]` and are
+auto-recorded `denied` (never presented for user decision). Missing/unparseable rebuttal or
+re-vote lines mean the original claim/verdict stands. No new decision vocabulary and no engine
+changes: `buildVerdict` and the renderer see a normal tally record.
+
+### 12.4 Chair verdict scale
+A chair-packet addendum: the chair closes with 3–5 hard questions plus a final parseable
+`VERDICT: Ship it | Fix these first | Fundamental rethink` line, surfaced verbatim at the top
+of `report.md` and in chat. Gives the council the single overall call its tier counts never
+provided. Possible later engine touch: an `overallVerdict` field in `verdict.json` + a
+renderer line (not needed for the prose version).
+
+### 12.5 Future work — parallel panels
+The debate agent's parallel-panels + cross-panel synthesis maps to two full council runs over
+different benches or lens-sets plus a super-chair synthesis of where the panels agree,
+disagree, and what each missed. Deferred: cost doubles and the run-folder conventions need
+namespacing. Revisit if lens runs prove valuable.

package/skills/second-opinion/MODEL-NOTES.md

@@ -4,8 +4,8 @@ This file is the `second-opinion` skill's evolving memory of **how to actually d
 well**. Read it before Stage 0 (council selection and launch); update it, with the user's
 approval, at the end of each run (Stage 6). Keep it tight — merge and prune rather than append.
 
-_Last updated: 2026-07-02 (runs 4-7 folded back from the field ledger: PowerShell quoting, date
-injection, long-read failures, judge tool-wander; see changelog)._
+_Last updated: 2026-07-14 (v2.2.0 fold-back: v3.1 optional council elements verified live;
+claim-class dedup adjudication limit; minimax and qwen-coder debut notes; see changelog)._
 
 ## Global operating rules (all models)
 - **Council runs are headless by design** (autonomous batch work): `fanout` is headless by
@@ -54,6 +54,19 @@ injection, long-read failures, judge tool-wander; see changelog)._
   no-tools preamble is now mandatory in SKILL.md Stage 2/3 — keep it verbatim.
 - The tally input needs **all five keys** (`meta` incl. `meta.models`, `findings`, `adjudications`,
   `rankings`, `runStats`) — see the SKILL.md Stage-2 recipe step 0.
+- **Claim-class dedup glosses rationale errors.** When several reviews raise the same claim, judges
+  adjudicate the class ("same as A3 — agree") and skip verifying each finding's own rationale text:
+  a direction-inverted arithmetic detail inside an otherwise-correct finding drew unanimous agrees
+  despite an explicit "dispute material factual errors in the rationale" instruction. If
+  rationale-level precision matters, instruct judges to independently verify the numbers in each
+  finding; otherwise expect class-level adjudication and let the chair reconcile details (a chair
+  has caught and corrected such an error unprompted).
+- **Optional council elements (v2.2.0) verified live:** critic seat (solo-alongside-fanout;
+  `role: "critic"` passes through `council tally` untouched), debate mode's nothing-to-debate path
+  (provisional `--no-ledger` tally → skip rebuttals → final ledger-recorded tally), and the chair
+  verdict scale (parseable `VERDICT:` line + hard questions) all behaved per SEAT-BRIEFS. **Debate
+  mode's defense/re-vote waves remain unexercised** — an all-Confirmed consensus run has no
+  rebuttal surface; exercise them on a contentious artifact before trusting that path blind.
 
 ## Per-model notes
 
@@ -102,6 +115,21 @@ injection, long-read failures, judge tool-wander; see changelog)._
 ### Claude  (in-council, when toggle on)
 - Consistently the most *calibrated* reviewer (no severity inflation; findings overwhelmingly Confirmed; bench-best street-cred in recent runs) but sometimes the least *original* — it can miss the boldest single catch. Treat as a reliability floor, not a discovery engine.
 
+### minimax  (`--model minimax` → via OpenRouter)
+- Fast (~2 min review legs), cheap, `clean` findings-JSON conformance on debut.
+- Took the **critic seat** brief exceptionally well: unanimously ranked #1 by its bench, full
+  coverage on a ground-truth test, zero padded findings under the anti-padding rule. A strong
+  default critic-seat candidate.
+- **Quirks:** occasional CJK character intrusions mid-English prose (cosmetic; the findings JSON
+  is unaffected); can invert the direction of an arithmetic detail inside an otherwise-correct
+  rationale — see the claim-class-dedup tip in Stage-2 briefing tips.
+
+### qwen-coder  (`--model qwen-coder` → via OpenRouter; distinct from `qwen`)
+- Very fast (16–22 s legs, review and judging alike), `clean` conformance, competent
+  core-blocker coverage on prose/PRD artifacts.
+- Ranked last on its debut bench — misses the offline, interruption-handling, and
+  test-methodology gap classes stronger seats catch. Fine budget-bench filler; do not chair it.
+
 ### (others — add as used)
 - Opus / o-series etc. are reachable via amicus **if their API keys are configured**. Add notes
   here the first time each is used.
@@ -160,3 +188,8 @@ This section keeps only per-model **qualitative quirks** and **structural-confor
   GUI-hangs-on-this-machine rule (resolved 2026-06-10; headless stays the council default by
   design). Config path updated to `~/.config/amicus/.env`.
 - **2026-07-02** — Folded back field lessons from runs 4-7 (AV-receiver, pork-shoulder, resume, novel ×2 councils): PowerShell `--models` quoting; current-date injection; long-read model selection; judge no-tools preamble; severity-inflation-justifies-dispute; five-keys tally schema; new Grok/Kimi/Mistral/Claude-in-council sections. Quantitative history stays in the ledger (`amicus council stats`).
+- **2026-07-14 (v2.2.0)** — Optional council elements shipped and verified on a planted-flaw
+  ground-truth council (critic seat, debate mode nothing-to-debate path, chair verdict scale;
+  expert lenses defined but not yet field-run). New lessons: claim-class dedup glosses
+  rationale-level errors in Stage-2; minimax debut (strong critic seat, CJK-intrusion quirk);
+  qwen-coder debut (fast budget filler). Debate mode's defense/re-vote waves still unexercised.