@polderlabs/bizar 10.19.1 → 10.19.3

package/cli/commands/models.mjs

@@ -418,14 +418,27 @@ export function defaultTierHint(modelId) {
   return 'mid';
 }
 
-function classifyKind(modelId) {
+export function classifyKind(modelId) {
   const id = String(modelId || '').toLowerCase();
+  // Order matters: more-specific legacy prefixes are checked BEFORE the
+  // generic `claude-` catch-all so `claude-qwen/...` and
+  // `claude-minimax/...` are tagged with their real family, not lumped
+  // into the generic `claude` bucket. Live gateway prefixes follow the
+  // same rule.
+  if (id.startsWith('claude-qwen/')) return 'claude-qwen';
+  if (id.startsWith('claude-minimax/')) return 'claude-minimax';
   if (id.startsWith('claude-')) return 'claude';
   if (id.startsWith('cx/')) return 'cx';
   if (id.startsWith('oc/')) return 'oc';
-  if (id.startsWith('claude-minimax/')) return 'claude-minimax';
-  if (id.startsWith('claude-qwen/')) return 'claude-qwen';
   if (id.startsWith('anthropic/')) return 'anthropic';
+  // Live gateway namespace (10.19.2+): bare provider/model forms exposed by
+  // OmniRoute at https://route.polderlabs.io/v1.
+  if (id.startsWith('minimax/')) return 'minimax';
+  if (id.startsWith('codex/')) return 'codex';
+  if (id.startsWith('glm/')) return 'glm';
+  if (id.startsWith('qct/')) return 'qct';
+  if (id.startsWith('openrouter/')) return 'openrouter';
+  if (id.startsWith('a/')) return 'a';
   return 'other';
 }
 
@@ -483,6 +496,205 @@ function writeAtomic(path, body) {
   renameSync(tmp, path);
 }
 
+// ── F-191 / 10.19.2 sync to Claude Code settings.json + stale-ID gating ─────
+
+/**
+ * Detect model IDs that the live gateway does not serve. Picker picks
+ * survive gateway 404s only when the user is warned at save time. Stale
+ * IDs are written into `userSelected.staleIds` (audit trail) and excluded
+ * from the settings.json sync so Claude Code does not see a `[claude-code:
+ * unrecognized_model]` for a known-dead ID on every turn.
+ *
+ * @param {{ liveIds: string[], pickedIds: string[] }} opts
+ * @returns {{ liveIds: string[], staleIds: string[], unknownIds: string[] }}
+ *   - `liveIds` — picks that exist in the live gateway pool.
+ *   - `staleIds` — picks that were never returned by the gateway in this run.
+ *   - `unknownIds` — picks whose live status is unknown (no live pool yet).
+ */
+export function partitionStalePicks({ liveIds, pickedIds }) {
+  const live = Array.isArray(liveIds) ? new Set(liveIds) : null;
+  const picks = Array.isArray(pickedIds) ? pickedIds.filter((id) => typeof id === 'string' && id.trim()) : [];
+  if (!live || live.size === 0) {
+    return { liveIds: [], staleIds: [], unknownIds: [...new Set(picks)] };
+  }
+  const liveOut = [];
+  const stale = [];
+  for (const id of picks) {
+    if (live.has(id)) liveOut.push(id);
+    else stale.push(id);
+  }
+  return { liveIds: liveOut, staleIds: stale, unknownIds: [] };
+}
+
+/**
+ * Sync `userSelected.models` into Claude Code's settings.json under
+ * `modelOverrides` using the self-map pattern (`<id>` → `<id>`). This
+ * suppresses `[claude-code:unrecognized_model]` diagnostics on every turn
+ * for any picked ID the gateway serves. Stale IDs (not returned by the
+ * gateway) are excluded so the diagnostic still surfaces them.
+ *
+ * Behavior:
+ *   - Reads `settings.json` if present; preserves every other field.
+ *   - Writes `modelOverrides` as a sparse object: only picked IDs that
+ *     are also in `liveIds`. Self-map pattern keeps dispatch behavior
+ *     identical (Claude Code dispatches the literal ID).
+ *   - Atomic replace via temp-file + rename (matches `applyModels`).
+ *   - When `settingsJsonPath` is provided (tests), uses that instead of
+ *     `~/.claude/settings.json`.
+ *   - When `settingsJsonPath` is `null`, skips the sync entirely — used
+ *     by tests that want to exercise the picker without touching Claude
+ *     Code's real settings file.
+ *
+ * @param {{
+ *   settingsJsonPath?: string|null,
+ *   pickedIds: string[],
+ *   liveIds?: string[],
+ * }} opts
+ * @returns {{
+ *   wrote: boolean,
+ *   syncedIds: string[],
+ *   skippedStale: string[],
+ *   settingsPath: string|null,
+ * }}
+ */
+export function applyModelOverrides({ settingsJsonPath, pickedIds, liveIds = [] }) {
+  const path = settingsJsonPath === undefined
+    ? join(homedir(), '.claude', 'settings.json')
+    : settingsJsonPath;
+  if (path === null) {
+    return { wrote: false, syncedIds: [], skippedStale: [], settingsPath: null };
+  }
+  const live = new Set(Array.isArray(liveIds) ? liveIds : []);
+  const picks = Array.isArray(pickedIds) ? pickedIds.filter((id) => typeof id === 'string' && id.trim()) : [];
+  // Self-map only IDs the live gateway serves. Stale IDs intentionally stay
+  // out so the `[claude-code:unrecognized_model]` diagnostic still fires
+  // for them — the operator should re-run `bizar models` to drop them.
+  const synced = live.size === 0
+    ? picks
+    : picks.filter((id) => live.has(id));
+  const skipped = live.size === 0 ? [] : picks.filter((id) => !live.has(id));
+
+  let settings = {};
+  if (existsSync(path)) {
+    try {
+      const parsed = JSON.parse(readFileSync(path, 'utf8'));
+      if (parsed && typeof parsed === 'object' && !Array.isArray(parsed)) settings = parsed;
+    } catch {
+      // Corrupt settings.json — refuse to overwrite it; surface the sync
+      // as a no-op so `bizar models` still completes.
+      return { wrote: false, syncedIds: [], skippedStale: skipped, settingsPath: path };
+    }
+  }
+  // Self-map pattern — Claude Code uses modelOverrides to suppress
+  // `[claude-code:unrecognized_model]` for any ID that maps to itself.
+  settings.modelOverrides = Object.fromEntries(synced.map((id) => [id, id]));
+  writeAtomic(path, JSON.stringify(settings, null, 2) + '\n');
+  return { wrote: true, syncedIds: synced, skippedStale: skipped, settingsPath: path };
+}
+
+/**
+ * Derive a human-readable label for a model ID. Used to populate the
+ * `modelPicker` array in settings.json so Claude Code's `/model` picker
+ * displays picked models with something nicer than the raw `provider/name`
+ * string.
+ *
+ *   minimax/MiniMax-M3              → "MiniMax M3"
+ *   codex/gpt-5.6-sol               → "GPT 5.6 Sol"
+ *   qct/qwen3.8-max-preview         → "Qwen3.8 Max Preview"
+ *   openrouter/nvidia/foo:free      → "nvidia/foo:free"
+ *
+ * The gateway-reported `name` (when available on the picked profile) always
+ * wins. Falls back to a title-cased rendering of the model segment.
+ *
+ * @param {string} modelId
+ * @param {object} [profile] Optional profile with `name` or `displayName`
+ * @returns {string}
+ */
+export function deriveModelLabel(modelId, profile) {
+  const name = profile && typeof profile.name === 'string' && profile.name.trim();
+  if (name) return name.trim();
+  const displayName = profile && typeof profile.displayName === 'string' && profile.displayName.trim();
+  if (displayName) return displayName.trim();
+  const id = String(modelId || '').trim();
+  if (!id) return '';
+  // Drop the leading provider segment (`minimax/MiniMax-M3` → `MiniMax-M3`)
+  // so the operator sees the model name, not the namespace.
+  const slash = id.indexOf('/');
+  const tail = slash >= 0 ? id.slice(slash + 1) : id;
+  // Split on word boundaries (hyphens / underscores / dots / colons / path
+  // separators) and join with spaces. Case is preserved verbatim — `gpt`
+  // stays `gpt`, `MiniMax` stays `MiniMax`, `M2.7` stays `M2.7`. Brand
+  // casing belongs to the gateway (`name` field), not us.
+  return tail
+    .replace(/[\\/]+/g, ' ')
+    .replace(/[-_]+/g, ' ')
+    .replace(/:/g, ' ')
+    .replace(/\s+/g, ' ')
+    .trim();
+}
+
+/**
+ * Sync `userSelected.models` into Claude Code's `modelPicker` array
+ * (settings.json). The picker is what populates `/model` — `modelOverrides`
+ * alone only silences diagnostics, it does NOT add entries to the picker.
+ *
+ * Per Claude Code's settings reference: `modelPicker` is an array of
+ * `{ id, label }` entries that replaces the gateway-discovered picker
+ * contents. Scope is User-or-managed (settings.json is in scope). Each
+ * entry preserves the operator's pick order from `userSelected.models`.
+ *
+ * Behavior:
+ *   - Reads settings.json; preserves every other field (env, mcpServers,
+ *     permissions, hooks, etc.).
+ *   - Writes `modelPicker` as an array of `{ id, label }`.
+ *   - Filters out picks the live gateway rejects (same stale-ID contract as
+ *     `applyModelOverrides`); the surviving picks fill the picker.
+ *   - Empty pick list → writes `modelPicker: []` so the picker falls back
+ *     to whatever Claude Code's defaults surface.
+ *   - Atomic replace via temp-file + rename (matches `applyModels`).
+ *   - Refuses to overwrite a corrupt settings.json.
+ *   - When `settingsJsonPath === null`, returns a no-op (tests).
+ *
+ * @param {{
+ *   settingsJsonPath?: string|null,
+ *   pickedIds: string[],
+ *   profiles?: Record<string, object>,
+ *   liveIds?: string[],
+ * }} opts
+ * @returns {{
+ *   wrote: boolean,
+ *   entries: Array<{id: string, label: string}>,
+ *   skippedStale: string[],
+ *   settingsPath: string|null,
+ * }}
+ */
+export function applyModelPicker({ settingsJsonPath, pickedIds, profiles = {}, liveIds = [] }) {
+  const path = settingsJsonPath === undefined
+    ? join(homedir(), '.claude', 'settings.json')
+    : settingsJsonPath;
+  if (path === null) {
+    return { wrote: false, entries: [], skippedStale: [], settingsPath: null };
+  }
+  const live = new Set(Array.isArray(liveIds) ? liveIds : []);
+  const picks = Array.isArray(pickedIds) ? pickedIds.filter((id) => typeof id === 'string' && id.trim()) : [];
+  const surviving = live.size === 0 ? picks : picks.filter((id) => live.has(id));
+  const skipped = live.size === 0 ? [] : picks.filter((id) => !live.has(id));
+  const entries = surviving.map((id) => ({ id, label: deriveModelLabel(id, profiles?.[id]) }));
+
+  let settings = {};
+  if (existsSync(path)) {
+    try {
+      const parsed = JSON.parse(readFileSync(path, 'utf8'));
+      if (parsed && typeof parsed === 'object' && !Array.isArray(parsed)) settings = parsed;
+    } catch {
+      return { wrote: false, entries: [], skippedStale: skipped, settingsPath: path };
+    }
+  }
+  settings.modelPicker = entries;
+  writeAtomic(path, JSON.stringify(settings, null, 2) + '\n');
+  return { wrote: true, entries, skippedStale: skipped, settingsPath: path };
+}
+
 // ── F-190 / IMP-017 refresh + operator-override preservation ────────────────
 
 /**
@@ -1294,8 +1506,15 @@ export async function run(name, args, isHelpRequest) {
       process.exit(2);
     }
     const block = applyModels({ routerPath, models: ids, source: 'cli-set' });
+    // Sync to Claude Code's settings.json — picks survive stale-ID
+    // filtering only when an empty liveIds array disables it (no
+    // candidates fetched for `--set`). The orchestrator will surface
+    // `[claude-code:unrecognized_model]` for any ID the gateway later
+    // rejects; the operator can re-run `bizar models` to drop them.
+    const sync = applyModelOverrides({ pickedIds: ids, liveIds: [] });
+    const picker = applyModelPicker({ pickedIds: ids, profiles: block.profiles || {}, liveIds: [] });
     if (wantJson) {
-      process.stdout.write(JSON.stringify({ applied: block }, null, 2) + '\n');
+      process.stdout.write(JSON.stringify({ applied: block, sync, picker }, null, 2) + '\n');
     } else {
       console.log(chalk.green(`  v ${block.models.length} model(s) saved to userSelected`));
       for (const id of block.models) {
@@ -1389,6 +1608,10 @@ export async function run(name, args, isHelpRequest) {
   const picked = await pickModels({ candidates, current });
   if (picked.length === 0) {
     const block = applyModels({ routerPath, models: [], source: 'live-pick' });
+    // Clear Claude Code modelOverrides + modelPicker when the picker is
+    // emptied so the session no longer claims to recognise removed IDs.
+    applyModelOverrides({ pickedIds: [], liveIds: [] });
+    applyModelPicker({ pickedIds: [], liveIds: [] });
     if (wantJson) {
       process.stdout.write(JSON.stringify({ applied: block, endpoint, endpointSource }, null, 2) + '\n');
     } else {
@@ -1403,9 +1626,23 @@ export async function run(name, args, isHelpRequest) {
     const profile = candidates.find((candidate) => candidate.id === id)?.profile;
     if (profile) profiles[id] = profile;
   }
+  // F-191 / 10.19.2 — stale-ID detection: the picker shows candidates
+  // filtered through the live gateway, but `--set` and the picker can both
+  // accept IDs the gateway later rejects (e.g. `a/1`). Persist them as
+  // `staleIds` for audit and skip them in the settings.json sync so Claude
+  // Code still surfaces the unrecognized_model diagnostic.
+  const liveIds = candidates.map((c) => c.id);
+  const partition = partitionStalePicks({ liveIds, pickedIds: picked });
   const block = applyModels({ routerPath, models: picked, tierHints, profiles, source: 'live-pick' });
+  if (partition.staleIds.length > 0) {
+    block.staleIds = partition.staleIds;
+  }
+  const sync = applyModelOverrides({ pickedIds: picked, liveIds });
+  // Sync the /model picker contents (`modelPicker` setting) so the user's
+  // picks drive the picker without relying on gateway discovery.
+  const picker = applyModelPicker({ pickedIds: picked, profiles, liveIds });
   if (wantJson) {
-    process.stdout.write(JSON.stringify({ applied: block, endpoint, endpointSource }, null, 2) + '\n');
+    process.stdout.write(JSON.stringify({ applied: block, endpoint, endpointSource, sync, picker }, null, 2) + '\n');
   } else {
     console.log(chalk.green(`\n  v Saved ${block.models.length} model(s) to ${routerPath}:`));
     console.log(chalk.dim(`    Models.dev profiles: ${Object.keys(block.profiles || {}).length}/${block.models.length}`));
@@ -1413,6 +1650,12 @@ export async function run(name, args, isHelpRequest) {
       const tier = block.tierHints[id] || defaultTierHint(id);
       console.log(chalk.dim(`    ${id}  (${tier})`));
     }
+    if (sync.wrote && sync.skippedStale.length > 0) {
+      console.log(chalk.yellow(`    Settings sync skipped ${sync.skippedStale.length} stale id(s): ${sync.skippedStale.join(', ')}`));
+    }
+    if (picker.wrote) {
+      console.log(chalk.dim(`    /model picker populated with ${picker.entries.length} entr${picker.entries.length === 1 ? 'y' : 'ies'}`));
+    }
   }
   return true;
 }

package/config/claude/commands/use-premium.md

@@ -1,11 +1,11 @@
 ---
-description: Launch Claude Code in this project on the premium tier (claude-qwen/qwen3.8-max). Prints the exact shell command and env vars to set.
+description: Launch Claude Code in this project on the premium tier (qct/qwen3.8-max-preview). Prints the exact shell command and env vars to set.
 allowed-tools: Bash
 ---
 
-# /use-premium — Run this session on claude-qwen/qwen3.8-max (premium)
+# /use-premium — Run this session on qct/qwen3.8-max-preview (premium)
 
-The provider gateway you have configured exposes `claude-qwen/qwen3.8-max`
+The provider gateway you have configured exposes `qct/qwen3.8-max-preview`
 as the premium-tier reasoning model. Bizar is provider-agnostic — set
 `BIZAR_MODEL_ROUTER_URL` (or `ANTHROPIC_BASE_URL`) to whatever gateway
 URL your operator runs, then point the model at the premium tier. Premium
@@ -13,7 +13,7 @@ is expensive — use it intentionally, not by default.
 
 ## What it does
 
-Sets `ANTHROPIC_MODEL=claude-qwen/qwen3.8-max` and forwards
+Sets `ANTHROPIC_MODEL=qct/qwen3.8-max-preview` and forwards
 `ANTHROPIC_BASE_URL` from your operator-configured gateway so the entire
 session loop runs on the premium model. The model-router is the same
 one your subagents (`@mike`, `@paul`, `@carl`) already hit.
@@ -24,7 +24,7 @@ one your subagents (`@mike`, `@paul`, `@carl`) already hit.
 
 ```bash
 ANTHROPIC_BASE_URL="${BIZAR_MODEL_ROUTER_URL:-http://your-gateway/v1}" \
-ANTHROPIC_MODEL=claude-qwen/qwen3.8-max \
+ANTHROPIC_MODEL=qct/qwen3.8-max-preview \
 claude
 ```
 
@@ -33,7 +33,7 @@ router as the model id):
 
 ```bash
 ANTHROPIC_BASE_URL="${BIZAR_MODEL_ROUTER_URL:-http://your-gateway/v1}" \
-claude --model claude-qwen/qwen3.8-max
+claude --model qct/qwen3.8-max-preview
 ```
 
 ### Make it the default for this project
@@ -44,7 +44,7 @@ Add to your local override `.claude/settings.local.json`:
 {
   "env": {
     "ANTHROPIC_BASE_URL": "${BIZAR_MODEL_ROUTER_URL}",
-    "ANTHROPIC_MODEL": "claude-qwen/qwen3.8-max"
+    "ANTHROPIC_MODEL": "qct/qwen3.8-max-preview"
   }
 }
 ```
@@ -54,9 +54,9 @@ Then every Claude Code session in this directory opens on premium.
 ### Available premium-tier models on the local router
 
 ```
-claude-qwen/qwen3.8-max  — strongest reasoning (orchestration, planning, debug)
-cx/gpt-5.6-terra        — high reasoning at lower cost (complex impl, review)
-cx/gpt-5.6-luna         — visually-oriented mid-tier (design)
+qct/qwen3.8-max-preview  — strongest reasoning (orchestration, planning, debug)
+codex/gpt-5.6-sol        — high reasoning at lower cost (complex impl, review)
+codex/gpt-5.6-luna       — visually-oriented mid-tier (design)
 ```
 
 See `.claude/model-router.json` for the full tier table.
@@ -70,9 +70,9 @@ See `.claude/model-router.json` for the full tier table.
 
 ## When NOT to use premium
 
-- Routine edits, mechanical work, single-file fixes (use default `claude-minimax/MiniMax-M3`).
+- Routine edits, mechanical work, single-file fixes (use default `minimax/MiniMax-M3`).
 - Research (always use `@greg` on default).
-- Implementation at moderate complexity (use `claude-minimax/MiniMax-M2.7` via `@todd`).
+- Implementation at moderate complexity (use `minimax/MiniMax-M2.7` via `@todd`).
 
 The harness's cost ceiling (`costCeilingPerSessionUsd: 5.0` in
 `model-router.json`) will downgrade you automatically if the budget is

package/config/claude/model-router.json

@@ -20,48 +20,51 @@
   "tiers": {
     "premium": {
       "models": [
-        "claude-qwen/qwen3.8-max",
-        "cx/gpt-5.6-terra"
+        "qct/qwen3.8-max-preview",
+        "codex/gpt-5.6-sol",
+        "codex/gpt-5.6-luna"
       ],
       "purpose": "Architecture, high-risk planning, adversarial verification, and last-resort debugging where mistakes are expensive.",
       "effort": "high"
     },
     "high": {
       "models": [
-        "cx/gpt-5.6-terra",
-        "claude-qwen/qwen3.8-max"
+        "codex/gpt-5.6-sol",
+        "qct/qwen3.8-max-preview"
       ],
       "purpose": "Complex implementation, security review, and cross-cutting reasoning.",
       "effort": "high"
     },
     "mid-design": {
       "models": [
-        "cx/gpt-5.6-luna",
-        "claude-minimax/MiniMax-M3"
+        "codex/gpt-5.6-luna",
+        "minimax/MiniMax-M3"
       ],
       "purpose": "UI/UX design and product-quality implementation where visual judgment matters.",
       "effort": "medium"
     },
     "default": {
       "models": [
-        "claude-minimax/MiniMax-M3",
-        "claude-minimax/MiniMax-M2.7"
+        "minimax/MiniMax-M3",
+        "minimax/MiniMax-M2.7"
       ],
       "purpose": "Broad research, git operations, and ordinary engineering work.",
       "effort": "medium"
     },
     "mid": {
       "models": [
-        "claude-minimax/MiniMax-M2.7",
-        "claude-minimax/MiniMax-M3"
+        "minimax/MiniMax-M2.7",
+        "minimax/MiniMax-M2.7-highspeed",
+        "qct/deepseek-v4-pro",
+        "openrouter/nvidia/nemotron-3-ultra-550b-a55b:free"
       ],
       "purpose": "Bounded technical analysis, code search, and moderate implementation.",
       "effort": "medium"
     },
     "budget": {
       "models": [
-        "claude-minimax/MiniMax-M2.5",
-        "claude-minimax/MiniMax-M2.7"
+        "glm/glm-5.3-flash",
+        "minimax/MiniMax-M2.7-highspeed"
       ],
       "purpose": "Routine deterministic edits, targeted clarification, and scripted verification.",
       "effort": "low"

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@polderlabs/bizar",
-  "version": "10.19.1",
+  "version": "10.19.3",
   "description": "Autonomous, human-in-the-loop multi-agent harness for Claude Code with guarded workflows, typed SDK primitives, and MCP tools.",
   "type": "module",
   "bin": {

package/packages/sdk/dist/version.d.ts

@@ -1,5 +1,5 @@
 /**
  * SDK version constant. Keep synchronized with the workspace package versions.
  */
-export declare const SDK_VERSION: "10.19.1";
+export declare const SDK_VERSION: "10.19.3";
 //# sourceMappingURL=version.d.ts.map

package/packages/sdk/dist/version.js

@@ -1,5 +1,5 @@
 /**
  * SDK version constant. Keep synchronized with the workspace package versions.
  */
-export const SDK_VERSION = "10.19.1";
+export const SDK_VERSION = "10.19.3";
 //# sourceMappingURL=version.js.map

package/packages/sdk/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@polderlabs/bizar-sdk",
-  "version": "10.19.1",
+  "version": "10.19.3",
   "type": "module",
   "main": "./dist/index.js",
   "types": "./dist/index.d.ts",