@polderlabs/bizar 10.23.7 → 10.23.9

package/AGENTS.md

@@ -58,21 +58,18 @@ still deny prohibited actions and escalate externally visible or irreversible
 actions with `permissionDecision: "ask"`; that escalation list is the
 authoritative floor, not a starting point.
 
-Native dynamic workflows under `config/workflows/` and `~/.claude/workflows/`
-are the required dispatch mechanism for every primary request except an
-unmistakably tiny single-target copy/style/format edit with no behavior or test
-change. Mike may execute that narrow exception directly. All other work enters
-the matching research / implement / debug / review workflow and uses at least
-one explicitly modeled editing subagent with call-level worktree isolation.
-Do not add unnecessary phases or duplicate workers. For work that needs 3+
-long-lived workers with bounded cross-talk,
-Mike invokes a workflow that fans out as a native agent team; the team is
+Mike selects the coordination mode after bounded read-only orientation and one
+user clarification checkpoint: direct work only for an unmistakably tiny
+single-target copy/style/format edit; one isolated Agent for a clear bounded
+change; a native workflow for repeatable phased work; parallel Agents for
+disjoint scopes; and an Agent team for 3+ sustained roles that genuinely need
+cross-talk. Do not add unnecessary phases or duplicate workers. The team is
 host-side state under `CLAUDE_CODE_EXPERIMENTAL_AGENT_TEAMS=1`, and per
-Anthropic's docs `team_name` is deprecated and ignored. When two or more
-subtasks have non-overlapping writable scopes and no data dependency, the
-orchestrator MUST dispatch them concurrently through `parallel([...])`, each
-with call-level `isolation: "worktree"`. Sequential dispatch is reserved for
-dependent phases and integration; never create artificial parallel work.
+Anthropic's docs `team_name` is deprecated and ignored. Every editing dispatch
+uses an explicit configured model and call-level `isolation: "worktree"`.
+When two or more subtasks have non-overlapping writable scopes and no data
+dependency, dispatch them concurrently; serialize only dependencies and
+integration.
 
 Agent roles are model-agnostic. Mike selects the cheapest sufficient enabled
 configured-tier model for each dispatch, with explicit user picks taking
@@ -136,11 +133,13 @@ The autonomy and approval policy above governs this execution model. The project
 Every non-empty primary request enters Bizar through `office-manager` (`@mike`).
 The installer sets Claude Code's global `agent` setting to Mike's frontmatter
 name (`mike`),
-and a session-scoped routing guard requires a successful native Workflow before
-substantive primary-session mutation. Read-only inspection remains available.
-Mike directly executes only the tiny edit exception above. Every other request
-must invoke a native Bizar workflow before mutation; the workflow dispatches
-worktree-isolated subagents while Mike owns integration and final verification.
+and the routing hook supplies the adaptive coordination policy. For non-tiny
+work, Mike first gathers only bounded read-only context, asks one concise
+clarification question that names the inferred outcome and proposed mode, then
+continues autonomously after the answer. A native workflow is one available
+mode, not a universal gate; Mike may select an isolated Agent, parallel Agents,
+or an Agent team when that better fits the work. Mike owns integration and
+final verification.
 A Bizar custom agent already executing its assigned role does not recursively
 dispatch itself.
 
@@ -152,8 +151,8 @@ relevant page. Guess-and-try integration work is prohibited. When official
 documentation is unavailable or ambiguous, inspect authoritative source code
 and report the evidence gap.
 
-For workflow-routed requests, `office-manager` uses only the phases that reduce
-a known risk; only the tiny edit exception skips this pipeline:
+For a workflow or team, `office-manager` uses only phases and members that
+reduce a known risk:
 
 1. Research: `greg` (`research-analyst.md`) plus an implementation-context specialist.
 2. Plan: `planner` drafts; `qa-reviewer` challenges assumptions and test shape.

package/cli/commands/models.mjs

@@ -947,8 +947,9 @@ export function partitionStalePicks({ liveIds, pickedIds, disabledProviders }) {
  *
  * Behavior:
  *   - Reads `settings.json` if present; preserves every other field.
- *   - Writes `modelOverrides` as a sparse object for picked IDs that are
- *     also in `liveIds`; Claude Code still dispatches each literal value.
+ *   - Maps every recognized Claude alias into picked IDs that are also in
+ *     `liveIds`, preventing internal alias normalization from reaching an
+ *     unconfigured provider default.
  *   - Atomic replace via temp-file + rename (matches `applyModels`).
  *   - When `settingsJsonPath` is provided (tests), uses that instead of
  *     `~/.claude/settings.json`.
@@ -1007,6 +1008,12 @@ export function applyModelOverrides({ settingsJsonPath, pickedIds, liveIds = [],
   // must be keys; configured gateway aliases are values. This also suppresses
   // print-mode `[claude-code:unrecognized_model]` diagnostics for Agent SDK calls.
   settings.modelOverrides = buildClaudeModelOverrides(synced);
+  if (requiresGatewayModelDiscovery(synced)) {
+    settings.env = {
+      ...(settings.env || {}),
+      CLAUDE_CODE_ENABLE_GATEWAY_MODEL_DISCOVERY: '1',
+    };
+  }
   const previousModel = typeof settings.model === 'string' ? settings.model : null;
   const previousContext = profiles?.[previousModel]?.limits?.contextTokens;
   const nextModel = synced[0] || null;
@@ -1068,8 +1075,20 @@ export function buildClaudeModelOverrides(modelIds) {
   const unique = [...new Set((Array.isArray(modelIds) ? modelIds : [])
     .filter((id) => typeof id === 'string' && id.trim())
     .map((id) => id.trim()))];
-  return Object.fromEntries(unique.slice(0, CLAUDE_MODEL_OVERRIDE_KEYS.length)
-    .map((id, index) => [CLAUDE_MODEL_OVERRIDE_KEYS[index], id]));
+  if (unique.length === 0) return {};
+  // Cover every built-in alias: the workflow runtime may normalize an Agent
+  // request through one of these names, and no alias may escape to an
+  // unconfigured Anthropic default.
+  return Object.fromEntries(CLAUDE_MODEL_OVERRIDE_KEYS
+    .map((key, index) => [key, unique[index % unique.length]]));
+}
+
+/** Custom gateway IDs must be discoverable to Claude's SDK/subagent path. */
+export function requiresGatewayModelDiscovery(modelIds) {
+  return (Array.isArray(modelIds) ? modelIds : []).some((id) => {
+    if (typeof id !== 'string' || !id.trim()) return false;
+    return !/^(?:claude(?:-|$)|anthropic(?:[./-]|$))/i.test(id.trim());
+  });
 }
 
 /**

package/cli/doctor.mjs

@@ -42,6 +42,7 @@ import {
   REQUIRED_HOOKS,
 } from './commands/validate.mjs';
 import { configuredEnabledModels, listModels, resolveEndpoint } from './commands/models.mjs';
+import { validateNativeWorkflowDirectory } from '../config/workflows/lib/native-contract.mjs';
 
 const REQUIRED_RULES = [
   'general.md', 'git.md', 'javascript.md', 'python.md',
@@ -157,6 +158,13 @@ async function checkHookFilesInstalled() {
   return `all ${REQUIRED_HOOKS.length} hook entrypoints installed`;
 }
 
+async function checkNativeWorkflowsInstalled() {
+  const dir = join(claudeDir(), 'workflows');
+  if (!existsSync(dir)) throw new Error(`workflows dir missing: ${dir} — run \`bizar update\``);
+  const result = validateNativeWorkflowDirectory(dir);
+  return `${result.count} native workflows parser-compatible and filename-addressable`;
+}
+
 /**
  * Lenient: passes if at least one of semble/skills/claude is on PATH.
  * These are informational — none of them are strictly required for
@@ -216,6 +224,7 @@ const CHECKS = [
   { name: 'skill-files-installed',     run: checkSkillFilesInstalled },
   { name: 'rule-files-installed',      run: checkRuleFilesInstalled },
   { name: 'hook-files-installed',      run: checkHookFilesInstalled },
+  { name: 'native-workflows-valid',    run: checkNativeWorkflowsInstalled },
   { name: 'tools-on-path',             run: checkToolsAvailable },
   { name: 'bizar-home',                run: checkBizarHome },
   { name: 'provider-reachable',        run: checkProviderReachable },

package/cli/provision.mjs

@@ -32,7 +32,12 @@ import { homedir } from 'node:os';
 import { dirname, join, resolve, sep } from 'node:path';
 import { fileURLToPath } from 'node:url';
 import { resolveBizarHome } from './config-paths.mjs';
-import { buildClaudeModelOverrides, configuredEnabledModels } from './commands/models.mjs';
+import {
+  buildClaudeModelOverrides,
+  configuredEnabledModels,
+  requiresGatewayModelDiscovery,
+} from './commands/models.mjs';
+import { validateNativeWorkflowDirectory } from '../config/workflows/lib/native-contract.mjs';
 
 const __filename = fileURLToPath(import.meta.url);
 const __dirname = dirname(__filename);
@@ -921,7 +926,6 @@ export function writeClaudeSettings({ dryRun = false, force = false } = {}) {
         pickEnv('CLAUDE_CODE_EXPERIMENTAL_AGENT_TEAMS')
         || shipped.env?.CLAUDE_CODE_EXPERIMENTAL_AGENT_TEAMS
         || '1',
-      ...(pickEnv('ANTHROPIC_MODEL') ? { ANTHROPIC_MODEL: pickEnv('ANTHROPIC_MODEL') } : {}),
     },
     hooks: {
       UserPromptSubmit: [{ hooks: [hook('user-prompt-submit', 10)] }],
@@ -968,6 +972,10 @@ export function writeClaudeSettings({ dryRun = false, force = false } = {}) {
   if (installModel) {
     bizarSettings.model = installModel;
     bizarSettings.modelOverrides = buildClaudeModelOverrides(installModels);
+    bizarSettings.env.ANTHROPIC_MODEL = installModel;
+    if (operatorGatewayUrl && requiresGatewayModelDiscovery(installModels)) {
+      bizarSettings.env.CLAUDE_CODE_ENABLE_GATEWAY_MODEL_DISCOVERY = '1';
+    }
     const configuredContext = pickEnv('CLAUDE_CODE_MAX_CONTEXT_TOKENS') || installContextTokens;
     if (configuredContext) bizarSettings.env.CLAUDE_CODE_MAX_CONTEXT_TOKENS = String(configuredContext);
   } else {
@@ -1013,6 +1021,9 @@ export function writeClaudeSettings({ dryRun = false, force = false } = {}) {
   if (installModel) {
     merged.model = installModel;
     merged.modelOverrides = bizarSettings.modelOverrides;
+    merged.env = { ...(merged.env || {}), ANTHROPIC_MODEL: installModel };
+  } else if (merged.env && typeof merged.env === 'object') {
+    delete merged.env.ANTHROPIC_MODEL;
   }
 
   // Auto-compaction is part of the Bizar reliability contract. Remove legacy
@@ -1348,9 +1359,9 @@ export async function syncConfigExtras({ dryRun = false } = {}) {
   const workflowsSrc = join(REPO_ROOT, 'config', 'workflows');
   if (existsSync(workflowsSrc)) {
     const workflowsDst = join(CLAUDE_DIR, 'workflows');
+    validateNativeWorkflowDirectory(workflowsSrc);
     await copyDirIfExists(workflowsSrc, workflowsDst);
-    counts.workflows = readdirSync(workflowsSrc, { withFileTypes: true })
-      .filter((entry) => entry.isFile() && entry.name.endsWith('.js')).length;
+    counts.workflows = validateNativeWorkflowDirectory(workflowsDst).count;
   }
 
   return { ok: true, message: `synced (${counts.commands} commands, ${counts.skills} skills, ${counts.hooks} hooks, ${counts.rules} rules, ${counts.workflows} workflows)`, counts };
@@ -1374,7 +1385,6 @@ export const FORCE_CLEAN_PRESERVE_ENV_KEYS = Object.freeze([
   'ANTHROPIC_AUTH_TOKEN',
   'BIZAR_MODEL_ROUTER_URL',
   'BIZAR_HOME',
-  'ANTHROPIC_MODEL',
   'CLAUDE_CODE_ENABLE_GATEWAY_MODEL_DISCOVERY',
   'CLAUDE_CODE_EXPERIMENTAL_AGENT_TEAMS',
   'CLAUDE_CODE_MAX_CONTEXT_TOKENS',

package/config/claude/CLAUDE.md

@@ -69,21 +69,18 @@ still deny prohibited actions and escalate externally visible or irreversible
 actions with `permissionDecision: "ask"`; that escalation list is the
 authoritative floor, not a starting point.
 
-Native dynamic workflows under `config/workflows/` and `~/.claude/workflows/`
-are the required dispatch mechanism for every primary request except an
-unmistakably tiny single-target copy/style/format edit with no behavior or test
-change. Mike may execute that narrow exception directly. All other work enters
-the matching research / implement / debug / review workflow and uses at least
-one explicitly modeled editing subagent with call-level worktree isolation.
-Do not add unnecessary phases or duplicate workers. For work that needs 3+
-long-lived workers with bounded cross-talk,
-Mike invokes a workflow that fans out as a native agent team; the team is
+Mike selects the coordination mode after bounded read-only orientation and one
+user clarification checkpoint: direct work only for an unmistakably tiny
+single-target copy/style/format edit; one isolated Agent for a clear bounded
+change; a native workflow for repeatable phased work; parallel Agents for
+disjoint scopes; and an Agent team for 3+ sustained roles that genuinely need
+cross-talk. Do not add unnecessary phases or duplicate workers. The team is
 host-side state under `CLAUDE_CODE_EXPERIMENTAL_AGENT_TEAMS=1`, and per
-Anthropic's docs `team_name` is deprecated and ignored. When two or more
-subtasks have non-overlapping writable scopes and no data dependency, the
-orchestrator MUST dispatch them concurrently through `parallel([...])`, each
-with call-level `isolation: "worktree"`. Sequential dispatch is reserved for
-dependent phases and integration; never create artificial parallel work.
+Anthropic's docs `team_name` is deprecated and ignored. Every editing dispatch
+uses an explicit configured model and call-level `isolation: "worktree"`.
+When two or more subtasks have non-overlapping writable scopes and no data
+dependency, dispatch them concurrently; serialize only dependencies and
+integration.
 
 Agent roles are model-agnostic. Mike selects the cheapest sufficient enabled
 configured-tier model for each dispatch, with explicit user picks taking
@@ -147,11 +144,13 @@ The autonomy and approval policy above governs this execution model. The project
 Every non-empty primary request enters Bizar through `office-manager` (`@mike`).
 The installer sets Claude Code's global `agent` setting to Mike's frontmatter
 name (`mike`),
-and a session-scoped routing guard requires a successful native Workflow before
-substantive primary-session mutation. Read-only inspection remains available.
-Mike directly executes only the tiny edit exception above. Every other request
-must invoke a native Bizar workflow before mutation; the workflow dispatches
-worktree-isolated subagents while Mike owns integration and final verification.
+and the routing hook supplies the adaptive coordination policy. For non-tiny
+work, Mike first gathers only bounded read-only context, asks one concise
+clarification question that names the inferred outcome and proposed mode, then
+continues autonomously after the answer. A native workflow is one available
+mode, not a universal gate; Mike may select an isolated Agent, parallel Agents,
+or an Agent team when that better fits the work. Mike owns integration and
+final verification.
 A Bizar custom agent already executing its assigned role does not recursively
 dispatch itself.
 
@@ -163,8 +162,8 @@ relevant page. Guess-and-try integration work is prohibited. When official
 documentation is unavailable or ambiguous, inspect authoritative source code
 and report the evidence gap.
 
-For workflow-routed requests, `office-manager` uses only the phases that reduce
-a known risk; only the tiny edit exception skips this pipeline:
+For a workflow or team, `office-manager` uses only phases and members that
+reduce a known risk:
 
 1. Research: `greg` (`research-analyst.md`) plus an implementation-context specialist.
 2. Plan: `planner` drafts; `qa-reviewer` challenges assumptions and test shape.

package/config/claude/agents/office-manager.md

@@ -1,7 +1,7 @@
 ---
 name: mike
-description: Mike — workflow-first orchestrator with a tiny direct-edit exception.
-tools: Workflow, Agent, Read, Write, Edit, Bash, Grep, Glob, WebFetch, WebSearch, Skill
+description: Mike — adaptive orchestrator that selects the lightest safe coordination mode.
+tools: Workflow, Agent, Read, Write, Edit, Bash, Grep, Glob, WebFetch, WebSearch, Skill, AskUserQuestion
 skills:
   - i-have-adhd
 ---
@@ -12,23 +12,45 @@ Follow `_shared/AGENT_BASELINE.md`. You own the user outcome, integration, and
 final verification. Direct execution is a narrow exception; workflows are the
 default for meaningful work.
 
-## Route, then reassess if scope expands
+## Orient, clarify, then select the coordination mode
 
 | Shape | Signals | Execution |
 |---|---|---|
 | Tiny direct | one obvious copy, typo, comment, whitespace, or single style-token edit; one target; no behavior or test change | inspect, make the micro-edit, run the smallest proving check yourself |
-| Bounded workflow | known non-trivial implementation, including a logical bug or any behavioral change | invoke `bizar-implement`; it dispatches at least one editing worker with `isolation: "worktree"`; merge and verify |
-| Debug workflow | failing behavior, unclear cause, regression, or interacting state | invoke `bizar-debug`; keep diagnosis and fix evidence separate |
-| Research/shaped | external/version-sensitive behavior, architecture/security, broad review, or interacting components | invoke `bizar-research` or the matching `ultracode*` workflow; parallelize independent lanes and serialize dependencies |
-
-If a request could reasonably require a regression test, touch multiple files,
-or needs inspection to discover its scope, it is not tiny: invoke a workflow
-before editing. The primary session does not substitute an ad-hoc Agent call
-for the workflow. Research current official docs only for external or
+| Single isolated worker | one bounded implementation after scope is clear | dispatch one worktree-isolated Agent with an explicit Bizar model; integrate and verify |
+| Native workflow | repeatable diagnosis, research, review, or an implementation needing visible phase barriers | invoke the matching Bizar workflow with explicit Bizar routing |
+| Agent team | three or more sustained, independent roles need bounded cross-talk or coordinated handoff | use the native Agent-team capability; writers use worktrees and explicit Bizar models |
+| Parallel agents | two disjoint writable scopes with no cross-talk needed | dispatch concurrently with explicit models and worktree isolation |
+
+For every non-tiny request, first make only enough read-only inspection to
+understand the repository boundary and current constraints. Then ask the user
+one concise clarification checkpoint: state the inferred outcome, the material
+choice or risk, and the proposed coordination mode. Wait for the answer before
+writing, dispatching editors, creating branches, or running tests. If the user
+explicitly says to proceed without questions, record that choice and continue.
+After the answer, work autonomously until the requested outcome and verification
+are complete. Research current official docs only for external or
 version-sensitive claims. Inspect installed skills before hard or specialized
 work; if stuck with no match, search skills.sh and review the candidate before
 proposing installation.
 
+Before every Workflow, Agent, or Agent-team call, read the global Bizar model router and construct a
+small `args.routing` object whose `default`, `medium`, and `high` values are
+explicit enabled configured model IDs (user picks win; otherwise use enabled
+tier candidates). Include the user's task in the same args object under the
+workflow's documented task field. Never pass only a string and never use
+`inherit`, `sonnet`, `opus`, or another provider default. If no configured ID
+exists, stop and ask the operator to run `bizar models`.
+
+Invoke the selected workflow by `name` first. If Claude reports that the Bizar
+name is unavailable, resolve the active Claude config directory and retry once
+with the absolute installed `scriptPath` at
+`<CLAUDE_CONFIG_DIR>/workflows/<name>.js` (normally
+`~/.claude/workflows/<name>.js`). Never retry a bare filename or a repository
+relative path. If that file is missing or invalid, stop with `bizar update`
+and `bizar doctor` as the repair commands; do not improvise a primary-session
+implementation around a broken workflow installation.
+
 ## Models
 
 For every Agent call, select the cheapest sufficient enabled configured model
@@ -41,9 +63,10 @@ only when the router explicitly supplies it.
 
 ## Worktree Discipline and integration
 
-Every editing subagent call uses call-level `isolation: "worktree"`. Parallel
-writers receive disjoint file ownership and sibling scopes. Read-only research
-stays foreground. When a writer finishes, merge its queued branch with
+Every editing subagent call uses call-level `isolation: "worktree"`. Use teams
+only when collaboration changes the result; do not manufacture a team or a
+workflow for a simple isolated task. Parallel writers receive disjoint file
+ownership and sibling scopes. Read-only research stays foreground. When a writer finishes, merge its queued branch with
 `bizar worktree-merge`; report conflicts instead of guessing. The integration
 branch runs final tests once after all required results are incorporated.
 Worktree branches use `wt/<agent_type>-<short-task-id>`.

package/config/claude/hooks/README.md

@@ -11,7 +11,7 @@ All hooks read Claude Code JSON from stdin and emit either no decision, addition
 | `simplify-guard.mjs` | PostToolUse Skill + PreToolUse Bash | require one `/simplify` per commit attempt |
 | `posttooluse-editwrite.mjs` | PostToolUse writes | local telemetry and test reminder |
 | `worker-suggest.mjs` | UserPromptSubmit | ranked skill/agent suggestions |
-| `workflow-route-guard.mjs` | UserPromptSubmit, PreToolUse, PostToolUse | requires a proven successful native workflow before substantive primary mutation; permits narrowly whitelisted, redirect-free Git inspection |
+| `workflow-route-guard.mjs` | UserPromptSubmit, PreToolUse, PostToolUse | records adaptive Bizar routing state without blocking Mike's selected coordination mode; workflow success clears the pending route record |
 | `thinking-route.mjs` | UserPromptSubmit | slash and mental-model routing |
 | `telemetry.mjs` | SessionStart/UserPromptSubmit | local correlation and rejection categories |
 | `sessionstart-prime.mjs` | SessionStart | bounded project and handoff context |

package/config/claude/hooks/agent-model-guard.mjs

@@ -36,11 +36,11 @@
  *     validated by `pickFailover` against the same registry.
  *   - The hard deny of out-of-pool models STILL APPLIES when
  *     `routingDecisionId` is absent — the contract is opt-in.
- *   - Registry load failures still fail open (F-176 advisory): the orchestrator
- *     already chose a model, let Claude Code validate it once.
+ *   - Registry load failures fail closed: allowing the dispatch would hand
+ *     model choice back to Claude Code's unconfigured provider default.
  *
- * The contract is intentionally additive. Existing callers that only pass
- * `model` see no behavior change.
+ * Callers that pass only `model` are accepted only when that literal ID is in
+ * the enabled configured pool. Missing or inherited model selection is denied.
  */
 
 import { readFileSync } from 'node:fs';
@@ -48,16 +48,6 @@ import { pathToFileURL } from 'node:url';
 
 import { loadModelRouter } from '../../../config/agents/model-assignment.mjs';
 
-function advise(reason) {
-  return {
-    hookSpecificOutput: {
-      hookEventName: 'PreToolUse',
-      permissionDecision: 'allow',
-      additionalContext: `🟡 Model override guidance: ${reason} The dispatch will proceed regardless.`,
-    },
-  };
-}
-
 function deny(reason) {
   return {
     hookSpecificOutput: {
@@ -72,8 +62,6 @@ function deny(reason) {
  * 10.22.0 / Phase 4: extract the operator's `disabledProviders` list from
  * the loaded registry. Whitespace-trimmed + lowercased at read time.
  * Returns `[]` for legacy configs that lack the key (no in-code default).
- * The hook MUST fail open on parse errors — the orchestrator already
- * chose a model; let Claude Code validate it once.
  */
 function readDisabledProvidersFromRegistry(registry) {
   if (!registry || typeof registry !== 'object') return [];
@@ -90,14 +78,15 @@ function readDisabledProvidersFromRegistry(registry) {
 }
 
 /**
- * 10.22.0 / Phase 4: case-sensitive prefix filter against the (lowercase)
- * disabled list. Empty / missing prefix list is a no-op (returns input).
+ * 10.22.0 / Phase 4: case-insensitive prefix filter against the normalized
+ * disabled list. Empty / missing prefix list is a no-op.
  */
 function isDisabledId(id, prefixes) {
   if (!Array.isArray(prefixes) || prefixes.length === 0) return false;
   if (typeof id !== 'string' || !id) return false;
+  const normalized = id.toLowerCase();
   for (const p of prefixes) {
-    if (typeof p === 'string' && p && id.startsWith(p)) return true;
+    if (typeof p === 'string' && p && normalized.startsWith(p)) return true;
   }
   return false;
 }
@@ -171,18 +160,15 @@ export async function guardAgentModel(input, options = {}) {
   const failoverBlock = readFailoverBlock(toolInput);
   const hasFailoverContract = Boolean(failoverBlock.routingDecisionId) && Boolean(failoverBlock.fallback);
 
-  // The workflow dispatcher enforces a model override. This permissive branch
-  // keeps the hook compatible with non-Bizar Agent callers.
-  if (!requested) return {};
-  if (requested === 'inherit') return {};
+  if (!requested || requested === 'inherit') {
+    return deny('Bizar Agent dispatch blocked: every Agent call requires an explicit enabled model from `bizar models`; session/provider inheritance is prohibited.');
+  }
 
   let registry;
   try {
     registry = options.registry || loadModelRouter(options.routerPath);
   } catch {
-    // A broken optional router must not strand subagents. The caller already
-    // chose a model; let Claude Code/provider validate it once.
-    return {};
+    return deny('Bizar Agent dispatch blocked: the global model router is missing or invalid. Run `bizar models`, then retry.');
   }
 
   // Disabled is an enforceable operator boundary. Falling through here would
@@ -201,17 +187,17 @@ export async function guardAgentModel(input, options = {}) {
   // already computed by `pickFailover` against this same registry.
   if (hasFailoverContract) {
     if (!userPicks.has(failoverBlock.fallback)) {
-      return advise(`Bizar Agent dispatch: fallback ${failoverBlock.fallback} is outside the user-selected pool; select an enabled configured fallback.`);
+      return deny(`Bizar Agent dispatch blocked: fallback ${failoverBlock.fallback} is outside the user-selected pool; select an enabled configured fallback.`);
     }
     if (!allowed.has(requested)) {
-      return advise(`Bizar Agent dispatch blocked: model override ${requested} is outside the configured dynamic tiers and the user-selected pool. Pick it via \`bizar models\`.`);
+      return deny(`Bizar Agent dispatch blocked: model override ${requested} is outside the configured dynamic tiers and the user-selected pool. Pick it via \`bizar models\`.`);
     }
     // Both IDs are user-selected. Accept without re-probing the gateway.
     return {};
   }
 
   if (!allowed.has(requested)) {
-    return advise(`Bizar Agent dispatch blocked: model override ${requested} is outside the configured dynamic tiers and the user-selected pool. Pick it via \`bizar models\`.`);
+    return deny(`Bizar Agent dispatch blocked: model override ${requested} is outside the configured dynamic tiers and the user-selected pool. Pick it via \`bizar models\`.`);
   }
 
   // User-selected models bypass live-discovery validation. The picker is the
@@ -220,7 +206,7 @@ export async function guardAgentModel(input, options = {}) {
   if (!fromUserPick && Array.isArray(options.availableModelIds)) {
     const available = new Set(options.availableModelIds);
     if (!available.has(requested)) {
-      return advise(`Bizar Agent dispatch blocked: ${requested} was not reported by live discovery. Select another enabled configured model; do not retry aliases.`);
+      return deny(`Bizar Agent dispatch blocked: ${requested} was not reported by live discovery. Select another enabled configured model; do not retry aliases.`);
     }
   }
 

package/config/claude/hooks/sessionstart-model-sync.mjs

@@ -122,6 +122,10 @@ function filterDisabled(ids, disabled) {
   return kept;
 }
 
+function requiresGatewayModelDiscovery(modelIds) {
+  return modelIds.some((id) => !/^(?:claude(?:-|$)|anthropic(?:[./-]|$))/i.test(id));
+}
+
 function readSettingsPath() {
   return join(resolveClaudeConfigDir(), 'settings.json');
 }
@@ -234,8 +238,14 @@ function syncOnce() {
   ];
   settings.modelPicker = { options };
   settings.modelOverrides = Object.fromEntries(
-    liveIds.slice(0, overrideKeys.length).map((id, index) => [overrideKeys[index], id]),
+    overrideKeys.map((key, index) => [key, liveIds[index % liveIds.length]]),
   );
+  if (requiresGatewayModelDiscovery(liveIds)) {
+    settings.env = {
+      ...(settings.env || {}),
+      CLAUDE_CODE_ENABLE_GATEWAY_MODEL_DISCOVERY: '1',
+    };
+  }
 
   let modelChanged = false;
   if (typeof settings.model !== 'string' || !liveIds.includes(settings.model)) {

package/config/claude/hooks/sessionstart-prime.mjs

@@ -186,13 +186,13 @@ function startupBriefing(cwd, featureBrief, recentCommits, projectLine, progress
     }
   }
   if (progressLast) lines.push(`- Progress: ${progressLast}.`);
-  lines.push('- You are @mike. Only an unmistakably tiny single-target copy/style/format edit is direct; every other change enters the matching native workflow before mutation. Use one isolated writer by default and parallel worktrees only for independent scopes.');
+  lines.push('- You are @mike. For non-tiny work, do bounded read-only orientation, ask one clarification checkpoint, then choose a single isolated worker, native workflow, parallel workers, or an Agent team by actual dependency. Use worktrees for editors and explicit Bizar models for every Agent.');
   lines.push('- External/version-sensitive work requires current official docs via WebSearch/WebFetch. Use relevant installed skills; apply i-have-adhd to user output. WIP=1.');
   lines.push('- TaskCompleted/SubagentStop/<task-notification> is terminal: consume <result>, mark done/failed, merge queued work, continue the objective.');
   // Default-first-stop hint when nothing is active yet.
   if (featureBrief && featureBrief.active.length === 0) {
     lines.push(
-      '- First move: read PROGRESS.md and feature_list.json; then choose the matching workflow, except for an unmistakably tiny direct edit.',
+      '- First move: read PROGRESS.md and feature_list.json, then do bounded orientation and ask the clarification checkpoint before choosing the fitting coordination mode.',
     );
   }
   return lines.join('\n');
@@ -203,14 +203,14 @@ function clearBriefing(cwd, recentCommits, progressLast) {
   if (progressLast) lines.push(`- Progress: ${progressLast}.`);
   if (recentCommits.length > 0) lines.push(`- Last commit: ${recentCommits[0]}.`);
   lines.push('- Context preserved in same repo / cwd — only the model turn was reset.');
-  lines.push('- You are @mike: continue through the active workflow; only an unmistakably tiny edit may stay direct.');
+  lines.push('- You are @mike: continue the active coordination plan; adapt it when new evidence changes the fit.');
   lines.push('- First move: continue from where the model left off; no need to reread project files.');
   return lines.join('\n');
 }
 
 function resumeBriefing(cwd, state) {
   const lines = ['Bizar SessionStart (resume):'];
-  lines.push('- You are @mike: restore state, then continue the active workflow; only an unmistakably tiny edit may stay direct.');
+  lines.push('- You are @mike: restore state, then continue the active coordination plan and adapt it if the evidence changed.');
   if (state) {
     if (state.activeFeature) lines.push(`- Last active feature: ${state.activeFeature}.`);
     if (state.reason) lines.push(`- Last session ended with: ${state.reason}.`);

package/config/claude/hooks/worker-suggest.mjs

@@ -6,7 +6,8 @@
  *
  * Runs on every user prompt. Only unmistakably tiny, single-scope edits take
  * a cheap fast path. Every other request is routed into a native Bizar
- * workflow that owns subagent dispatch.
+ * adaptive coordination mode selected by Mike after bounded orientation and a
+ * clarification checkpoint.
  *
  * Uses import.meta.url + dynamic import() to resolve the sibling CLI module so
  * the hook works regardless of install path (fixes ERR_MODULE_NOT_FOUND after
@@ -56,10 +57,10 @@ const FAST_ROUTE_POLICY = [
 ].join('\n');
 
 const ROUTE_POLICY = [
-  'Workflow-required Bizar routing policy:',
-  '- If this is the primary session, you ARE @mike. Do not implement this request directly in the primary session. Before any edit or mutation, invoke the matching native Bizar workflow; the primary owns routing, integration, and final verification.',
-  '- Use bizar-implement for known bounded changes, bizar-debug for bugs needing diagnosis, bizar-research for external or uncertain implementation context, and ultracode / ultracode-research / ultracode-review for broad, high-risk, or review-heavy objectives. Use only phases that reduce a concrete risk.',
-  '- The workflow must dispatch at least one implementation subagent with an explicit configured model and call-level `isolation: "worktree"`. For genuinely disjoint writable scopes, dispatch them concurrently; otherwise use one isolated writer. Never create duplicate workers merely to satisfy fan-out.',
+  'Adaptive Bizar routing policy:',
+  '- If this is the primary session, you ARE @mike. First do only bounded read-only orientation. Then ask one concise clarification question describing the inferred outcome, the material choice/risk, and your proposed coordination mode. Wait for the answer before edits, branches, tests, or editor dispatch. If the user explicitly waives questions, continue autonomously.',
+  '- After clarification, choose the lightest coordination mode: a direct tiny edit, one isolated Agent for a bounded change, a native Bizar Workflow for repeatable phased work, parallel Agents for disjoint scopes, or an Agent team only when 3+ sustained roles need cross-talk. Do not force a workflow or team when it adds no value.',
+  '- Every Agent or team member receives an explicit enabled configured Bizar model; use no provider alias, inheritance, or default. Every editing worker uses call-level `isolation: "worktree"`. For genuinely disjoint writable scopes, dispatch concurrently; otherwise use one owner.',
   '- Consume terminal agent results, merge queued worktrees with bizar worktree-merge, and run integration checks in the primary session. A subagent may not recursively dispatch itself.',
   '- Do NOT execute any tool you do not have. If a tool you need is missing from your tools list, dispatch to a subagent that has it — do not pretend you have it.',
   '- If you are already running as a Bizar custom agent, follow your assigned role and do not recursively dispatch yourself.',