fix(agent-core-v2): align ModelResolver with production v1 behavior

- return auth as { apiKey } (kosong reads apiKey, not headers) - apply default thinking from defaultThinking/thinking config + always_thinking clamp - treat empty apiKey as absent so OAuth is not short-circuited - strip trailing /v1 from anthropic baseUrl to avoid double /v1/v1/messages - add model-providers example (sandboxed ~/.kimi-code ping-pong) and resolver tests
2026-08-19 13:45:28 +00:00 · 2026-07-02 18:42:13 +08:00 · 2026-07-02 18:42:13 +08:00 · 643acf454a
commit 643acf454a
parent ec865ed89c
3 changed files with 670 additions and 5 deletions
--- a/packages/agent-core-v2/examples/model-providers.example.ts
+++ b/packages/agent-core-v2/examples/model-providers.example.ts
@ -0,0 +1,373 @@
+/**
+ * Scenario: the **Provider / Platform / Protocol / Model** slice, driven from
+ * a real `~/.kimi-code/config.toml` and its credentials, and exercised through
+ * the new `IModelResolver` → `Model` god-object path introduced in the
+ * "Model god-object and protocol domains" change.
+ *
+ * Goals of this example:
+ *  1. **Sandbox the real config.** At runtime, copy `~/.kimi-code/config.toml`
+ *     and `~/.kimi-code/credentials/` into the per-run `KIMI_CODE_HOME` the
+ *     example harness provisions (`.vitest-results/kimi-code-{ts}/`). The real
+ *     home is never read or written directly — even an OAuth token refresh
+ *     lands in the sandbox copy.
+ *  2. **List everything.** Enumerate every `[providers.*]`, `[platforms.*]`,
+ *     supported `Protocol`, and `[models.*]` entry, then resolve each Model id
+ *     through `IModelResolver` and report whether it produces a runnable
+ *     `Model` (protocol, base URL, auth mode) — a concrete compatibility
+ *     matrix for the new god-object resolver.
+ *  3. **Ping every Model.** Send a "ping" → expect a streamed response
+ *     against **every** Model that resolved (bounded concurrency, per-request
+ *     timeout), and report which ones actually answer — an end-to-end reachability
+ *     check for the whole configured catalogue, not just the default model.
+ *
+ * All Services come from `src/`; nothing here defines a new Service.
+ */
+
+import { copyFileSync, existsSync, mkdirSync, readdirSync, statSync } from 'node:fs';
+import { homedir } from 'node:os';
+import { join } from 'node:path';
+
+import { afterAll, beforeAll, describe, expect, test } from 'vitest';
+
+import '#/index';
+
+import { type Scope, type ScopeSeed } from '#/_base/di/scope';
+import { bootstrap } from '#/app/bootstrap';
+import { IConfigService } from '#/app/config';
+import { createUserMessage, isContentPart, type TokenUsage } from '#/app/llmProtocol';
+import { IModelResolver, IModelService, type Model, type ModelConfig } from '#/app/model';
+import { IPlatformService } from '#/app/platform';
+import { IProviderService, type ProviderConfig } from '#/app/provider';
+import { IProtocolAdapterRegistry } from '#/app/protocol';
+import { ILogOptions, resolveLoggingConfig } from '#/app/log/logConfig';
+
+const PER_REQUEST_TIMEOUT_MS = 30_000;
+const PING_CONCURRENCY = 4;
+const ALL_MODELS_TEST_TIMEOUT_MS = 300_000;
+
+interface ModelReport {
+  readonly id: string;
+  readonly name?: string;
+  readonly protocol?: string;
+  readonly baseUrl?: string;
+  readonly authMode: string;
+  readonly resolved: boolean;
+  readonly error?: string;
+}
+
+describe('model / provider / platform / protocol slice (resolved from a sandboxed ~/.kimi-code)', () => {
+  let app: Scope | undefined;
+  let sandboxHome = '';
+  let configCopied = false;
+  let credentialsCopied = 0;
+  let reports: readonly ModelReport[] = [];
+  let useRealHome = false;
+
+  beforeAll(() => {
+    useRealHome = process.env['KIMI_CODE_EXAMPLE_USE_REAL_HOME'] === '1';
+    sandboxHome = useRealHome ? join(homedir(), '.kimi-code') : resolveSandboxHome();
+    if (useRealHome) {
+      // Run directly against the real home so OAuth token refresh can read AND
+      // write back the real credentials. Read-only for config — this example
+      // never calls IConfigService.set/replace.
+      configCopied = true;
+      credentialsCopied = 0;
+    } else {
+      const mirror = mirrorRealKimiHome(sandboxHome);
+      configCopied = mirror.configCopied;
+      credentialsCopied = mirror.credentialsCopied;
+    }
+
+    const logSeed: ScopeSeed = [
+      [ILogOptions, resolveLoggingConfig({ homeDir: sandboxHome, env: process.env })],
+    ];
+    app = bootstrap({ homeDir: sandboxHome }, logSeed).app;
+  });
+
+  afterAll(() => app?.dispose());
+
+  test('lists every Provider / Platform / Protocol and resolves every Model', async () => {
+    const host = requireApp(app);
+    const config = host.accessor.get(IConfigService);
+    await config.ready;
+
+    const providers = host.accessor.get(IProviderService);
+    const platforms = host.accessor.get(IPlatformService);
+    const models = host.accessor.get(IModelService);
+    const resolver = host.accessor.get(IModelResolver);
+    const protocols = host.accessor.get(IProtocolAdapterRegistry);
+
+    // Touch each registry so its config section is registered before we read.
+    const providerMap = providers.list();
+    const platformMap = platforms.list();
+    const modelMap = models.list();
+    const supportedProtocols = protocols.supportedProtocols();
+
+    console.log(`\nhome:                ${sandboxHome}${useRealHome ? ' (REAL ~/.kimi-code)' : ' (sandbox copy)'}`);
+    console.log(`config.toml copied:  ${useRealHome ? 'n/a (using real)' : configCopied}`);
+    console.log(`credentials copied:  ${useRealHome ? 'n/a (using real)' : credentialsCopied}`);
+    console.log(`\nsupported protocols: ${supportedProtocols.join(', ') || '(none)'}`);
+
+    console.log(`\n[providers.*] (${Object.keys(providerMap).length}):`);
+    for (const [id, p] of Object.entries(providerMap)) {
+      console.log(`  - ${id}: type=${p.type ?? '-'} baseUrl=${p.baseUrl ?? '-'} auth=${providerAuthMode(p)} platform=${p.platformId ?? '-'}`);
+    }
+
+    console.log(`\n[platforms.*] (${Object.keys(platformMap).length}):`);
+    if (Object.keys(platformMap).length === 0) console.log('  (none configured)');
+    for (const [id, pl] of Object.entries(platformMap)) {
+      const auth = pl.auth?.apiKey !== undefined ? 'apiKey' : pl.auth?.oauth !== undefined ? 'oauth' : pl.auth?.env !== undefined ? 'env' : '-';
+      console.log(`  - ${id}: auth=${auth} displayName=${pl.displayName ?? '-'}`);
+    }
+
+    reports = Object.entries(modelMap).map(([id, m]) => resolveOne(id, m, providerMap, resolver));
+
+    console.log(`\n[models.*] (${reports.length}) — resolve compatibility:`);
+    for (const r of reports) {
+      const head = r.resolved ? 'OK ' : 'FAIL';
+      const detail = r.resolved
+        ? `protocol=${r.protocol} baseUrl=${r.baseUrl} auth=${r.authMode} name=${r.name}`
+        : `auth=${r.authMode} error=${r.error}`;
+      console.log(`  [${head}] ${r.id} → ${detail}`);
+    }
+
+    // The example is meaningful even on a machine without the real config: it
+    // simply reports an empty registry instead of failing.
+    if (!configCopied) {
+      console.log('\n(no ~/.kimi-code/config.toml found — reporting an empty registry)');
+      return;
+    }
+
+    expect(Object.keys(providerMap).length).toBeGreaterThan(0);
+    expect(reports.length).toBeGreaterThan(0);
+    // Every configured Model must at least resolve into a god-object; a
+    // resolution failure here is a real compatibility regression.
+    const failures = reports.filter((r) => !r.resolved);
+    expect(
+      failures,
+      `models that failed to resolve: ${failures.map((f) => `${f.id}(${f.error})`).join(', ')}`,
+    ).toEqual([]);
+  });
+
+  test('sends a ping → pong request through EVERY resolvable Model', async () => {
+    const host = requireApp(app);
+    if (!configCopied || reports.length === 0) {
+      console.log('skipped: no ~/.kimi-code/config.toml or no models configured');
+      return;
+    }
+
+    const resolver = host.accessor.get(IModelResolver);
+    const candidates = reports.filter((r) => r.resolved);
+    if (candidates.length === 0) {
+      console.log('skipped: no resolvable models');
+      return;
+    }
+
+    console.log(
+      `\npinging ${candidates.length} resolvable models ` +
+        `(concurrency=${PING_CONCURRENCY}, per-request timeout=${PER_REQUEST_TIMEOUT_MS}ms):`,
+    );
+
+    const outcomes = await mapPool(candidates, PING_CONCURRENCY, async (report) => {
+      const model = resolver.resolve(report.id);
+      const controller = new AbortController();
+      const timer = setTimeout(() => controller.abort(), PER_REQUEST_TIMEOUT_MS);
+      try {
+        const result = await collectResponse(model, controller.signal);
+        return {
+          report,
+          ok: true as const,
+          text: result.text,
+          finishReason: result.finishReason,
+        };
+      } catch (error) {
+        return {
+          report,
+          ok: false as const,
+          error: error instanceof Error ? error.message : String(error),
+        };
+      } finally {
+        clearTimeout(timer);
+      }
+    });
+
+    for (const o of outcomes) {
+      if (o.ok) {
+        console.log(
+          `  [OK  ] ${o.report.id} → ${JSON.stringify(truncate(o.text, 40))} ` +
+            `(finish=${o.finishReason ?? '-'})`,
+        );
+      } else {
+        console.log(`  [FAIL] ${o.report.id} → ${truncate(o.error, 140)}`);
+      }
+    }
+
+    const passed = outcomes.filter((o) => o.ok).length;
+    console.log(`\nping-pong summary: ${passed}/${outcomes.length} models responded.`);
+
+    const failed = outcomes.filter((o) => !o.ok);
+    expect(
+      failed,
+      `models that failed to respond: ${failed.map((f) => `${f.report.id}(${f.error})`).join(', ')}`,
+    ).toEqual([]);
+  }, ALL_MODELS_TEST_TIMEOUT_MS);
+});
+
+function resolveOne(
+  id: string,
+  model: ModelConfig,
+  providers: Readonly<Record<string, ProviderConfig>>,
+  resolver: IModelResolver,
+): ModelReport {
+  const authMode = modelAuthMode(model, providers);
+  try {
+    const resolved = resolver.resolve(id);
+    return {
+      id,
+      name: resolved.name,
+      protocol: resolved.protocol,
+      baseUrl: resolved.baseUrl,
+      authMode,
+      resolved: true,
+    };
+  } catch (error) {
+    return {
+      id,
+      name: model.name ?? model.model,
+      authMode,
+      resolved: false,
+      error: error instanceof Error ? error.message : String(error),
+    };
+  }
+}
+
+/** Mirror the resolver's auth-precedence to label where each Model's
+ *  credential comes from, without ever reading the secret itself. */
+function modelAuthMode(
+  model: ModelConfig,
+  providers: Readonly<Record<string, ProviderConfig>>,
+): string {
+  if (model.apiKey !== undefined && model.apiKey.length > 0) return 'model.apiKey';
+  if (model.oauth !== undefined) return 'model.oauth';
+  const providerId = model.providerId ?? model.provider;
+  const provider = providerId === undefined ? undefined : providers[providerId];
+  const platformId = provider?.platformId;
+  if (platformId !== undefined && platformId !== '__unknown__') {
+    return `platform(${platformId})`;
+  }
+  if (provider?.apiKey !== undefined && provider.apiKey.length > 0) return 'provider.apiKey';
+  if (provider?.oauth !== undefined) return 'provider.oauth';
+  return 'none';
+}
+
+function providerAuthMode(provider: ProviderConfig): string {
+  if (provider.apiKey !== undefined && provider.apiKey.length > 0) return 'apiKey';
+  if (provider.oauth !== undefined) return 'oauth';
+  if (provider.platformId !== undefined) return `platform(${provider.platformId})`;
+  if (provider.env !== undefined) return 'env';
+  return 'none';
+}
+
+async function collectResponse(
+  model: Model,
+  signal: AbortSignal,
+): Promise<{ text: string; finishReason?: string; usage?: TokenUsage }> {
+  let text = '';
+  let think = '';
+  let finishReason: string | undefined;
+  let usage: TokenUsage | undefined;
+
+  const stream = model.request(
+    {
+      systemPrompt:
+        'You are a connectivity check. The user will say "ping". Reply with the single word: pong',
+      tools: [],
+      messages: [createUserMessage('ping')],
+    },
+    signal,
+  );
+
+  for await (const event of stream) {
+    if (event.type === 'part') {
+      const part = event.part;
+      if (isContentPart(part) && part.type === 'text') text += part.text;
+      else if (isContentPart(part) && part.type === 'think') think += part.think;
+    } else if (event.type === 'usage') {
+      usage = event.usage;
+    } else if (event.type === 'finish') {
+      finishReason = event.rawFinishReason ?? event.providerFinishReason;
+    }
+  }
+  // Thinking models may put the answer in `think`; surface whichever carried
+  // content so the report shows what came back.
+  return { text: text.trim().length > 0 ? text : think, finishReason, usage };
+}
+
+/** Run `fn` over `items` with at most `size` in flight, preserving order. */
+async function mapPool<T, R>(
+  items: readonly T[],
+  size: number,
+  fn: (item: T) => Promise<R>,
+): Promise<R[]> {
+  const results: R[] = new Array(items.length);
+  let next = 0;
+  const worker = async (): Promise<void> => {
+    while (next < items.length) {
+      const i = next++;
+      results[i] = await fn(items[i] as T);
+    }
+  };
+  await Promise.all(Array.from({ length: Math.min(size, items.length) }, worker));
+  return results;
+}
+
+function truncate(s: string, max: number): string {
+  const oneLine = s.replaceAll(/\s+/g, ' ').trim();
+  return oneLine.length > max ? `${oneLine.slice(0, max)}…` : oneLine;
+}
+
+function resolveSandboxHome(): string {
+  const fromEnv = process.env['KIMI_CODE_HOME'];
+  if (fromEnv !== undefined && fromEnv.length > 0) return fromEnv;
+  // Fallback for running this file outside the example harness: mirror into a
+  // fresh temp dir so the real home is still never touched.
+  const dir = join(homedir(), '.kimi-code-example-sandbox');
+  mkdirSync(dir, { recursive: true });
+  process.env['KIMI_CODE_HOME'] = dir;
+  return dir;
+}
+
+/** Copy `~/.kimi-code/config.toml` and `~/.kimi-code/credentials/*` into the
+ *  sandbox home. Never reads credential contents — only copies bytes. */
+function mirrorRealKimiHome(sandboxHome: string): {
+  configCopied: boolean;
+  credentialsCopied: number;
+} {
+  const realHome = join(homedir(), '.kimi-code');
+  let configCopied = false;
+  const srcConfig = join(realHome, 'config.toml');
+  if (existsSync(srcConfig)) {
+    copyFileSync(srcConfig, join(sandboxHome, 'config.toml'));
+    configCopied = true;
+  }
+
+  let credentialsCopied = 0;
+  const srcCreds = join(realHome, 'credentials');
+  if (existsSync(srcCreds) && statSync(srcCreds).isDirectory()) {
+    const dstCreds = join(sandboxHome, 'credentials');
+    mkdirSync(dstCreds, { recursive: true });
+    for (const entry of readdirSync(srcCreds)) {
+      const src = join(srcCreds, entry);
+      if (statSync(src).isFile()) {
+        copyFileSync(src, join(dstCreds, entry));
+        credentialsCopied++;
+      }
+    }
+  }
+  return { configCopied, credentialsCopied };
+}
+
+function requireApp(app: Scope | undefined): Scope {
+  if (app === undefined) throw new Error('App scope was not initialized in beforeAll');
+  return app;
+}
--- a/packages/agent-core-v2/src/app/model/modelResolverService.ts
+++ b/packages/agent-core-v2/src/app/model/modelResolverService.ts
@ -25,6 +25,7 @@ import {
  UNKNOWN_CAPABILITY,
  type ModelCapability,
  type ProviderRequestAuth,
+  type ThinkingEffort,
 } from '#/app/llmProtocol';
 import { IPlatformService, UNKNOWN_PLATFORM_KEY } from '#/app/platform';
 import type { OAuthRef, ProviderConfig } from '#/app/provider';
@ -38,6 +39,23 @@ import type { AuthProvider, Model } from './modelInstance';
 import { IModelResolver } from './modelResolver';
 import { ModelImpl, StaticAuthProvider } from './modelImpl';

+/**
+ * Default thinking effort applied when the user has not disabled thinking
+ * (matches `profile`'s `DEFAULT_THINKING_EFFORT`). Read here rather than
+ * imported so `model` (L2) does not depend on `profile` (L4); the source of
+ * truth for the value is the `thinking` / `defaultThinking` config sections,
+ * which are shared via `IConfigService`.
+ */
+const DEFAULT_THINKING_EFFORT: ThinkingEffort = 'high';
+const THINKING_EFFORTS: readonly ThinkingEffort[] = ['low', 'medium', 'high', 'xhigh', 'max'];
+
+/** Shape of the `thinking` config section (owned by `profile`); only the
+ *  fields the resolver needs to mirror the production default are read here. */
+interface ThinkingSection {
+  readonly mode?: string;
+  readonly effort?: string;
+}
+
 interface ResolvedAuthMaterial {
  readonly apiKey?: string;
  readonly oauth?: OAuthRef;
@ -68,11 +86,16 @@ export class ModelResolverService extends Disposable implements IModelResolver {
      );
    }

-    const { providerConfig, providerName, resolvedBaseUrl } = this.resolveProviderContext(id, model);
+    const { providerConfig, providerName, resolvedBaseUrl: rawBaseUrl } = this.resolveProviderContext(id, model);
    const auth = this.resolveAuth(model, providerConfig);
    const authProvider = this.buildAuthProvider(providerName, auth);

    const protocol = this.resolveProtocol(id, model, providerConfig);
+    // The Anthropic SDK appends `/v1/messages` to the baseUrl, so a provider
+    // whose baseUrl already ends in `/v1` (e.g. the managed Kimi endpoint) would
+    // otherwise produce a double `/v1/v1/messages` → 404. Match production v1
+    // (`provider-manager` strips a trailing `/v1` for the anthropic transport).
+    const resolvedBaseUrl = protocol === 'anthropic' ? stripTrailingV1(rawBaseUrl) : rawBaseUrl;
    const wireName = model.name ?? model.model;
    if (wireName === undefined) {
      throw new KimiError(
@ -97,7 +120,7 @@ export class ModelResolverService extends Disposable implements IModelResolver {
    };
    const alwaysThinking = declared.has('always_thinking');

-    return new ModelImpl({
+    const impl = new ModelImpl({
      id,
      name: wireName,
      aliases: model.aliases ?? [],
@ -113,7 +136,35 @@ export class ModelResolverService extends Disposable implements IModelResolver {
      providerName,
      authProvider,
      protocolRegistry: this.protocolRegistry as ProtocolAdapterRegistry,
+      extras: buildProviderExtras(model),
    });
+
+    // Apply the production default thinking effort so a plain `model.request()`
+    // behaves like the agent path (which routes through `profile` and reads the
+    // same `thinking` / `defaultThinking` config). Required for models whose
+    // endpoint rejects a request that omits thinking (e.g. kimi-k2.7 over the
+    // Anthropic protocol returns 400 unless `thinking.type === 'enabled'`).
+    const effort = this.resolveDefaultThinking(alwaysThinking);
+    return effort === 'off' ? impl : impl.withThinking(effort);
+  }
+
+  /**
+   * Mirror `profile`'s `resolveThinkingLevel` / `resolveThinkingEffort` so the
+   * god-object's default matches the production agent path:
+   *   - an explicit `defaultThinking === false` or `thinking.mode === 'off'`
+   *     turns thinking off;
+   *   - otherwise the configured `thinking.effort` (default 'high') is used;
+   *   - an `always_thinking` model clamps an explicit "off" back to on.
+   */
+  private resolveDefaultThinking(alwaysThinking: boolean): ThinkingEffort {
+    const defaultThinking = this.config.get<boolean | undefined>('defaultThinking');
+    const thinking = this.config.get<ThinkingSection | undefined>('thinking');
+    const turnedOff = defaultThinking === false || thinking?.mode === 'off';
+    const configured = parseThinkingEffort(thinking?.effort) ?? DEFAULT_THINKING_EFFORT;
+    if (turnedOff && !alwaysThinking) {
+      return 'off';
+    }
+    return configured;
  }

  findByName(name: string): readonly string[] {
@ -198,12 +249,18 @@ export class ModelResolverService extends Disposable implements IModelResolver {
   *   1. Model-inline `apiKey` / `oauth` (flat-case override).
   *   2. Provider.platformId → Platform.auth (structured shared auth).
   *   3. Provider-legacy `apiKey` / `oauth` (pre-migration configs).
+   *
+   * An empty / whitespace `apiKey` is treated as absent (matching production's
+   * `nonEmptyString`), so a provider that carries both `api_key = ""` and an
+   * `oauth` block correctly falls through to OAuth instead of producing an
+   * empty bearer token.
   */
  private resolveAuth(
    model: ModelConfig,
    provider: ProviderConfig | undefined,
  ): ResolvedAuthMaterial {
-    if (model.apiKey !== undefined) return { apiKey: model.apiKey };
+    const modelApiKey = nonEmpty(model.apiKey);
+    if (modelApiKey !== undefined) return { apiKey: modelApiKey };
    if (model.oauth !== undefined) {
      return { oauth: model.oauth, oauthProviderKey: model.providerId ?? model.provider };
    }
@ -211,7 +268,8 @@ export class ModelResolverService extends Disposable implements IModelResolver {
    const platformId = provider?.platformId;
    if (platformId !== undefined && platformId !== UNKNOWN_PLATFORM_KEY) {
      const platform = this.platforms.get(platformId);
-      if (platform?.auth?.apiKey !== undefined) return { apiKey: platform.auth.apiKey };
+      const platformApiKey = nonEmpty(platform?.auth?.apiKey);
+      if (platformApiKey !== undefined) return { apiKey: platformApiKey };
      if (platform?.auth?.oauth !== undefined) {
        return {
          oauth: platform.auth.oauth,
@ -221,7 +279,8 @@ export class ModelResolverService extends Disposable implements IModelResolver {
    }

    // Legacy: provider carried auth directly (pre-Phase 4 migration).
-    if (provider?.apiKey !== undefined) return { apiKey: provider.apiKey };
+    const providerApiKey = nonEmpty(provider?.apiKey);
+    if (providerApiKey !== undefined) return { apiKey: providerApiKey };
    if (provider?.oauth !== undefined) {
      return { oauth: provider.oauth, oauthProviderKey: model.providerId ?? model.provider };
    }
@ -250,6 +309,43 @@ export class ModelResolverService extends Disposable implements IModelResolver {
  }
 }

+function parseThinkingEffort(value: string | undefined): ThinkingEffort | undefined {
+  const normalized = value?.trim().toLowerCase();
+  return normalized !== undefined && (THINKING_EFFORTS as readonly string[]).includes(normalized)
+    ? (normalized as ThinkingEffort)
+    : undefined;
+}
+
+/** Treat an empty / whitespace string as absent (matches production's
+ *  `nonEmptyString` used by the session resolver). */
+function nonEmpty(value: string | undefined): string | undefined {
+  const trimmed = value?.trim();
+  return trimmed === undefined || trimmed.length === 0 ? undefined : trimmed;
+}
+
+/** Strip a trailing `/v1` (with optional trailing slash) from a baseUrl, matching
+ *  production v1's anthropic-transport normalization so the Anthropic SDK's
+ *  `/v1/messages` suffix does not produce a double `/v1/v1/messages`. */
+function stripTrailingV1(baseUrl: string): string {
+  return baseUrl.replace(/\/v1\/?$/, '');
+}
+
+/** Provider knobs the wire adapter needs that aren't first-class ModelImpl
+ *  fields. `adaptiveThinking` changes how the Anthropic adapter encodes the
+ *  thinking param, so it must reach the provider for the default-thinking
+ *  transform to produce the right shape on adaptive models. */
+function buildProviderExtras(model: ModelConfig): Readonly<Record<string, unknown>> | undefined {
+  const extras: Record<string, unknown> = {};
+  if (model.adaptiveThinking !== undefined) {
+    extras['adaptiveThinking'] = model.adaptiveThinking;
+  }
+  const betaApi = (model as Record<string, unknown>)['betaApi'];
+  if (betaApi !== undefined) {
+    extras['betaApi'] = betaApi;
+  }
+  return Object.keys(extras).length > 0 ? extras : undefined;
+}
+
 /**
 * Derive a synthetic Provider id from a Model's flat baseUrl. Uses only the
 * origin (host, optionally port) per Phase 2 decision "a=origin only" — two
--- a/packages/agent-core-v2/test/model/modelResolver.test.ts
+++ b/packages/agent-core-v2/test/model/modelResolver.test.ts
@ -0,0 +1,196 @@
+/**
+ * `model` domain — `ModelResolverService` regression tests.
+ *
+ * Covers two resolver responsibilities:
+ *  1. Auth shape — the resolved `Model` god-object drives real requests through
+ *     kosong, which reads the bearer/api token from `ProviderRequestAuth.apiKey`
+ *     (`requireProviderApiKey`). The resolver's `AuthProvider` must return the
+ *     token as `apiKey` (not wrapped in `headers`), so a resolved Model can
+ *     authenticate against its endpoint.
+ *  2. Default thinking — the resolver reads the `thinking` / `defaultThinking`
+ *     config sections and applies the same default effort the production agent
+ *     path (via `profile`) does, so a plain `model.request()` behaves
+ *     identically (some endpoints reject a request that omits thinking).
+ */
+
+import { afterEach, beforeEach, describe, expect, it, vi } from 'vitest';
+
+import { DisposableStore } from '#/_base/di/lifecycle';
+import { createServices, type TestInstantiationService } from '#/_base/di/test';
+import { IOAuthService } from '#/app/auth';
+import { IConfigService } from '#/app/config/config';
+import { type ModelConfig, IModelResolver, IModelService } from '#/app/model';
+import { ModelResolverService } from '#/app/model/modelResolverService';
+import { IPlatformService } from '#/app/platform';
+import { type ProviderConfig, IProviderService } from '#/app/provider';
+import { IProtocolAdapterRegistry } from '#/app/protocol';
+
+describe('ModelResolverService', () => {
+  let disposables: DisposableStore;
+  let ix: TestInstantiationService;
+  let providers: Record<string, ProviderConfig>;
+  let models: Record<string, ModelConfig>;
+  let configValues: Record<string, unknown>;
+  let resolveTokenProvider: ReturnType<typeof vi.fn>;
+
+  beforeEach(() => {
+    disposables = new DisposableStore();
+    providers = {};
+    models = {};
+    configValues = {};
+    resolveTokenProvider = vi.fn();
+    ix = createServices(disposables, {
+      additionalServices: (reg) => {
+        reg.definePartialInstance(IConfigService, {
+          get: ((domain: string) => configValues[domain]) as unknown as IConfigService['get'],
+        });
+        reg.definePartialInstance(IProviderService, {
+          get: ((name: string) => providers[name]) as IProviderService['get'],
+          list: (() => providers) as IProviderService['list'],
+        });
+        reg.definePartialInstance(IPlatformService, {
+          get: (() => undefined) as IPlatformService['get'],
+          list: (() => ({})) as IPlatformService['list'],
+        });
+        reg.definePartialInstance(IModelService, {
+          get: ((id: string) => models[id]) as IModelService['get'],
+          list: (() => models) as IModelService['list'],
+        });
+        reg.definePartialInstance(IOAuthService, {
+          resolveTokenProvider: resolveTokenProvider as unknown as IOAuthService['resolveTokenProvider'],
+        });
+        reg.definePartialInstance(IProtocolAdapterRegistry, {
+          supportedProtocols: () => [],
+        });
+        reg.define(IModelResolver, ModelResolverService);
+      },
+    });
+  });
+
+  afterEach(() => disposables.dispose());
+
+  it('returns the provider apiKey as ProviderRequestAuth.apiKey', async () => {
+    providers['p'] = { type: 'kimi', baseUrl: 'https://example.test/v1', apiKey: 'sk-test' };
+    models['m'] = { provider: 'p', model: 'wire-name', maxContextSize: 1000 };
+
+    const auth = await ix.get(IModelResolver).resolve('m').authProvider.getAuth();
+
+    expect(auth).toEqual({ apiKey: 'sk-test' });
+  });
+
+  it('prefers a model-inline apiKey override as ProviderRequestAuth.apiKey', async () => {
+    providers['p'] = { type: 'kimi', baseUrl: 'https://example.test/v1', apiKey: 'sk-provider' };
+    models['m'] = {
+      provider: 'p',
+      model: 'wire-name',
+      maxContextSize: 1000,
+      apiKey: 'sk-model',
+    };
+
+    const auth = await ix.get(IModelResolver).resolve('m').authProvider.getAuth();
+
+    expect(auth).toEqual({ apiKey: 'sk-model' });
+  });
+
+  it('returns an OAuth access token as ProviderRequestAuth.apiKey', async () => {
+    providers['p'] = {
+      type: 'kimi',
+      baseUrl: 'https://example.test/v1',
+      oauth: { storage: 'file', key: 'oauth/test' },
+    };
+    models['m'] = { provider: 'p', model: 'wire-name', maxContextSize: 1000 };
+    resolveTokenProvider.mockReturnValue({ getAccessToken: async () => 'oauth-token' });
+
+    const auth = await ix.get(IModelResolver).resolve('m').authProvider.getAuth();
+
+    expect(auth).toEqual({ apiKey: 'oauth-token' });
+    expect(resolveTokenProvider).toHaveBeenCalledWith('p', { storage: 'file', key: 'oauth/test' });
+  });
+
+  it('returns undefined when the model carries no auth material', async () => {
+    providers['p'] = { type: 'kimi', baseUrl: 'https://example.test/v1' };
+    models['m'] = { provider: 'p', model: 'wire-name', maxContextSize: 1000 };
+
+    const auth = await ix.get(IModelResolver).resolve('m').authProvider.getAuth();
+
+    expect(auth).toBeUndefined();
+  });
+
+  it('falls through an empty-string provider apiKey to OAuth', async () => {
+    providers['p'] = {
+      type: 'kimi',
+      baseUrl: 'https://example.test/v1',
+      apiKey: '',
+      oauth: { storage: 'file', key: 'oauth/test' },
+    };
+    models['m'] = { provider: 'p', model: 'wire-name', maxContextSize: 1000 };
+    resolveTokenProvider.mockReturnValue({ getAccessToken: async () => 'oauth-token' });
+
+    const auth = await ix.get(IModelResolver).resolve('m').authProvider.getAuth();
+
+    expect(auth).toEqual({ apiKey: 'oauth-token' });
+  });
+
+  describe('default thinking', () => {
+    function resolveEffort(capabilities?: string[]): string | null {
+      providers['p'] = { type: 'kimi', baseUrl: 'https://example.test/v1', apiKey: 'sk' };
+      models['m'] = {
+        provider: 'p',
+        model: 'wire-name',
+        maxContextSize: 1000,
+        ...(capabilities === undefined ? {} : { capabilities }),
+      };
+      return ix.get(IModelResolver).resolve('m').thinkingEffort;
+    }
+
+    it('defaults to "high" when thinking is not disabled', () => {
+      expect(resolveEffort()).toBe('high');
+    });
+
+    it('is off (null) when defaultThinking is false', () => {
+      configValues['defaultThinking'] = false;
+      expect(resolveEffort()).toBeNull();
+    });
+
+    it('is off (null) when thinking.mode is "off"', () => {
+      configValues['thinking'] = { mode: 'off' };
+      expect(resolveEffort()).toBeNull();
+    });
+
+    it('uses the configured thinking.effort', () => {
+      configValues['thinking'] = { effort: 'medium' };
+      expect(resolveEffort()).toBe('medium');
+    });
+
+    it('clamps an explicit off back to on for always_thinking models', () => {
+      configValues['defaultThinking'] = false;
+      expect(resolveEffort(['always_thinking'])).toBe('high');
+    });
+  });
+
+  describe('baseUrl normalization', () => {
+    function resolveBaseUrl(protocol: string, providerType: string, baseUrl: string): string {
+      providers['p'] = { type: providerType, baseUrl, apiKey: 'sk' } as ProviderConfig;
+      models['m'] = { provider: 'p', model: 'wire-name', maxContextSize: 1000, protocol } as ModelConfig;
+      return ix.get(IModelResolver).resolve('m').baseUrl;
+    }
+
+    it('strips a trailing /v1 for the anthropic protocol', () => {
+      expect(resolveBaseUrl('anthropic', 'kimi', 'https://example.test/coding/v1')).toBe(
+        'https://example.test/coding',
+      );
+    });
+
+    it('strips a trailing /v1/ (with slash) for the anthropic protocol', () => {
+      expect(resolveBaseUrl('anthropic', 'kimi', 'https://example.test/coding/v1/')).toBe(
+        'https://example.test/coding',
+      );
+    });
+
+    it('does not strip /v1 for non-anthropic protocols', () => {
+      expect(resolveBaseUrl('kimi', 'kimi', 'https://example.test/coding/v1')).toBe(
+        'https://example.test/coding/v1',
+      );
+    });
+  });
+});