guard model context before launch

(cherry picked from commit 14ccef8)

Author: Siri-Ray

apps/daemon/src/integrations/provider-models.ts

@@ -141,7 +141,7 @@ function openAiModelOption(item: unknown): ProviderModelOption | null {
   const obj = item as { id?: unknown; metadata?: unknown };
   const id = typeof obj?.id === 'string' ? obj.id : '';
   if (!id || !isOpenAiChatModelId(id)) return null;
-  const metadata = extractModelMetadata(obj.metadata);
+  const metadata = extractModelMetadata(obj);
   return {
     id,
     label: id,
@@ -151,16 +151,38 @@ function openAiModelOption(item: unknown): ProviderModelOption | null {
 
 function extractModelMetadata(value: unknown): ModelMetadata | null {
   if (!value || typeof value !== 'object' || Array.isArray(value)) return null;
-  const metadata = value as { cost?: unknown; capability?: unknown };
+  const record = value as Record<string, unknown>;
+  const metadata = (
+    record.metadata && typeof record.metadata === 'object' && !Array.isArray(record.metadata)
+      ? record.metadata
+      : record
+  ) as Record<string, unknown>;
   const cost = parseModelCost(metadata.cost);
   const capability = parseModelCapability(metadata.capability);
-  if (!cost && !capability) return null;
+  const contextWindowTokens = positiveInteger(
+    metadata.contextWindowTokens ??
+      metadata.context_window_tokens ??
+      record.context_length ??
+      record.contextWindow,
+  );
+  const maxOutputTokens = positiveInteger(
+    metadata.maxOutputTokens ?? metadata.max_output_tokens ?? record.max_output_tokens,
+  );
+  if (!cost && !capability && !contextWindowTokens && !maxOutputTokens) return null;
   return {
     ...(cost ? { cost } : {}),
     ...(capability ? { capability } : {}),
+    ...(contextWindowTokens ? { contextWindowTokens } : {}),
+    ...(maxOutputTokens ? { maxOutputTokens } : {}),
   };
 }
 
+function positiveInteger(value: unknown): number | null {
+  return typeof value === 'number' && Number.isInteger(value) && value > 0
+    ? value
+    : null;
+}
+
 function parseModelCost(value: unknown): ModelCost | null {
   return value === 'low' ||
     value === 'medium' ||

apps/daemon/src/langfuse-bridge.ts

@@ -1100,6 +1100,7 @@ export async function reportRunCompletedFromDaemon(
         ...(stderr ? { stderr } : {}),
         ...(stdout ? { stdout } : {}),
         diagnostics,
+        ...(run.contextBudget ? { contextBudget: run.contextBudget } : {}),
       },
       message: {
         messageId: run.assistantMessageId ?? '',

apps/daemon/src/runtimes/defs/amr.ts

@@ -198,13 +198,45 @@ function extractModelMetadata(item: unknown): ModelMetadata | null {
   const metadata = isRecord(item.metadata) ? item.metadata : item;
   const cost = parseModelCost(metadata.cost);
   const capability = parseModelCapability(metadata.capability);
-  if (!cost && !capability) return null;
+  const contextWindowTokens = extractPositiveModelLimit(item, [
+    'contextWindowTokens',
+    'context_window_tokens',
+    'context_length',
+    'contextLength',
+    'context',
+  ]);
+  const maxOutputTokens = extractPositiveModelLimit(item, [
+    'maxOutputTokens',
+    'max_output_tokens',
+    'output',
+  ]);
+  if (!cost && !capability && !contextWindowTokens && !maxOutputTokens) return null;
   return {
     ...(cost ? { cost } : {}),
     ...(capability ? { capability } : {}),
+    ...(contextWindowTokens ? { contextWindowTokens } : {}),
+    ...(maxOutputTokens ? { maxOutputTokens } : {}),
   };
 }
 
+function extractPositiveModelLimit(
+  item: Record<string, unknown>,
+  keys: string[],
+): number | null {
+  const metadata = isRecord(item.metadata) ? item.metadata : null;
+  const limit = isRecord(item.limit) ? item.limit : null;
+  for (const source of [metadata, limit, item]) {
+    if (!source) continue;
+    for (const key of keys) {
+      const value = source[key];
+      if (typeof value === 'number' && Number.isInteger(value) && value > 0) {
+        return value;
+      }
+    }
+  }
+  return null;
+}
+
 function withPriceDerivedCostMetadata(
   metadata: ModelMetadata | null,
   inputPriceUsdPerMillion: number | undefined,
@@ -462,14 +494,20 @@ function openCodeModelPrice(
 > | null {
   if (!isRecord(model)) return null;
   const inputPriceUsdPerMillion = extractInputPriceUsdPerMillion(model);
-  if (inputPriceUsdPerMillion === undefined) return null;
   const outputPriceUsdPerMillion = extractOutputPriceUsdPerMillion(model);
   const metadata = withPriceDerivedCostMetadata(
     extractModelMetadata(model),
     inputPriceUsdPerMillion,
   );
+  if (
+    inputPriceUsdPerMillion === undefined &&
+    outputPriceUsdPerMillion === undefined &&
+    metadata === null
+  ) {
+    return null;
+  }
   return {
-    inputPriceUsdPerMillion,
+    ...(inputPriceUsdPerMillion === undefined ? {} : { inputPriceUsdPerMillion }),
     ...(outputPriceUsdPerMillion === undefined ? {} : { outputPriceUsdPerMillion }),
     ...(metadata === null ? {} : { metadata }),
   };

apps/daemon/src/runtimes/model-context-budget.ts

@@ -132,7 +132,6 @@ export function compactTranscriptForSessionRollover(
     omittedMessageBlocks,
   };
 }
-
 export function evaluateModelContextBudget({
   prompt,
   modelId,

apps/daemon/src/runtimes/models.ts

@@ -45,6 +45,19 @@ export function getRememberedLiveModels(agentId: string, scope?: string | null):
   return liveModelOrder.get(liveModelCacheKey(agentId, scope)) ?? [];
 }
 
+export function getKnownModelOption(
+  def: RuntimeAgentDef,
+  modelId: string | null | undefined,
+  scope?: string | null,
+): RuntimeModelOption | null {
+  if (!modelId) return null;
+  const live = getRememberedLiveModels(def.id, scope).find(
+    (model) => model.id === modelId,
+  );
+  if (live) return live;
+  return def.fallbackModels.find((model) => model.id === modelId) ?? null;
+}
+
 export function preferFreshLiveModels(
   freshModels: RuntimeModelOption[],
   rememberedModels: RuntimeModelOption[],

apps/daemon/src/server.ts

@@ -192,6 +192,7 @@ import {
   spawnEnvForAgent,
 } from './agents.js';
 import {
+  getKnownModelOption,
   getRememberedLiveModels,
   preferFreshLiveModels,
   rememberLiveModels,

apps/daemon/tests/amr-acp-integration.test.ts

@@ -226,7 +226,12 @@ describe('AMR runtime def', () => {
           enabled: true,
           default: true,
           cost: { input: 0.14, output: 0.28 },
-          metadata: { cost: 'low', capability: 'standard' },
+          metadata: {
+            cost: 'low',
+            capability: 'standard',
+            context_window_tokens: 204800,
+            max_output_tokens: 32768,
+          },
         },
         { id: 'gpt-image-2' },
         { id: 'deepseek-v4-flash' },
@@ -240,7 +245,12 @@ describe('AMR runtime def', () => {
         default: true,
         inputPriceUsdPerMillion: 0.14,
         outputPriceUsdPerMillion: 0.28,
-        metadata: { cost: 'low', capability: 'standard' },
+        metadata: {
+          cost: 'low',
+          capability: 'standard',
+          contextWindowTokens: 204800,
+          maxOutputTokens: 32768,
+        },
       },
       { id: 'deepseek-v3.2', label: 'deepseek-v3.2' },
       { id: 'kimi-k2.7-code', label: 'kimi-k2.7-code', enabled: false },

apps/daemon/tests/runtimes/model-context-budget.test.ts

@@ -55,7 +55,6 @@ describe('model context budget', () => {
       error: { code: 'AGENT_PROMPT_TOO_LARGE' },
     });
   });
-
   it('uses provider metadata and preserves output plus safety headroom', () => {
     const decision = evaluateModelContextBudget({
       prompt: 'short prompt',

apps/daemon/tests/runtimes/run-failure-telemetry-smoke.test.ts

@@ -131,7 +131,20 @@ describe('run failure telemetry smoke', () => {
         expectedDetail: 'prompt_too_large',
         expectedDiagnosticSource: 'error_event',
         expectStderr: false,
-        message: `od-failure-smoke-context ${'large-context '.repeat(4000)}`,
+        message: `od-failure-smoke-context ${'large-context '.repeat(10_000)}`,
+      },
+      {
+        id: 'model_context_budget',
+        agentId: 'claude',
+        config: { agentCliEnv: { claude: { CLAUDE_BIN: path.join(binDir, 'claude-auth') } } },
+        model: 'claude-sonnet-4-5',
+        expectedCode: 'AGENT_PROMPT_TOO_LARGE',
+        expectedCategory: 'prompt_too_large',
+        expectedDetail: 'prompt_too_large',
+        expectedDiagnosticSource: 'error_event',
+        expectedContextBudgetAction: 'blocked',
+        expectStderr: false,
+        message: `od-failure-smoke-model-context ${'x'.repeat(650_000)}`,
       },
       {
         id: 'hang_timeout',
@@ -151,6 +164,7 @@ describe('run failure telemetry smoke', () => {
         caseId: item.id,
         agentId: item.agentId,
         message: 'message' in item ? item.message : `od-failure-smoke-${item.id}`,
+        ...('model' in item ? { model: item.model } : {}),
       });
       const events = await readRunEvents(run.eventsLogPath);
       const errorCode = deriveRunErrorCode(run);
@@ -167,20 +181,33 @@ describe('run failure telemetry smoke', () => {
         signal: run.signal,
       });
 
-      expect(run.status).toBe('failed');
+      expect(run.status, item.id).toBe('failed');
       expect('expectedCodes' in item ? item.expectedCodes : [item.expectedCode])
         .toContain(errorCode);
       expect(failure?.failure_category).toBe(item.expectedCategory);
       expect(failure?.failure_detail).toBe(item.expectedDetail);
       expect(diagnostics.diagnostic_source).toBe(item.expectedDiagnosticSource);
       expect(diagnostics.stderr_present).toBe(item.expectStderr);
+      if ('expectedContextBudgetAction' in item) {
+        expect(events).toContainEqual(expect.objectContaining({
+          event: 'diagnostic',
+          data: expect.objectContaining({
+            type: 'model_context_budget',
+            action: item.expectedContextBudgetAction,
+          }),
+        }));
+      }
 
       await finalizeAssistantMessage(started.url, run);
       const trace = await ingestion.waitForTrace(run.id);
       expect('expectedCodes' in item ? item.expectedCodes : [item.expectedCode])
         .toContain(trace.body.metadata.error_code);
       expect(trace.body.metadata.failure_category).toBe(item.expectedCategory);
       expect(trace.body.metadata.failure_detail).toBe(item.expectedDetail);
+      if ('expectedContextBudgetAction' in item) {
+        expect(trace.body.metadata.contextBudgetAction).toBe(item.expectedContextBudgetAction);
+        expect(trace.body.metadata.contextWindowTokens).toBe(204_800);
+      }
       if (item.expectStderr) {
         expect(trace.body.metadata.stderr.lineCount).toBeGreaterThan(0);
       } else {
@@ -479,6 +506,7 @@ async function createAndWaitForRun(url: string, input: {
   caseId: string;
   agentId: string;
   message: string;
+  model?: string;
 }): Promise<RunStatus> {
   const projectId = `failure_smoke_${input.caseId}_${randomUUID()}`;
   const projectResponse = await fetch(`${url}/api/projects`, {
@@ -503,6 +531,7 @@ async function createAndWaitForRun(url: string, input: {
       assistantMessageId,
       clientRequestId: `client_${input.caseId}_${randomUUID()}`,
       agentId: input.agentId,
+      ...(input.model ? { model: input.model } : {}),
       message: input.message,
       currentPrompt: input.message,
     }),

apps/web/src/types.ts

@@ -531,6 +531,8 @@ export type ModelCapability = 'standard' | 'advanced' | 'best_quality';
 export interface ModelMetadata {
   cost?: ModelCost;
   capability?: ModelCapability;
+  contextWindowTokens?: number;
+  maxOutputTokens?: number;
 }
 
 export interface AgentModelOption {