Q76: runs call the agent's own model, or fail plainly where the provider cannot

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

src/main/services/provider-response-runner.ts

@@ -17,6 +17,7 @@ import { selectToolTurn, supportsToolTurn } from "./provider-tool-turns/router";
 import { resolveTurnCredential } from "./provider-tool-turns/credential-resolver";
 import { currentRunId } from "./run-control";
 import { budgetBlock, isPriced, promptChars, providerPrices } from "./spend";
+import { effectiveModel } from "./task-run-preparation";
 import {
   HEALTH_STATE_HEALTHY,
   WORKSPACE_SCOPES,
@@ -116,15 +117,19 @@ export class ProviderResponseRunner {
     if (!supportsToolTurn(provider.provider_key)) {
       throw new Error("Agent execution is not available for this provider; assign the agent an OpenAI, Anthropic, xAI, Groq, Mistral, Cerebras, OpenRouter, GPT account, GitHub Copilot, Google Antigravity, or Claude Pro provider.");
     }
-    const model = provider.model;
+    // A call made for a run works for its agent: the agent's model, its budget, its attribution.
+    const taskRunId = currentRunId();
+    const agent = taskRunId
+      ? this.deps.db.prepare("select a.id, a.department, a.model_id from task_runs r join agents a on a.id = r.agent_id where r.id = ?").get(taskRunId) as { id: string; department: string; model_id: string | null }
+      : null;
+    // Azure calls the deployment stored with its credential, so an agent there cannot pick another model.
+    if (provider.provider_key === "azure-openai" && agent?.model_id && agent.model_id !== provider.model) {
+      throw new Error(`Azure OpenAI calls the deployment configured on the provider; this agent asks for "${agent.model_id}". Clear the agent's model or set it to "${provider.model}".`);
+    }
+    const model = effectiveModel(agent?.model_id ?? null, provider.model);
     if (!model) throw new Error("Provider model is required.");
     const owner: ProviderRow & { model: string } = { ...provider, model };
     const turn = this.deps.toolTurn ?? selectToolTurn(provider.provider_key);
-    // A call made for a run is spent by its agent: refuse before calling once its budget is gone.
-    const taskRunId = currentRunId();
-    const agent = taskRunId
-      ? this.deps.db.prepare("select a.id, a.department from task_runs r join agents a on a.id = r.agent_id where r.id = ?").get(taskRunId) as { id: string; department: string }
-      : null;
     const blocked = agent ? budgetBlock(this.deps.db, agent.id, isPriced(providerPrices(this.deps.db, provider.id)), new Date()) : null;
     if (blocked) throw new Error(blocked);
     // Everything the model reads: instructions, the conversation, and the tool definitions.

src/main/services/task-execution-service.ts

@@ -37,7 +37,7 @@ import { ConnectorGrantsService } from "./composio/connector-grants";
 import { ProviderCooldowns } from "./provider-cooldowns";
 import { HeartbeatScheduler } from "./heartbeat-scheduler";
 import {
-  assertExecutionApproval, createIsolationContext, executionApprovalProblem, loadGoal, loadProject, loadWorkspace,
+  assertExecutionApproval, createIsolationContext, effectiveModel, executionApprovalProblem, loadGoal, loadProject, loadWorkspace,
   requireAgentProvider, requireAssignedAgent, requireTask, sameWorkspaceAccess, workspaceAccess,
 } from "./task-run-preparation";
 
@@ -175,7 +175,7 @@ export class TaskExecutionService {
     let providerRequestId: string | null = null;
     let skillRuntimeSnapshot: SkillRuntimeSnapshotRow | null = null;
     try {
-      skillRuntimeSnapshot = this.resolveSkillRuntimeSnapshot(orgId, agent.id, runId, skillBudgetTokens(provider.context_window_tokens), `${task.title}\n${task.description}`, { providerKey: provider.provider_key, model: provider.model }, task.writing_mode === WRITING_MODE.CONVERSATIONAL);
+      skillRuntimeSnapshot = this.resolveSkillRuntimeSnapshot(orgId, agent.id, runId, skillBudgetTokens(provider.context_window_tokens), `${task.title}\n${task.description}`, { providerKey: provider.provider_key, model: effectiveModel(agent.model_id, provider.model) }, task.writing_mode === WRITING_MODE.CONVERSATIONAL);
       // Watermark for reply delivery and mid-run message polling.
       const runStartedAt = new Date().toISOString();
       const chatContext = loadChatWakeContext(this.db, orgId, agent.id, isolation.secretScopeJson);

src/main/services/task-prompts.ts

@@ -1,5 +1,6 @@
 import { WRITING_MODE, type WritingMode } from "../../shared/domain";
 import { CEO_FIRST_HIRES, firstHiresFor, leadsPeople } from "./org-building";
+import { effectiveModel } from "./task-run-preparation";
 import { canonicalMentionHandle } from "../../shared/mention-handles";
 
 const RECORD_LEARNINGS_INSTRUCTION =
@@ -220,7 +221,7 @@ export function buildTaskPrompt(input: {
     `Requires filesystem: ${input.task.requires_filesystem === 1 ? "yes" : "no"}`,
     `Assignee: ${input.agent.name} (${input.agent.role})`,
     ...(input.managerName ? [`Reports to: ${input.managerName}`] : []),
-    `Provider: ${input.provider.label} ${input.provider.model ?? input.agent.model_id ?? ""}`.trim(),
+    `Provider: ${input.provider.label} ${effectiveModel(input.agent.model_id, input.provider.model) ?? ""}`.trim(),
     `Execution context: ${input.isolation.executionContextId}`,
     `Service identity: ${input.isolation.serviceIdentityId}`,
   ];

src/main/services/task-run-preparation.ts

@@ -67,6 +67,11 @@ export function sameWorkspaceAccess(mountJson: string, current: Array<{ workspac
   return JSON.stringify(recorded) === JSON.stringify(current);
 }
 
+/** The model a run calls: the agent's own choice, or else its provider's default. */
+export function effectiveModel(agentModelId: string | null, providerModel: string | null): string | null {
+  return agentModelId?.trim() || providerModel;
+}
+
 export function requireAgentProvider(db: DatabaseSync, orgId: string, agent: TaskExecutionAgentRow): TaskExecutionProviderRow {
   if (!agent.model_provider_id) throw new Error("Task assignee requires a model provider.");
   const provider = db.prepare(`