feat(backends): add resolveModel() to AgentEngine interface

src/backends/adapter.ts

@@ -133,6 +133,7 @@ async function buildExecutionPlan(
 	gitHubToken: string | undefined,
 	isGitHubAck: boolean,
 	engineId: string,
+	engine: AgentEngine,
 ): Promise<
 	Omit<AgentExecutionPlan, 'progressReporter'> & {
 		reviewSidecarPath?: string;
@@ -173,7 +174,7 @@ async function buildExecutionPlan(
 	const {
 		systemPrompt,
 		taskPrompt: taskPromptOverride,
-		model,
+		model: rawModel,
 		maxIterations,
 		contextFiles,
 	} = await resolveModelConfig({
@@ -186,6 +187,9 @@ async function buildExecutionPlan(
 		agentInput: input,
 	});
 
+	// Allow the engine to resolve/validate the model string (e.g. strip provider prefix)
+	const model = engine.resolveModel ? engine.resolveModel(rawModel) : rawModel;
+
 	const profile = await getAgentProfile(agentType);
 
 	// Use profile to fetch agent-specific context injections
@@ -412,6 +416,7 @@ async function resolvePartialExecutionPlan(
 			gitHubToken,
 			isGitHubAck,
 			engine.definition.id,
+			engine,
 		);
 
 	const partialInput = gitHubToken

src/backends/claude-code/index.ts

@@ -456,6 +456,10 @@ export class ClaudeCodeEngine implements AgentEngine {
 		return true;
 	}
 
+	resolveModel(cascadeModel: string): string {
+		return resolveClaudeModel(cascadeModel);
+	}
+
 	async execute(input: AgentExecutionPlan): Promise<AgentEngineResult> {
 		const startTime = Date.now();
 		const systemPrompt = buildSystemPrompt(input.systemPrompt, input.availableTools);
@@ -464,6 +468,11 @@ export class ClaudeCodeEngine implements AgentEngine {
 			input.contextInjections,
 			input.repoDir,
 		);
+		// Resolve model again here for backward compatibility: execute() may be called
+		// directly (e.g. in tests) without going through the adapter, so we cannot rely
+		// solely on the adapter's engine.resolveModel() pre-resolution. Since
+		// resolveClaudeModel() is idempotent, calling it twice via the normal adapter path
+		// is safe.
 		const model = resolveClaudeModel(input.model);
 
 		input.logWriter('INFO', 'Starting Claude Code SDK execution', {

src/backends/codex/index.ts

@@ -480,6 +480,10 @@ export class CodexEngine implements AgentEngine {
 		return true;
 	}
 
+	resolveModel(cascadeModel: string): string {
+		return resolveCodexModel(cascadeModel);
+	}
+
 	async execute(input: AgentExecutionPlan): Promise<AgentEngineResult> {
 		const startTime = Date.now();
 		const systemPrompt = buildSystemPrompt(input.systemPrompt, input.availableTools);
@@ -488,6 +492,11 @@ export class CodexEngine implements AgentEngine {
 			input.contextInjections,
 			input.repoDir,
 		);
+		// Resolve model again here for backward compatibility: execute() may be called
+		// directly (e.g. in tests) without going through the adapter, so we cannot rely
+		// solely on the adapter's engine.resolveModel() pre-resolution. Since
+		// resolveCodexModel() is idempotent, calling it twice via the normal adapter path
+		// is safe.
 		const model = resolveCodexModel(input.model);
 		const settings = resolveCodexSettings(input.project, input.nativeToolCapabilities);
 		assertHeadlessCodexSettings(settings);

src/backends/opencode/index.ts

@@ -801,9 +801,18 @@ export class OpenCodeEngine implements AgentEngine {
 		return true;
 	}
 
+	resolveModel(cascadeModel: string): string {
+		return resolveOpenCodeModel(cascadeModel);
+	}
+
 	async execute(input: AgentExecutionPlan): Promise<AgentEngineResult> {
 		const settings = resolveOpenCodeSettings(input.project);
 		const agent = 'build' as const;
+		// Resolve model again here for backward compatibility: execute() may be called
+		// directly (e.g. in tests) without going through the adapter, so we cannot rely
+		// solely on the adapter's engine.resolveModel() pre-resolution. Since
+		// resolveOpenCodeModel() is idempotent, calling it twice via the normal adapter path
+		// is safe.
 		const model = resolveOpenCodeModel(input.model);
 		const config = buildConfig(input, model, settings);
 		const { prompt: taskPrompt, hasOffloadedContext } = await buildTaskPrompt(

src/backends/types.ts

@@ -147,4 +147,10 @@ export interface AgentEngine {
 
 	execute(input: AgentExecutionPlan): Promise<AgentEngineResult>;
 	supportsAgentType(agentType: string): boolean;
+	/**
+	 * Optionally resolve a CASCADE model string to the engine-specific model identifier.
+	 * Engines that need model validation (e.g., Claude Code, Codex) implement this method.
+	 * Engines that pass the model through unchanged (e.g., LLMist) do not need to implement it.
+	 */
+	resolveModel?(cascadeModel: string): string;
 }