What causes "No response returned by agent creation architect." in can1357/oh-my-pi?

Retry the agent creation — transient empty responses usually succeed on a second run

No response returned by agent creation architect.

Error message

No response returned by agent creation architect.

What it means

The architect sub-session's prompt() completed but extractAssistantText found no assistant text in the session message state, so there is nothing to parse into an agent spec. This guards against empty/degenerate model responses being treated as a valid spec.

Source

Thrown at packages/coding-agent/src/modes/components/agents-hub.ts:797

			skills: [],
			contextFiles: [],
			promptTemplates: [],
			slashCommands: [],
		});
		const unsubscribe = session.subscribe(event => {
			if (event.type === "message_update" && "assistantMessageEvent" in event) {
				const ame = event.assistantMessageEvent;
				if (ame.type === "text_delta") {
					this.#createStreamingText += ame.delta;
					this.#tui.requestRender();
				}
			}
		});
		try {
			await session.prompt(userPrompt, { expandPromptTemplates: false });
			const raw = extractAssistantText(session.state.messages);
			if (!raw) {
				throw new Error("No response returned by agent creation architect.");
			}
			return parseGeneratedAgentSpec(raw);
		} finally {
			unsubscribe();
			await session.dispose();
		}
	}

	async #saveGeneratedAgent(): Promise<void> {
		const spec = this.#createSpec;
		if (!spec) return;
		const dirs = getConfigDirs("agents", {
			user: this.#createScope === "user",
			project: this.#createScope === "project",
			cwd: this.#cwd,
		});
		const targetDir = dirs[0]?.path;
		if (!targetDir) {

View on GitHub (pinned to 9690622007)

Solutions

Retry the agent creation — transient empty responses usually succeed on a second run
Check the selected model actually produces text output (not tool-only) and is configured correctly
Inspect logs/provider response for content-filter or abort causes
Switch to a different available model for the architect

Example fix

// before
selectedModel: tool-only endpoint model -> empty assistant text
// after
selectedModel: standard chat model -> assistant text parsed into spec

Defensive patterns

Strategy: retry

Validate before calling

const raw = extractAssistantText(session.state.messages);
if (!raw) throw new Error("Empty architect response — retry needed");

Type guard

function hasAssistantText(messages: unknown): boolean {
  return extractAssistantText(messages as never).trim().length > 0;
}

Try / catch

try {
  return await hub.runAgentCreationArchitect(desc);
} catch (err) {
  if (err instanceof Error && err.message.includes("No response returned")) {
    return await hub.runAgentCreationArchitect(desc); // one retry
  }
  throw err;
}

Prevention

Retry transient empty completions automatically with backoff
Select a chat-capable (non-tool-only) model for the architect
Monitor provider status for content-filter/abort incidents
Log raw session messages to diagnose empty responses

When it happens

Trigger: session.prompt(userPrompt) resolves but the session state contains no assistant message text — e.g. the model returned only tool calls, the response was empty, or messages were filtered out before extraction.

Common situations: Model hits an output-content filter; provider returns an empty completion; the session aborted mid-turn leaving no assistant text; a non-chat model misconfigured for the sub-session.

Related errors

AI-assisted analysis of can1357/oh-my-pi@9690622007 (2026-08-31). Data as JSON: /api/errors/338bdce71b2fe9d6. Report an issue: GitHub.