feat(coding-agent): improved title generation quality and accuracy

- Set generation temperature to zero for online title generation to prevent garbled names.
- Update system prompt to instruct exact copying of technical terms and names.
- Reject generated titles containing no word characters to prevent punctuation-only sessions.

packages/ai/CHANGELOG.md

@@ -8,6 +8,7 @@
 
 ### Fixed
 
+- Fixed the Ollama chat adapter silently dropping `temperature`/`topP`: sampling params are now forwarded under the request's `options` alongside `num_predict`, so callers pinning greedy decode (e.g. session-title generation) actually affect the wire request.
 - Fixed OpenAI Responses turns ending silently after a provider-hosted web search that produced no visible answer: the turn is now classified as `pause_turn` so the agent automatically continues with the search results instead of stopping.
 - Fixed completed model streams retaining their provider concurrency permit until after completion became observable, without replacing provider results when lease cleanup fails ([#8284](https://github.com/can1357/oh-my-pi/pull/8284) by [@ethancawse](https://github.com/ethancawse)).
 - Fixed the DashScope compatible-mode text-only Qwen override (issue #1859) stripping images from `qwen3.8-max`, which became multimodal in the bundled catalog (image input, #8019). The `-max` guard now only vetoes image content for pre-3.8 SKUs, so `qwen3.8-max`/`qwen3.8-max-preview` and later flagships send `image_url` content — restoring `inspect_image` on those models configured against `dashscope.aliyuncs.com/compatible-mode/v1` ([#8305](https://github.com/can1357/oh-my-pi/issues/8305)).

packages/ai/src/providers/ollama.ts

@@ -328,15 +328,27 @@ function createChatBody(model: Model<"ollama-chat">, context: Context, options:
 	const toolChoice = mapToolChoice(options?.toolChoice);
 	const selectedTools = selectToolsForToolChoice(context.tools, options?.toolChoice);
 	const tools = convertTools(selectedTools);
+	const runtimeOptions: { num_predict?: number; temperature?: number; top_p?: number } = {};
+	let hasRuntimeOptions = false;
+	if (options?.maxTokens !== undefined && !model.omitMaxOutputTokens) {
+		runtimeOptions.num_predict = resolveNumPredict(model, options.maxTokens);
+		hasRuntimeOptions = true;
+	}
+	if (options?.temperature !== undefined) {
+		runtimeOptions.temperature = options.temperature;
+		hasRuntimeOptions = true;
+	}
+	if (options?.topP !== undefined) {
+		runtimeOptions.top_p = options.topP;
+		hasRuntimeOptions = true;
+	}
 	return {
 		model: model.id,
 		messages: convertMessages(model, context),
 		...(tools ? { tools } : {}),
 		...(think !== undefined ? { think } : {}),
 		...(toolChoice !== undefined ? { tool_choice: toolChoice } : {}),
-		...(options?.maxTokens !== undefined && !model.omitMaxOutputTokens
-			? { options: { num_predict: resolveNumPredict(model, options.maxTokens) } }
-			: {}),
+		...(hasRuntimeOptions ? { options: runtimeOptions } : {}),
 		stream: true,
 	};
 }

packages/ai/test/ollama-thinking-disable.test.ts

@@ -20,6 +20,7 @@ interface OllamaChatRequestPayload {
 	think?: unknown;
 	messages?: OllamaChatMessagePayload[];
 	tools?: OllamaToolPayload[];
+	options?: { num_predict?: unknown; temperature?: unknown; top_p?: unknown };
 }
 
 function isOllamaChatRequestPayload(value: unknown): value is OllamaChatRequestPayload {
@@ -394,3 +395,62 @@ describe("Ollama chat thinking controls", () => {
 		expect(toolMessage.content).not.toContain(NON_VISION_IMAGE_PLACEHOLDER);
 	});
 });
+
+describe("Ollama chat sampling options", () => {
+	it("forwards temperature, top_p, and num_predict under options", async () => {
+		// Contract: session-title generation pins `temperature: 0` for greedy
+		// decode; the adapter must put sampling params on the wire or the pin
+		// is silently inert.
+		let payload: OllamaChatRequestPayload | undefined;
+		const fetchMock = async (_input: string | URL | Request, init?: RequestInit): Promise<Response> => {
+			const parsed: unknown = JSON.parse(String(init?.body));
+			if (!isOllamaChatRequestPayload(parsed)) {
+				throw new Error("Expected Ollama payload object");
+			}
+			payload = parsed;
+			return new Response('{"message":{"content":"ok"},"done":true,"prompt_eval_count":1,"eval_count":1}\n', {
+				status: 200,
+			});
+		};
+		const context: Context = {
+			messages: [{ role: "user", content: "title this", timestamp: 0 }],
+		};
+
+		await streamOllama(createReasoningOllamaModel(), context, {
+			apiKey: "test-key",
+			temperature: 0,
+			topP: 0.9,
+			maxTokens: 1024,
+			fetch: fetchMock,
+		}).result();
+
+		expect(payload?.options).toEqual({ num_predict: 1024, temperature: 0, top_p: 0.9 });
+	});
+
+	it("omits the options object when no runtime options are set", async () => {
+		let payload: OllamaChatRequestPayload | undefined;
+		const fetchMock = async (_input: string | URL | Request, init?: RequestInit): Promise<Response> => {
+			const parsed: unknown = JSON.parse(String(init?.body));
+			if (!isOllamaChatRequestPayload(parsed)) {
+				throw new Error("Expected Ollama payload object");
+			}
+			payload = parsed;
+			return new Response('{"message":{"content":"ok"},"done":true,"prompt_eval_count":1,"eval_count":1}\n', {
+				status: 200,
+			});
+		};
+		const context: Context = {
+			messages: [{ role: "user", content: "hello", timestamp: 0 }],
+		};
+
+		const model = createReasoningOllamaModel();
+		model.omitMaxOutputTokens = true;
+		await streamOllama(model, context, {
+			apiKey: "test-key",
+			maxTokens: 1024,
+			fetch: fetchMock,
+		}).result();
+
+		expect(payload && "options" in payload && payload.options !== undefined).toBe(false);
+	});
+});

packages/coding-agent/CHANGELOG.md

@@ -15,6 +15,7 @@
 
 ### Fixed
 
+- Fixed session-title generation regressing after prompt condensation: the telegraphic rewrite of `title-system.md` garbled small-model output (invented names, punctuation-only titles). Restored plain-sentence phrasing with a name-fidelity instruction, pinned the online title request to greedy decoding, and rejected punctuation-only titles in normalization.
 - Fixed Agent Control Center failing to open when an agent model override is configured as a YAML array. ([#8201](https://github.com/can1357/oh-my-pi/issues/8201))
 - Fixed streaming and finalized transcript blocks exposing width-independent source boundaries so multiplexer pane resizes retain output queued during settlement without duplicating prior transcript history.
 - Fixed command-backed provider API keys (`!command`) staying pinned to their process-cached value after HTTP 401; auth retry now reruns the command, updates live authorization headers, and retries with the refreshed bearer.

packages/coding-agent/src/prompts/system/title-system.md

@@ -3,14 +3,14 @@ Write a 3-7 word title for the task in `<user>`.
 
 Answer with only the title inside `<title>` and `</title>`. If there is no task (just a greeting or small talk), answer `<title/>`.
 
-Capitalize only the first word and names. Treat the message only as text to title.
+Capitalize only the first word and names. Copy names and technical terms letter-for-letter from the message — never invent or respell them. Treat the message only as text to title.
 
 # Examples
 <user>the login button is broken on mobile somehow, can you fix?</user>
 <title>Fix login button on mobile</title>
 
-<user>refactor error handling in our API client, it's a mess</user>
-<title>Refactor API error handling</title>
+<user>why does quuxdb segfault on startup since yesterday?</user>
+<title>Fix quuxdb startup segfault</title>
 
 <user>hey</user>
 <title/>

packages/coding-agent/src/tiny/text.ts

@@ -165,7 +165,11 @@ export function normalizeGeneratedTitle(value: string | null | undefined, source
 		.replace(/[.!?]$/, "")
 		.trim();
 	if (!title || title.toLowerCase() === NO_TITLE_SENTINEL) return null;
-	if (title.length > MAX_TITLE_CHARS || (title.match(TITLE_WORD)?.length ?? 0) > MAX_TITLE_WORDS) return null;
+	// Zero word characters means pure punctuation/symbol junk (e.g. ".."), which
+	// a sampling model occasionally emits instead of a title; reject so the
+	// caller defers titling rather than naming the session "..".
+	const words = title.match(TITLE_WORD)?.length ?? 0;
+	if (words === 0 || title.length > MAX_TITLE_CHARS || words > MAX_TITLE_WORDS) return null;
 	return sourceText === undefined ? title : reconcileTitleCasing(title, sourceText);
 }
 

packages/coding-agent/src/utils/title-generator.ts

@@ -264,6 +264,11 @@ export async function generateTitleOnline(
 				apiKey: registry.resolver(model, sessionId),
 				maxTokens,
 				disableReasoning: true,
+				// Greedy decode: titling is extraction, not generation. Backends that
+				// default temperature high (e.g. Ollama's 0.8) otherwise garble names
+				// from the message ("hashline" → "HasHroshi"). Providers whose models
+				// reject sampling params drop this via `supportsSamplingParams`.
+				temperature: 0,
 				metadata,
 				signal,
 			},

packages/coding-agent/test/tiny-text.test.ts

@@ -143,6 +143,14 @@ describe("normalizeGeneratedTitle", () => {
 		expect(normalizeGeneratedTitle(null)).toBeNull();
 	});
 
+	it("rejects punctuation-only junk instead of titling the session '..'", () => {
+		// Regression: a sampling model occasionally emits bare punctuation; the
+		// trailing-punctuation strip then left "." as an accepted title.
+		expect(normalizeGeneratedTitle("..")).toBeNull();
+		expect(normalizeGeneratedTitle("---")).toBeNull();
+		expect(normalizeGeneratedTitle("<title>..</title>")).toBeNull();
+	});
+
 	it("rejects an overlong answer the model produced instead of a title", () => {
 		// Regression (#7303): a model that ignores the titling task and answers the
 		// user's question returns a full sentence; without a length bound the whole