gsd-pi 2.63.0 → 2.64.0-dev.1a85e85

package/src/resources/extensions/ollama/index.ts

@@ -17,19 +17,10 @@
  */
 
 import { importExtensionModule, type ExtensionAPI } from "@gsd/pi-coding-agent";
-import type { OpenAICompletionsCompat } from "@gsd/pi-ai";
 import * as client from "./ollama-client.js";
-import { discoverModels, getOllamaOpenAIBaseUrl } from "./ollama-discovery.js";
+import { discoverModels } from "./ollama-discovery.js";
 import { registerOllamaCommands } from "./ollama-commands.js";
-
-/** Default compat settings for Ollama models via OpenAI-compat endpoint */
-const OLLAMA_COMPAT: OpenAICompletionsCompat = {
-	supportsDeveloperRole: false,
-	supportsReasoningEffort: false,
-	supportsUsageInStreaming: false,
-	maxTokensField: "max_tokens",
-	supportsStore: false,
-};
+import { streamOllamaChat } from "./ollama-chat-provider.js";
 
 let toolsPromise: Promise<void> | null = null;
 
@@ -68,12 +59,18 @@ async function probeAndRegister(pi: ExtensionAPI): Promise<boolean> {
 	const models = await discoverModels();
 	if (models.length === 0) return true; // Running but no models pulled
 
-	const baseUrl = getOllamaOpenAIBaseUrl();
+	const baseUrl = client.getOllamaHost();
 
+	// Use authMode "apiKey" with a dummy key (#3440).
+	// authMode "none" requires a custom streamSimple handler, but Ollama uses
+	// the standard OpenAI-compatible streaming endpoint. Ollama ignores the
+	// Authorization header so the dummy key is harmless.
 	pi.registerProvider("ollama", {
-		authMode: "none",
+		authMode: "apiKey",
+		apiKey: "ollama",
 		baseUrl,
-		api: "openai-completions",
+		api: "ollama-chat",
+		streamSimple: streamOllamaChat,
 		isReady: () => true,
 		models: models.map((m) => ({
 			id: m.id,
@@ -83,7 +80,7 @@ async function probeAndRegister(pi: ExtensionAPI): Promise<boolean> {
 			cost: m.cost,
 			contextWindow: m.contextWindow,
 			maxTokens: m.maxTokens,
-			compat: OLLAMA_COMPAT,
+			providerOptions: (m.ollamaOptions ?? {}) as Record<string, unknown>,
 		})),
 	});
 
@@ -108,16 +105,20 @@ export default function ollama(pi: ExtensionAPI) {
 			await registerOllamaTools(pi);
 		}
 
-		// Async probe — don't block startup
-		probeAndRegister(pi)
-			.then((found) => {
-				if (found && ctx.hasUI) {
-					ctx.ui.setStatus("ollama", "Ollama");
-				}
-			})
-			.catch(() => {
-				// Silently ignore probe failures
-			});
+		// In headless/auto mode, await the probe so the fallback resolver can
+		// see Ollama before the first LLM call (#3531 race condition).
+		// In interactive mode, keep it async for fast startup.
+		if (!ctx.hasUI) {
+			try {
+				await probeAndRegister(pi);
+			} catch { /* non-fatal */ }
+		} else {
+			probeAndRegister(pi)
+				.then((found) => {
+					if (found) ctx.ui.setStatus("ollama", "Ollama");
+				})
+				.catch(() => {});
+		}
 	});
 
 	pi.on("session_shutdown", async () => {

package/src/resources/extensions/ollama/model-capabilities.ts

@@ -8,11 +8,15 @@
  * Fallback: estimate from parameter count if model isn't in the table.
  */
 
+import type { OllamaChatOptions } from "./types.js";
+
 export interface ModelCapability {
 	contextWindow?: number;
 	maxTokens?: number;
 	input?: ("text" | "image")[];
 	reasoning?: boolean;
+	/** Ollama-specific default inference options for this model family. */
+	ollamaOptions?: OllamaChatOptions;
 }
 
 /**
@@ -20,58 +24,61 @@ export interface ModelCapability {
  * Keys are matched as prefixes against the model name (before the colon/tag).
  * More specific entries should appear first.
  */
+// Note: ollamaOptions.num_ctx is set for known model families where the context
+// window is authoritative. For unknown/estimated models, num_ctx is NOT sent
+// to avoid OOM risk — Ollama uses its own safe default instead.
 const KNOWN_MODELS: Array<[pattern: string, caps: ModelCapability]> = [
 	// ─── Reasoning models ───────────────────────────────────────────────
-	["deepseek-r1", { contextWindow: 131072, reasoning: true }],
-	["qwq", { contextWindow: 131072, reasoning: true }],
+	["deepseek-r1", { contextWindow: 131072, reasoning: true, ollamaOptions: { num_ctx: 131072 } }],
+	["qwq", { contextWindow: 131072, reasoning: true, ollamaOptions: { num_ctx: 131072 } }],
 
 	// ─── Vision models ──────────────────────────────────────────────────
-	["llava", { contextWindow: 4096, input: ["text", "image"] }],
-	["bakllava", { contextWindow: 4096, input: ["text", "image"] }],
-	["moondream", { contextWindow: 8192, input: ["text", "image"] }],
-	["llama3.2-vision", { contextWindow: 131072, input: ["text", "image"] }],
-	["minicpm-v", { contextWindow: 4096, input: ["text", "image"] }],
+	["llava", { contextWindow: 4096, input: ["text", "image"], ollamaOptions: { num_ctx: 4096 } }],
+	["bakllava", { contextWindow: 4096, input: ["text", "image"], ollamaOptions: { num_ctx: 4096 } }],
+	["moondream", { contextWindow: 8192, input: ["text", "image"], ollamaOptions: { num_ctx: 8192 } }],
+	["llama3.2-vision", { contextWindow: 131072, input: ["text", "image"], ollamaOptions: { num_ctx: 131072 } }],
+	["minicpm-v", { contextWindow: 4096, input: ["text", "image"], ollamaOptions: { num_ctx: 4096 } }],
 
 	// ─── Code models ────────────────────────────────────────────────────
-	["codestral", { contextWindow: 262144, maxTokens: 32768 }],
-	["qwen2.5-coder", { contextWindow: 131072, maxTokens: 32768 }],
-	["deepseek-coder-v2", { contextWindow: 131072, maxTokens: 16384 }],
-	["starcoder2", { contextWindow: 16384, maxTokens: 8192 }],
-	["codegemma", { contextWindow: 8192, maxTokens: 8192 }],
-	["codellama", { contextWindow: 16384, maxTokens: 8192 }],
-	["devstral", { contextWindow: 131072, maxTokens: 32768 }],
+	["codestral", { contextWindow: 262144, maxTokens: 32768, ollamaOptions: { num_ctx: 262144 } }],
+	["qwen2.5-coder", { contextWindow: 131072, maxTokens: 32768, ollamaOptions: { num_ctx: 131072 } }],
+	["deepseek-coder-v2", { contextWindow: 131072, maxTokens: 16384, ollamaOptions: { num_ctx: 131072 } }],
+	["starcoder2", { contextWindow: 16384, maxTokens: 8192, ollamaOptions: { num_ctx: 16384 } }],
+	["codegemma", { contextWindow: 8192, maxTokens: 8192, ollamaOptions: { num_ctx: 8192 } }],
+	["codellama", { contextWindow: 16384, maxTokens: 8192, ollamaOptions: { num_ctx: 16384 } }],
+	["devstral", { contextWindow: 131072, maxTokens: 32768, ollamaOptions: { num_ctx: 131072 } }],
 
 	// ─── Llama family ───────────────────────────────────────────────────
-	["llama3.3", { contextWindow: 131072, maxTokens: 16384 }],
-	["llama3.2", { contextWindow: 131072, maxTokens: 16384 }],
-	["llama3.1", { contextWindow: 131072, maxTokens: 16384 }],
-	["llama3", { contextWindow: 8192, maxTokens: 8192 }],
-	["llama2", { contextWindow: 4096, maxTokens: 4096 }],
+	["llama3.3", { contextWindow: 131072, maxTokens: 16384, ollamaOptions: { num_ctx: 131072 } }],
+	["llama3.2", { contextWindow: 131072, maxTokens: 16384, ollamaOptions: { num_ctx: 131072 } }],
+	["llama3.1", { contextWindow: 131072, maxTokens: 16384, ollamaOptions: { num_ctx: 131072 } }],
+	["llama3", { contextWindow: 8192, maxTokens: 8192, ollamaOptions: { num_ctx: 8192 } }],
+	["llama2", { contextWindow: 4096, maxTokens: 4096, ollamaOptions: { num_ctx: 4096 } }],
 
 	// ─── Qwen family ────────────────────────────────────────────────────
-	["qwen3", { contextWindow: 131072, maxTokens: 32768 }],
-	["qwen2.5", { contextWindow: 131072, maxTokens: 32768 }],
-	["qwen2", { contextWindow: 131072, maxTokens: 32768 }],
+	["qwen3", { contextWindow: 131072, maxTokens: 32768, ollamaOptions: { num_ctx: 131072 } }],
+	["qwen2.5", { contextWindow: 131072, maxTokens: 32768, ollamaOptions: { num_ctx: 131072 } }],
+	["qwen2", { contextWindow: 131072, maxTokens: 32768, ollamaOptions: { num_ctx: 131072 } }],
 
 	// ─── Gemma family ───────────────────────────────────────────────────
-	["gemma3", { contextWindow: 131072, maxTokens: 16384 }],
-	["gemma2", { contextWindow: 8192, maxTokens: 8192 }],
+	["gemma3", { contextWindow: 131072, maxTokens: 16384, ollamaOptions: { num_ctx: 131072 } }],
+	["gemma2", { contextWindow: 8192, maxTokens: 8192, ollamaOptions: { num_ctx: 8192 } }],
 
 	// ─── Mistral family ─────────────────────────────────────────────────
-	["mistral-large", { contextWindow: 131072, maxTokens: 16384 }],
-	["mistral-small", { contextWindow: 131072, maxTokens: 16384 }],
-	["mistral-nemo", { contextWindow: 131072, maxTokens: 16384 }],
-	["mistral", { contextWindow: 32768, maxTokens: 8192 }],
-	["mixtral", { contextWindow: 32768, maxTokens: 8192 }],
+	["mistral-large", { contextWindow: 131072, maxTokens: 16384, ollamaOptions: { num_ctx: 131072 } }],
+	["mistral-small", { contextWindow: 131072, maxTokens: 16384, ollamaOptions: { num_ctx: 131072 } }],
+	["mistral-nemo", { contextWindow: 131072, maxTokens: 16384, ollamaOptions: { num_ctx: 131072 } }],
+	["mistral", { contextWindow: 32768, maxTokens: 8192, ollamaOptions: { num_ctx: 32768 } }],
+	["mixtral", { contextWindow: 32768, maxTokens: 8192, ollamaOptions: { num_ctx: 32768 } }],
 
 	// ─── Phi family ─────────────────────────────────────────────────────
-	["phi4", { contextWindow: 16384, maxTokens: 16384 }],
-	["phi3.5", { contextWindow: 131072, maxTokens: 16384 }],
-	["phi3", { contextWindow: 131072, maxTokens: 4096 }],
+	["phi4", { contextWindow: 16384, maxTokens: 16384, ollamaOptions: { num_ctx: 16384 } }],
+	["phi3.5", { contextWindow: 131072, maxTokens: 16384, ollamaOptions: { num_ctx: 131072 } }],
+	["phi3", { contextWindow: 131072, maxTokens: 4096, ollamaOptions: { num_ctx: 131072 } }],
 
 	// ─── Command R ──────────────────────────────────────────────────────
-	["command-r-plus", { contextWindow: 131072, maxTokens: 16384 }],
-	["command-r", { contextWindow: 131072, maxTokens: 16384 }],
+	["command-r-plus", { contextWindow: 131072, maxTokens: 16384, ollamaOptions: { num_ctx: 131072 } }],
+	["command-r", { contextWindow: 131072, maxTokens: 16384, ollamaOptions: { num_ctx: 131072 } }],
 ];
 
 /**

package/src/resources/extensions/ollama/ndjson-stream.ts

@@ -0,0 +1,63 @@
+// GSD2 — Ollama Extension: NDJSON streaming parser
+
+/**
+ * Parses a streaming NDJSON (newline-delimited JSON) response body into
+ * typed objects. Used for Ollama's /api/chat and /api/pull endpoints.
+ *
+ * @param strict When true, malformed JSON lines throw instead of being skipped.
+ *   Use strict mode for inference streams where silent data loss is unacceptable.
+ *   Use permissive mode (default) for progress endpoints like /api/pull.
+ */
+
+export async function* parseNDJsonStream<T>(
+	body: ReadableStream<Uint8Array>,
+	signal?: AbortSignal,
+	strict = false,
+): AsyncGenerator<T> {
+	const reader = body.getReader();
+	const decoder = new TextDecoder();
+	let buffer = "";
+
+	try {
+		while (true) {
+			if (signal?.aborted) break;
+
+			const { done, value } = await reader.read();
+			if (done) break;
+
+			buffer += decoder.decode(value, { stream: true });
+			const lines = buffer.split("\n");
+			buffer = lines.pop() ?? "";
+
+			for (const line of lines) {
+				const trimmed = line.trim();
+				if (!trimmed) continue;
+				try {
+					yield JSON.parse(trimmed) as T;
+				} catch (err) {
+					if (strict) {
+						throw new Error(
+							`Malformed NDJSON line from Ollama: ${trimmed.slice(0, 200)}`,
+						);
+					}
+					// Permissive mode: skip malformed lines
+				}
+			}
+		}
+
+		// Flush remaining buffer (skip if aborted)
+		if (buffer.trim() && !signal?.aborted) {
+			try {
+				yield JSON.parse(buffer.trim()) as T;
+			} catch (err) {
+				if (strict) {
+					throw new Error(
+						`Malformed NDJSON line from Ollama: ${buffer.trim().slice(0, 200)}`,
+					);
+				}
+			}
+		}
+	} finally {
+		reader.releaseLock();
+	}
+}

package/src/resources/extensions/ollama/ollama-auth-mode.test.ts

@@ -0,0 +1,20 @@
+/**
+ * Regression test for #3440: Ollama extension must register with
+ * authMode "apiKey" (not "none") to avoid streamSimple requirement.
+ */
+import { test } from "node:test";
+import assert from "node:assert/strict";
+import { readFileSync } from "node:fs";
+import { join, dirname } from "node:path";
+import { fileURLToPath } from "node:url";
+
+const __dirname = dirname(fileURLToPath(import.meta.url));
+
+test("Ollama registers with authMode apiKey, not none (#3440)", () => {
+  const src = readFileSync(join(__dirname, "index.ts"), "utf-8");
+  // Find the registerProvider call
+  const registerBlock = src.slice(src.indexOf("pi.registerProvider(\"ollama\""));
+  const authLine = registerBlock.match(/authMode:\s*"(\w+)"/);
+  assert.ok(authLine, "registerProvider must specify authMode");
+  assert.equal(authLine![1], "apiKey", "authMode must be apiKey, not none");
+});