@opengsd/gsd-pi 1.0.1-dev.bf2899e → 1.0.2-dev.0ee79e5

package/src/resources/extensions/gsd/uok/timeline.ts

@@ -1,4 +1,4 @@
-// GSD2 UOK Timeline Reconstruction from Authoritative DB Records
+// gsd-pi UOK Timeline Reconstruction from Authoritative DB Records
 
 import { existsSync, readFileSync } from "node:fs";
 import { join } from "node:path";

package/src/resources/extensions/gsd/validation-block-guard.ts

@@ -1,4 +1,4 @@
-// Project/App: GSD-2
+// Project/App: gsd-pi
 // File Purpose: Shared command gate for validation-blocked milestones.
 
 import { existsSync } from "node:fs";

package/src/resources/extensions/gsd/verification-gate.ts

@@ -41,7 +41,7 @@ const PACKAGE_SCRIPT_KEYS = ["typecheck", "lint", "test"] as const;
 
 /**
  * Discover verification commands using the first-non-empty-wins strategy (D003):
- *   1. Task plan verify field (split on && and newlines)
+ *   1. Task plan verify field (split on newlines)
  *   2. Explicit preference commands
  *   3. package.json scripts (typecheck, lint, test)
  *   4. Python pytest project markers
@@ -59,7 +59,7 @@ export function discoverCommands(options: DiscoverCommandsOptions): DiscoveredCo
   if (taskPlanVerify) {
     const commands: string[] = [];
     const candidates = taskPlanVerify
-      .split(/&&|\r?\n/)
+      .split(/\r?\n/)
       .map(c => c.trim())
       .filter(Boolean);
     for (const candidate of candidates) {

package/src/resources/extensions/gsd/verification-verdict.ts

@@ -1,4 +1,4 @@
-// Project/App: GSD-2
+// Project/App: gsd-pi
 // File Purpose: Host-owned verification verdict policy for auto-mode units.
 
 import type { VerificationResult as VerificationGateResult } from "./types.js";

package/src/resources/extensions/gsd/watch/header-renderer.ts

@@ -1,4 +1,4 @@
-// Project/App: GSD-2
+// Project/App: gsd-pi
 // File Purpose: Watch-mode terminal header and splash renderer for GSD project status.
 
 import { execFileSync } from "node:child_process";

package/src/resources/extensions/gsd/watch/splash-palette.ts

@@ -1,4 +1,4 @@
-// Project/App: GSD-2
+// Project/App: gsd-pi
 // File Purpose: Shared RGB palette for the pre-theme GSD watch splash.
 
 export const splashPalette = {

package/src/resources/extensions/gsd/workflow-manifest.ts

@@ -1,4 +1,4 @@
-// Project/App: GSD-2
+// Project/App: gsd-pi
 // File Purpose: State manifest snapshot and restore orchestration for GSD workflow data.
 
 import {

package/src/resources/extensions/gsd/workflow-projections.ts

@@ -1,4 +1,4 @@
-// Project/App: GSD-2
+// Project/App: gsd-pi
 // File Purpose: Projection renderers for GSD workflow database rows.
 // GSD Extension — Projection Renderers (DB -> Markdown)
 // Renders PLAN.md, ROADMAP.md, SUMMARY.md, and STATE.md from database rows.

package/src/resources/extensions/gsd/workflow-protocol.ts

@@ -1,4 +1,4 @@
-// Project/App: GSD-2
+// Project/App: gsd-pi
 // File Purpose: Shared capped workflow protocol and doctor-heal prompt payload helpers.
 
 const DEFAULT_WORKFLOW_PROTOCOL_EXCERPT_CHARS = 4_000;

package/src/resources/extensions/gsd/workflow-reconcile.ts

@@ -105,6 +105,13 @@ function replayEvents(events: WorkflowEvent[]): void {
         updateTaskStatus(milestoneId, sliceId, taskId, "in-progress", event.ts);
         break;
       }
+      case "skip_task": {
+        const milestoneId = p["milestoneId"] as string;
+        const sliceId = p["sliceId"] as string;
+        const taskId = p["taskId"] as string;
+        updateTaskStatus(milestoneId, sliceId, taskId, "skipped", event.ts);
+        break;
+      }
       case "report_blocker": {
         const milestoneId = p["milestoneId"] as string;
         const sliceId = p["sliceId"] as string;
@@ -236,6 +243,7 @@ export function extractEntityKey(
   switch (cmd) {
     case "complete_task":
     case "start_task":
+    case "skip_task":
     case "report_blocker":
     case "record_verification":
     case "plan_task":

package/src/resources/extensions/gsd/workflow-templates.ts

@@ -1,4 +1,4 @@
-// Project/App: GSD-2
+// Project/App: gsd-pi
 // File Purpose: Workflow template registry loading, matching, and display helpers.
 
 /**

package/src/resources/extensions/gsd/working-output-messages.ts

@@ -1,4 +1,4 @@
-// GSD-2 + src/resources/extensions/gsd/working-output-messages.ts - Formats and audits user-facing working-state messages.
+// gsd-pi + src/resources/extensions/gsd/working-output-messages.ts - Formats and audits user-facing working-state messages.
 
 export type WorkingOutputSurface =
   | "loader"

package/src/resources/extensions/gsd/workspace.ts

@@ -1,4 +1,4 @@
-// GSD-2 + Workspace handle: single source of truth for path resolution per milestone
+// gsd-pi + Workspace handle: single source of truth for path resolution per milestone
 
 import { join, resolve } from "node:path";
 import { type GsdPathContract, resolveGsdPathContract, normalizeRealPath } from "./paths.js";

package/src/resources/extensions/gsd/worktree-lifecycle.ts

@@ -1,4 +1,4 @@
-// GSD-2 — Worktree Lifecycle module: owns milestone entry/exit lifecycle behind a small, typed Interface.
+// gsd-pi — Worktree Lifecycle module: owns milestone entry/exit lifecycle behind a small, typed Interface.
 /**
  * Worktree Lifecycle module — first-class Module for worktree create/enter/exit/merge.
  *

package/src/resources/extensions/gsd/worktree-manager.ts

@@ -1,4 +1,4 @@
-// Project/App: GSD-2
+// Project/App: gsd-pi
 // File Purpose: Creates, resolves, and reconciles GSD milestone worktrees.
 
 /**

package/src/resources/extensions/gsd/worktree-safety.ts

@@ -1,4 +1,4 @@
-// Project/App: GSD-2
+// Project/App: gsd-pi
 // File Purpose: Worktree Safety module contract for validating source-writing Unit roots.
 
 import { existsSync, lstatSync, type Stats } from "node:fs";

package/src/resources/extensions/gsd/worktree-state-projection.ts

@@ -1,4 +1,4 @@
-// GSD-2 — Worktree State Projection module: directional state-flow rules between project root and auto-worktree.
+// gsd-pi — Worktree State Projection module: directional state-flow rules between project root and auto-worktree.
 /**
  * Worktree State Projection module — first-class Module for directional
  * state-file flow between the project root and the auto-worktree.

package/src/resources/extensions/mcp-client/manager.ts

@@ -1,5 +1,5 @@
 /**
- * GSD-2 — Shared MCP connection management.
+ * gsd-pi — Shared MCP connection management.
  *
  * File Purpose: Reads, writes, and tests user-configured MCP server entries
  * for runtime tools, TUI commands, and app settings surfaces.

package/src/resources/extensions/mcp-client/tests/manager.test.ts

@@ -1,5 +1,5 @@
 /**
- * GSD-2 — MCP manager tests.
+ * gsd-pi — MCP manager tests.
  *
  * File Purpose: Behaviour coverage for shared MCP config management and
  * side-effect-free connection testing.

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

@@ -1,4 +1,4 @@
-// GSD2 — Ollama Extension: First-class local LLM support
+// gsd-pi — Ollama Extension: First-class local LLM support
 /**
  * Ollama Extension
  *

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

@@ -1,4 +1,4 @@
-// GSD2 — Known model capability table for Ollama models
+// gsd-pi — Known model capability table for Ollama models
 
 /**
  * Maps well-known Ollama model families to their capabilities.
@@ -24,11 +24,18 @@ 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.
+// Note: ollamaOptions.num_ctx is set when the context window has an authoritative
+// source — either a KNOWN_MODELS table entry, or /api/show returning context_length
+// at runtime (ollama-discovery.ts syncs num_ctx with the /api/show value when present).
+// When neither source provides a context window, num_ctx is NOT sent and ollama
+// uses its own safe default to avoid OOM on constrained hosts.
 const KNOWN_MODELS: Array<[pattern: string, caps: ModelCapability]> = [
 	// ─── Reasoning models ───────────────────────────────────────────────
+	// Long-variants listed before the bare `deepseek-v4` base to avoid prefix shadowing.
+	// Same invariant as qwen3-coder / glm / kimi / minimax families.
+	["deepseek-v4-pro",   { contextWindow: 1048576, reasoning: true, ollamaOptions: { num_ctx: 1048576 } }],
+	["deepseek-v4-flash", { contextWindow: 1048576, reasoning: true, ollamaOptions: { num_ctx: 1048576 } }],
+	["deepseek-v4",       { contextWindow: 1048576, reasoning: true, ollamaOptions: { num_ctx: 1048576 } }],
 	["deepseek-r1", { contextWindow: 131072, reasoning: true, ollamaOptions: { num_ctx: 131072 } }],
 	["qwq", { contextWindow: 131072, reasoning: true, ollamaOptions: { num_ctx: 131072 } }],
 
@@ -90,11 +97,15 @@ const KNOWN_MODELS: Array<[pattern: string, caps: ModelCapability]> = [
 
 	// ─── MiniMax M2 (Ollama Cloud) ─────────────────────────────────────
 	// ref: minimax-m2 1M ctx — https://www.minimax.io/news/minimax-m2
-	["minimax-m2.7", { contextWindow: 1048576, maxTokens: 16384, ollamaOptions: { num_ctx: 1048576 } }],
+	// minimax-m2.7:cloud reports 196608 via /api/show despite the M2 announcement
+	// quoting 1M context. Cloud deployment truncates / OOMs at the announced
+	// number; trust the deployed backend.
+	["minimax-m2.7", { contextWindow: 196608, maxTokens: 16384, reasoning: true, ollamaOptions: { num_ctx: 196608 } }],
 	["minimax-m2.5", { contextWindow: 1048576, maxTokens: 16384, ollamaOptions: { num_ctx: 1048576 } }],
 	["minimax-m2", { contextWindow: 1048576, maxTokens: 16384, ollamaOptions: { num_ctx: 1048576 } }],
 
 	// ─── Gemma family ───────────────────────────────────────────────────
+	["gemma4", { contextWindow: 262144, reasoning: true, ollamaOptions: { num_ctx: 262144 } }],
 	["gemma3", { contextWindow: 131072, maxTokens: 16384, ollamaOptions: { num_ctx: 131072 } }],
 	["gemma2", { contextWindow: 8192, maxTokens: 8192, ollamaOptions: { num_ctx: 8192 } }],
 

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

@@ -1,4 +1,4 @@
-// GSD2 — Ollama Extension: NDJSON streaming parser
+// gsd-pi — Ollama Extension: NDJSON streaming parser
 
 /**
  * Parses a streaming NDJSON (newline-delimited JSON) response body into

package/src/resources/extensions/ollama/ollama-chat-provider.ts

@@ -1,4 +1,4 @@
-// GSD2 — Ollama Extension: Native /api/chat stream provider
+// gsd-pi — Ollama Extension: Native /api/chat stream provider
 
 /**
  * Implements the "ollama-chat" API provider, streaming responses directly
@@ -21,6 +21,7 @@ import {
 	type StopReason,
 	type TextContent,
 	type ThinkingContent,
+	type ThinkingLevel,
 	type Tool,
 	type ToolCall,
 	type Usage,
@@ -146,6 +147,13 @@ export function streamOllamaChat(
 			}
 
 			for await (const chunk of chat(request, options?.signal)) {
+				// Native thinking field (ollama 0.4+ for reasoning models). Emit before
+				// content so blocks open in the natural reasoning-then-answer order.
+				const thinking = chunk.message?.thinking ?? "";
+				if (thinking) {
+					emitDelta("thinking", thinking);
+				}
+
 				// Handle text content — process independently of tool_calls
 				// (a chunk may contain both content and tool_calls)
 				const content = chunk.message?.content ?? "";
@@ -189,7 +197,31 @@ export function streamOllamaChat(
 
 // ─── Request building ───────────────────────────────────────────────────────
 
-function buildRequest(
+/**
+ * Map a SimpleStreamOptions.reasoning ThinkingLevel to ollama's `think` field.
+ *
+ * - Returns `undefined` when the model doesn't support reasoning, or when no
+ *   level was requested (preserving existing default behaviour: ollama uses
+ *   the model's built-in default).
+ * - `"minimal"` turns thinking off (`think: false`) — universally supported.
+ * - `"low" | "medium" | "high"` are passed through as strings; gpt-oss honors
+ *   them as strength levels, other thinking models treat any present value
+ *   as "thinking enabled".
+ * - `"xhigh"` is collapsed to `"high"` (ollama caps strength at high).
+ */
+export function buildThinkParam(
+	model: Model<Api>,
+	options?: SimpleStreamOptions,
+): boolean | "low" | "medium" | "high" | undefined {
+	if (!model.reasoning) return undefined;
+	const level: ThinkingLevel | undefined = options?.reasoning;
+	if (level === undefined) return undefined;
+	if (level === "minimal") return false;
+	if (level === "xhigh") return "high";
+	return level;
+}
+
+export function buildRequest(
 	model: Model<Api>,
 	context: Context,
 	options?: SimpleStreamOptions,
@@ -245,6 +277,13 @@ function buildRequest(
 		request.tools = convertTools(context.tools);
 	}
 
+	// Thinking strength (ollama 0.4+). Only set when both the model is reasoning-capable
+	// and the caller explicitly requested a level — otherwise ollama uses its default.
+	const think = buildThinkParam(model, options);
+	if (think !== undefined) {
+		request.think = think;
+	}
+
 	return request;
 }
 

package/src/resources/extensions/ollama/ollama-client.ts

@@ -1,4 +1,4 @@
-// GSD2 — HTTP client for Ollama REST API
+// gsd-pi — HTTP client for Ollama REST API
 
 /**
  * Low-level HTTP client for the Ollama REST API.

package/src/resources/extensions/ollama/ollama-commands.ts

@@ -1,4 +1,4 @@
-// GSD2 — Ollama slash commands
+// gsd-pi — Ollama slash commands
 
 /**
  * Registers /ollama slash commands for managing local Ollama models.

package/src/resources/extensions/ollama/ollama-discovery.ts

@@ -1,4 +1,4 @@
-// GSD2 — Ollama model discovery and capability detection
+// gsd-pi — Ollama model discovery and capability detection
 
 /**
  * Discovers locally available Ollama models and enriches them with
@@ -15,7 +15,7 @@ import {
 	getModelCapabilities,
 	humanizeModelName,
 } from "./model-capabilities.js";
-import type { OllamaChatOptions, OllamaModelInfo } from "./types.js";
+import type { OllamaChatOptions, OllamaModelInfo, OllamaShowResponse } from "./types.js";
 
 /**
  * Extract context window from /api/show model_info.
@@ -57,35 +57,59 @@ async function enrichModel(info: OllamaModelInfo, deps: ClientDeps): Promise<Dis
 	const caps = getModelCapabilities(info.name);
 	const parameterSize = info.details?.parameter_size ?? "";
 
-	// /api/tags doesn't include context length; /api/show does via "{arch}.context_length" in model_info.
+	// Always call /api/show — it carries two pieces of info absent from /api/tags:
+	// the per-architecture context_length, and (ollama 0.4+) a `capabilities` array
+	// like ["thinking", "completion", "tools"]. The capabilities array is the
+	// authoritative source for reasoning detection on cloud-routed models that
+	// don't appear in the static KNOWN_MODELS table.
 	let showContextWindow: number | undefined;
-	if (caps.contextWindow === undefined) {
-		try {
-			const showData = await deps.showModel(info.name);
-			showContextWindow = extractContextFromModelInfo(showData.model_info);
-		} catch (err) {
-			// non-fatal: fall through to estimate
-			if (process.env.GSD_DEBUG) console.warn(`[ollama] /api/show failed for ${info.name}:`, err instanceof Error ? err.message : String(err));
-		}
+	let showCapabilities: string[] | undefined;
+	try {
+		const showData = await deps.showModel(info.name);
+		showContextWindow = extractContextFromModelInfo(showData.model_info);
+		showCapabilities = showData.capabilities;
+	} catch (err) {
+		// non-fatal: fall through to table/estimate
+		if (process.env.GSD_DEBUG) console.warn(`[ollama] /api/show failed for ${info.name}:`, err instanceof Error ? err.message : String(err));
 	}
 
-	// Determine context window: known table > /api/show > estimate from param size > default
+	// Determine context window: /api/show (authoritative ollama metadata) >
+	// known table (fallback for old ollama versions / network failure) >
+	// estimate from parameter size > default. Earlier priority order put
+	// known table first, but the table fell behind reality on several
+	// model families (deepseek-v4-* 131072 vs real 1048576; minimax-m2.7
+	// 1048576 vs real 196608). /api/show is the source of truth when
+	// reachable; the table only fills the gap when it isn't.
 	const contextWindow =
-		caps.contextWindow ??
 		showContextWindow ??
+		caps.contextWindow ??
 		(parameterSize ? estimateContextFromParams(parameterSize) : 8192);
 
 	// Determine max tokens: known table > fraction of context > default
 	const maxTokens =
 		caps.maxTokens ?? Math.min(Math.floor(contextWindow / 4), 16384);
 
-	// Detect vision from families or known table
+	// Detect vision: /api/show capabilities > known table > model families heuristic
 	const hasVision =
+		showCapabilities?.includes("vision") ??
 		caps.input?.includes("image") ??
 		(info.details?.families?.some((f) => f === "clip" || f === "mllama") ?? false);
 
-	// Detect reasoning from known table
-	const reasoning = caps.reasoning ?? false;
+	// Detect reasoning: /api/show capabilities (authoritative) > known table fallback
+	const reasoning =
+		showCapabilities?.includes("thinking") ??
+		caps.reasoning ??
+		false;
+
+	// Sync num_ctx with the authoritative contextWindow. When /api/show
+	// wins, the table's static num_ctx would otherwise be stale and sent
+	// on every chat request — the very drift this commit's priority flip
+	// was designed to eliminate. Keep all other ollamaOptions (num_gpu,
+	// sampling params, keep_alive) from the table.
+	const ollamaOptions =
+		showContextWindow !== undefined
+			? { ...caps.ollamaOptions, num_ctx: showContextWindow }
+			: caps.ollamaOptions;
 
 	return {
 		id: info.name,
@@ -97,12 +121,20 @@ async function enrichModel(info: OllamaModelInfo, deps: ClientDeps): Promise<Dis
 		maxTokens,
 		sizeBytes: info.size,
 		parameterSize,
-		ollamaOptions: caps.ollamaOptions,
+		ollamaOptions,
 	};
 }
 
 /**
  * Discover all locally available Ollama models with enriched capabilities.
+ *
+ * /api/show is now invoked for every model (capabilities + context_length
+ * both live there). Requests are dispatched concurrently here, but ollama's
+ * local server serializes model-metadata calls internally, so wall time
+ * scales linearly with model count — empirically ~50-100 ms/model on a warm
+ * ollama instance. A user with 30+ models pays ~3 s on app start. If this
+ * becomes a complaint, gate the call on `caps.contextWindow === undefined`
+ * and `caps.reasoning === undefined` (i.e., unknown-model fast path).
  */
 export async function discoverModels(deps: ClientDeps = { listModels, showModel }): Promise<DiscoveredOllamaModel[]> {
 	const tags = await deps.listModels();

package/src/resources/extensions/ollama/ollama-tool.ts

@@ -1,4 +1,4 @@
-// GSD2 — LLM-callable Ollama management tool
+// gsd-pi — LLM-callable Ollama management tool
 /**
  * Registers an ollama_manage tool that the LLM can call to interact
  * with the local Ollama instance — list models, pull new ones, check status.

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

@@ -1,4 +1,4 @@
-// GSD2 — Tests for Ollama model capability detection
+// gsd-pi — Tests for Ollama model capability detection
 import { describe, it } from "node:test";
 import assert from "node:assert/strict";
 import {
@@ -149,9 +149,8 @@ describe("getModelCapabilities — long-variant overrides aren't shadowed (#4991
 		assert.equal(caps.contextWindow, 262144);
 	});
 
-	it("minimax-m2.5:cloud and minimax-m2.7:cloud report 1M", () => {
+	it("minimax-m2.5:cloud reports 1M", () => {
 		assert.equal(getModelCapabilities("minimax-m2.5:cloud").contextWindow, 1048576);
-		assert.equal(getModelCapabilities("minimax-m2.7:cloud").contextWindow, 1048576);
 	});
 
 	it("minimax-m2 base resolves to 1M", () => {
@@ -256,3 +255,49 @@ describe("formatModelSize", () => {
 		assert.equal(formatModelSize(500_000), "500 KB");
 	});
 });
+
+// ─── deepseek-v4 prefix-shadowing regression ────────────────────────────────
+//
+// deepseek-v4-pro:cloud and deepseek-v4-flash:cloud must be listed before the
+// bare `deepseek-v4` entry in KNOWN_MODELS, otherwise the linear startsWith
+// scan resolves any deepseek-v4-* query to the family base. Same invariant
+// as the qwen3-coder / glm / kimi families already pin elsewhere.
+
+describe("getModelCapabilities — deepseek-v4 long-variants aren't shadowed", () => {
+	it("deepseek-v4-pro:cloud and deepseek-v4-flash:cloud resolve to 1M (long-variants beat deepseek-v4 base)", () => {
+		assert.equal(getModelCapabilities("deepseek-v4-pro:cloud").contextWindow, 1048576);
+		assert.equal(getModelCapabilities("deepseek-v4-flash:cloud").contextWindow, 1048576);
+	});
+
+	it("deepseek-v4 base also resolves to 1M (parity with long-variants)", () => {
+		const caps = getModelCapabilities("deepseek-v4:671b");
+		assert.equal(caps.contextWindow, 1048576);
+	});
+
+	it("ollamaOptions.num_ctx mirrors contextWindow for all deepseek-v4 / gemma4 entries", () => {
+		// Inference time: num_ctx is what gets sent to Ollama on each chat.
+		// If contextWindow is right but num_ctx is stale, the model still
+		// gets truncated. Pin both sides.
+		for (const name of [
+			"deepseek-v4-pro:cloud",
+			"deepseek-v4-flash:cloud",
+			"deepseek-v4:671b",
+			"gemma4:31b",
+		]) {
+			const caps = getModelCapabilities(name);
+			assert.equal(caps.ollamaOptions?.num_ctx, caps.contextWindow,
+				`${name}: num_ctx ${caps.ollamaOptions?.num_ctx} != contextWindow ${caps.contextWindow}`);
+		}
+	});
+});
+
+describe("getModelCapabilities — minimax-m2.7 reflects /api/show truth", () => {
+	it("minimax-m2.7 contextWindow is 196608, not the official-spec 1048576", () => {
+		// minimax-m2.7:cloud reports 196608 via /api/show even though the
+		// MiniMax M2 announcement quoted 1M context. Trust the deployed
+		// backend, not marketing — a 1M num_ctx would silently truncate
+		// or OOM under cloud-routing.
+		assert.equal(getModelCapabilities("minimax-m2.7:cloud").contextWindow, 196608);
+		assert.equal(getModelCapabilities("minimax-m2.7:cloud").ollamaOptions?.num_ctx, 196608);
+	});
+});

package/src/resources/extensions/ollama/tests/ollama-chat-provider-stream.test.ts

@@ -1,4 +1,4 @@
-// GSD2 — Regression test: Ollama streaming must not drop content on done:true chunks (#3576)
+// gsd-pi — Regression test: Ollama streaming must not drop content on done:true chunks (#3576)
 // Copyright (c) 2026 Jeremy McSpadden <jeremy@fluxlabs.net>
 
 import { describe, it } from "node:test";

package/src/resources/extensions/ollama/tests/ollama-chat-provider-think.test.ts

@@ -0,0 +1,87 @@
+// gsd-pi — Tests for ollama think-parameter mapping
+import { describe, it } from "node:test";
+import assert from "node:assert/strict";
+import type { Api, Context, Model, SimpleStreamOptions } from "@gsd/pi-ai";
+import { buildRequest, buildThinkParam } from "../ollama-chat-provider.js";
+
+function modelStub(reasoning: boolean, id = "gpt-oss:20b"): Model<Api> {
+	return {
+		id,
+		name: id,
+		api: "openai-completions" as Api,
+		provider: "ollama",
+		baseUrl: "http://localhost:11434",
+		reasoning,
+		input: ["text"],
+		cost: { input: 0, output: 0, cacheRead: 0, cacheWrite: 0 },
+		contextWindow: 131072,
+		maxTokens: 32768,
+	};
+}
+
+describe("buildThinkParam — gating", () => {
+	it("returns undefined when model.reasoning is false", () => {
+		const m = modelStub(false);
+		const opts: SimpleStreamOptions = { reasoning: "high" };
+		assert.equal(buildThinkParam(m, opts), undefined);
+	});
+
+	it("returns undefined when options.reasoning is undefined (preserve existing default)", () => {
+		const m = modelStub(true);
+		assert.equal(buildThinkParam(m, {}), undefined);
+		assert.equal(buildThinkParam(m, undefined), undefined);
+	});
+});
+
+describe("buildThinkParam — ThinkingLevel mapping", () => {
+	const m = modelStub(true);
+
+	it("maps 'minimal' to false (turn thinking off)", () => {
+		assert.equal(buildThinkParam(m, { reasoning: "minimal" }), false);
+	});
+
+	it("passes 'low' through as string", () => {
+		assert.equal(buildThinkParam(m, { reasoning: "low" }), "low");
+	});
+
+	it("passes 'medium' through as string", () => {
+		assert.equal(buildThinkParam(m, { reasoning: "medium" }), "medium");
+	});
+
+	it("passes 'high' through as string", () => {
+		assert.equal(buildThinkParam(m, { reasoning: "high" }), "high");
+	});
+
+	it("collapses 'xhigh' to 'high' (ollama caps at high)", () => {
+		assert.equal(buildThinkParam(m, { reasoning: "xhigh" }), "high");
+	});
+});
+
+describe("buildRequest — wire-level think field", () => {
+	const emptyContext: Context = { messages: [] } as unknown as Context;
+
+	it("sets request.think = false when reasoning='minimal' on a reasoning model", () => {
+		// Wire-level coverage: buildThinkParam returns false for 'minimal',
+		// and the buildRequest guard must let falsy-but-not-undefined values
+		// through. A naive `if (think) request.think = think` would drop
+		// 'minimal' entirely and leave the model's default thinking on.
+		const req = buildRequest(modelStub(true), emptyContext, { reasoning: "minimal" });
+		assert.equal(req.think, false);
+		assert.ok('think' in req, "'think' key must be present in request when minimal is set — guards against future `delete request.think` refactors");
+	});
+
+	it("sets request.think = 'high' when reasoning='high'", () => {
+		const req = buildRequest(modelStub(true), emptyContext, { reasoning: "high" });
+		assert.equal(req.think, "high");
+	});
+
+	it("omits request.think when reasoning is unset (preserve model default)", () => {
+		const req = buildRequest(modelStub(true), emptyContext, {});
+		assert.equal(req.think, undefined);
+	});
+
+	it("omits request.think on non-reasoning models even when reasoning is requested", () => {
+		const req = buildRequest(modelStub(false), emptyContext, { reasoning: "high" });
+		assert.equal(req.think, undefined);
+	});
+});

package/src/resources/extensions/ollama/tests/ollama-client-timeout-env.test.ts

@@ -1,4 +1,4 @@
-// GSD2 — Tests for OLLAMA_PROBE_TIMEOUT_MS / OLLAMA_REQUEST_TIMEOUT_MS env vars (#5003 / #4982)
+// gsd-pi — Tests for OLLAMA_PROBE_TIMEOUT_MS / OLLAMA_REQUEST_TIMEOUT_MS env vars (#5003 / #4982)
 //
 // Pinned defaults: 1500 ms probe, 10 000 ms request. The defaults must be
 // preserved when the env var is unset, empty, non-numeric, zero, or negative

package/src/resources/extensions/ollama/tests/ollama-client.test.ts

@@ -1,4 +1,4 @@
-// GSD2 — Tests for Ollama HTTP client
+// gsd-pi — Tests for Ollama HTTP client
 import { describe, it, beforeEach, afterEach } from "node:test";
 import assert from "node:assert/strict";
 import { getOllamaHost } from "../ollama-client.js";