pi-ui-extend 0.1.60 → 0.1.63

package/dist/app/commands/command-model-actions.js

@@ -63,8 +63,8 @@ export class ModelCommandActions {
         const parsed = parseScopedModelRef(modelRef);
         if (!parsed)
             throw new Error("Model must use provider/model[:thinking] format");
-        runtime.services.modelRegistry.refresh();
-        const model = runtime.services.modelRegistry.find(parsed.provider, parsed.modelId);
+        await runtime.services.modelRuntime.reloadConfig();
+        const model = runtime.services.modelRuntime.getModel(parsed.provider, parsed.modelId);
         if (!model)
             throw new Error(`Model not found: ${parsed.provider}/${parsed.modelId}`);
         await this.runModelCommand(model);
@@ -97,8 +97,8 @@ export class ModelCommandActions {
         const parsed = parseScopedModelRef(modelRef);
         if (!parsed)
             throw new Error("Model must use provider/model[:thinking] format");
-        runtime.services.modelRegistry.refresh();
-        const model = runtime.services.modelRegistry.find(parsed.provider, parsed.modelId);
+        await runtime.services.modelRuntime.reloadConfig();
+        const model = runtime.services.modelRuntime.getModel(parsed.provider, parsed.modelId);
         if (!model)
             throw new Error(`Model not found: ${parsed.provider}/${parsed.modelId}`);
         this.saveDefaultModel(modelRef);
@@ -118,8 +118,8 @@ export class ModelCommandActions {
         const parsed = parseScopedModelRef(modelRef);
         if (!parsed)
             throw new Error("Model must use provider/model[:thinking] format, or run /autocomplete with no arguments to disable");
-        runtime.services.modelRegistry.refresh();
-        const model = runtime.services.modelRegistry.find(parsed.provider, parsed.modelId);
+        await runtime.services.modelRuntime.reloadConfig();
+        const model = runtime.services.modelRuntime.getModel(parsed.provider, parsed.modelId);
         if (!model)
             throw new Error(`Model not found: ${parsed.provider}/${parsed.modelId}`);
         const saved = savePixAutocompleteModel(modelRef);
@@ -199,10 +199,10 @@ export class ModelCommandActions {
         const refs = value.split(/[,\s]+/).map((ref) => ref.trim()).filter(Boolean);
         const scopedModels = [];
         const invalidRefs = [];
-        runtime.services.modelRegistry.refresh();
+        await runtime.services.modelRuntime.reloadConfig();
         for (const ref of refs) {
             const parsed = parseScopedModelRef(ref);
-            const model = parsed ? runtime.services.modelRegistry.find(parsed.provider, parsed.modelId) : undefined;
+            const model = parsed ? runtime.services.modelRuntime.getModel(parsed.provider, parsed.modelId) : undefined;
             if (!parsed || !model) {
                 invalidRefs.push(ref);
                 continue;

package/dist/app/commands/command-session-actions.js

@@ -12,7 +12,7 @@ import { checkPixUpdate, formatPixUpdateCheck, parsePixUpdateArgs, pixUpdateUsag
 import { createStartupInfoMessage } from "../cli/startup-info.js";
 import { getCompleteSessionStats } from "../session/session-stats.js";
 import { loadSessionTitleConfig } from "../../bundled-extensions/session-title/config.js";
-import { fallbackSessionTitleFromInput, firstUserMessageText, generateSessionTitle, sessionTitleModelRefs, } from "../../bundled-extensions/session-title/title-generation.js";
+import { fallbackSessionTitleFromInput, firstUserMessageText, generateSessionTitleWithRuntime, sessionTitleModelRefs, } from "../../bundled-extensions/session-title/title-generation.js";
 export class SessionCommandActions {
     host;
     constructor(host) {
@@ -116,7 +116,7 @@ export class SessionCommandActions {
             if (config.enabled) {
                 for (const modelRef of modelRefs) {
                     for (let attempt = 0; attempt < config.generationAttempts; attempt++) {
-                        generatedName = await generateSessionTitle(firstPrompt.slice(0, config.maxInputChars).trim(), runtime.services.modelRegistry, config, modelRef, AbortSignal.timeout(config.timeoutMs));
+                        generatedName = await generateSessionTitleWithRuntime(firstPrompt.slice(0, config.maxInputChars).trim(), runtime.services.modelRuntime, config, modelRef, AbortSignal.timeout(config.timeoutMs));
                         if (generatedName)
                             break;
                     }

package/dist/app/input/autocomplete-controller.js

@@ -1,4 +1,3 @@
-import { streamSimple } from "@earendil-works/pi-ai/compat";
 import { isRecord } from "../guards.js";
 import { parseModelRef } from "../model/model-ref.js";
 const AUTOCOMPLETE_DEBOUNCE_MS = 350;
@@ -140,17 +139,14 @@ export class AppAutocompleteController {
 }
 export async function completeInputWithPi(runtime, draft, config, signal) {
     const parsedModel = parseModelRef(config.modelRef);
-    const registry = runtime.services.modelRegistry;
-    let model = registry.find(parsedModel.provider, parsedModel.modelId);
+    const modelRuntime = runtime.services.modelRuntime;
+    let model = modelRuntime.getModel(parsedModel.provider, parsedModel.modelId);
     if (!model) {
-        registry.refresh();
-        model = registry.find(parsedModel.provider, parsedModel.modelId);
+        await modelRuntime.reloadConfig();
+        model = modelRuntime.getModel(parsedModel.provider, parsedModel.modelId);
     }
     if (!model)
         throw new Error(`Model not found: ${parsedModel.provider}/${parsedModel.modelId}`);
-    const auth = await registry.getApiKeyAndHeaders(model);
-    if (!auth.ok)
-        throw new Error(auth.error);
     const timeoutMs = numberInRange(config.timeoutMs, AUTOCOMPLETE_TIMEOUT_MS, 250, 10_000);
     const maxTokens = numberInRange(config.maxTokens, AUTOCOMPLETE_MAX_TOKENS, 8, 256);
     const maxPromptTokens = numberInRange(config.maxPromptTokens, AUTOCOMPLETE_MAX_PROMPT_TOKENS, 256, 16_000);
@@ -165,12 +161,10 @@ export async function completeInputWithPi(runtime, draft, config, signal) {
     let output = "";
     let streamError;
     try {
-        const stream = streamSimple(requestModel, {
+        const stream = modelRuntime.streamSimple(requestModel, {
             systemPrompt: AUTOCOMPLETE_SYSTEM_PROMPT,
             messages: [{ role: "user", content: prompt, timestamp: Date.now() }],
         }, {
-            ...(auth.apiKey === undefined ? {} : { apiKey: auth.apiKey }),
-            ...(auth.headers === undefined ? {} : { headers: auth.headers }),
             cacheRetention: "none",
             maxRetryDelayMs: 0,
             maxRetries: 0,

package/dist/app/input/prompt-enhancer-controller.js

@@ -127,9 +127,8 @@ async function enhancePromptWithPi(runtime, draft, config) {
     const services = await promptEnhancerPiDeps.createAgentSessionServices({
         cwd: runtime.cwd,
         agentDir: runtime.services.agentDir,
-        authStorage: runtime.services.authStorage,
         settingsManager: runtime.services.settingsManager,
-        modelRegistry: runtime.services.modelRegistry,
+        modelRuntime: runtime.services.modelRuntime,
         resourceLoaderOptions: {
             noExtensions: true,
             noSkills: true,
@@ -139,10 +138,10 @@ async function enhancePromptWithPi(runtime, draft, config) {
             systemPrompt: PROMPT_ENHANCER_SYSTEM_PROMPT,
         },
     });
-    services.modelRegistry.refresh();
-    const model = services.modelRegistry.find(parsedModel.provider, parsedModel.modelId);
+    await services.modelRuntime.reloadConfig();
+    const model = services.modelRuntime.getModel(parsedModel.provider, parsedModel.modelId);
     if (!model) {
-        throw new Error(modelNotFoundMessage(parsedModel.provider, parsedModel.modelId, services.modelRegistry.getAll()));
+        throw new Error(modelNotFoundMessage(parsedModel.provider, parsedModel.modelId, services.modelRuntime.getModels()));
     }
     const { session } = await promptEnhancerPiDeps.createAgentSessionFromServices({
         services,

package/dist/app/model/model-usage-status.js

@@ -2,7 +2,7 @@ import { createHash } from "node:crypto";
 import { readFile } from "node:fs/promises";
 import { homedir } from "node:os";
 import { join } from "node:path";
-import { AuthStorage } from "@earendil-works/pi-coding-agent";
+import { ModelRuntime, readStoredCredential } from "@earendil-works/pi-coding-agent";
 import { formatCompactProgressBar } from "../../context-progress-bar.js";
 import { APP_ICONS } from "../icons.js";
 const OPENAI_USAGE_URL = "https://chatgpt.com/backend-api/wham/usage";
@@ -222,17 +222,27 @@ async function queryOpenAIAccountUsage(now) {
 async function refreshOpenAICodexAuth() {
     // Delegate to pi core so refresh-token rotation is persisted under the same
     // cross-process auth.json lock used by model requests.
-    const authStorage = AuthStorage.create(getPiAuthPath());
-    const access = await authStorage.getApiKey("openai-codex", { includeFallback: false });
-    const credential = authStorage.get("openai-codex");
+    const modelRuntime = await ModelRuntime.create({
+        authPath: getPiAuthPath(),
+        allowModelNetwork: false,
+    });
+    let resolvedAuth;
+    try {
+        resolvedAuth = await modelRuntime.getAuth("openai-codex");
+    }
+    catch (error) {
+        throw new Error("OpenAI Codex OAuth token refresh failed", { cause: error });
+    }
+    const credential = readStoredCredential("openai-codex", getPiAuthPath());
+    const access = resolvedAuth?.auth.apiKey ?? credential?.access;
     if (!access || credential?.type !== "oauth" || !credential.access || isExpired(credential)) {
         throw new Error("OpenAI Codex OAuth token refresh failed");
     }
     return {
         type: "oauth",
         access: credential.access,
-        refresh: credential.refresh,
-        expires: credential.expires,
+        ...(credential.refresh === undefined ? {} : { refresh: credential.refresh }),
+        ...(credential.expires === undefined ? {} : { expires: credential.expires }),
     };
 }
 async function readOpenAIAuth() {

package/dist/app/popup/menu-items-controller.js

@@ -173,16 +173,15 @@ export class AppMenuItemsController {
         return current?.provider === model.provider && current.id === model.id;
     }
     resolveScopedModelRefs(modelRefs) {
-        const registry = this.host.runtime()?.services.modelRegistry;
-        if (!registry)
+        const modelRuntime = this.host.runtime()?.services.modelRuntime;
+        if (!modelRuntime)
             return [];
-        registry.refresh();
         const scopedModels = [];
         for (const modelRef of modelRefs) {
             const parsed = parseScopedModelRef(modelRef);
             if (!parsed)
                 continue;
-            const model = registry.find(parsed.provider, parsed.modelId);
+            const model = modelRuntime.getModel(parsed.provider, parsed.modelId);
             if (!model)
                 continue;
             scopedModels.push({
@@ -197,10 +196,9 @@ export class AppMenuItemsController {
         const scopedModels = session?.scopedModels.length ? session.scopedModels : this.getFavoriteScopedModels();
         if (!scopedModels.length)
             return [];
-        const registry = this.host.runtime()?.services.modelRegistry;
-        registry?.refresh();
+        const modelRuntime = this.host.runtime()?.services.modelRuntime;
         return scopedModels.map((scoped) => {
-            const refreshed = registry?.find(scoped.model.provider, scoped.model.id);
+            const refreshed = modelRuntime?.getModel(scoped.model.provider, scoped.model.id);
             return (refreshed ?? scoped.model);
         });
     }

package/dist/app/runtime.js

@@ -239,8 +239,8 @@ export async function createPixRuntime(options, runtimeOptions = {}) {
                 config,
                 ...(runtimeOptions.eventBus === undefined ? {} : { eventBus: runtimeOptions.eventBus }),
             });
-        services.modelRegistry.refresh();
-        const model = parsedModel ? services.modelRegistry.find(parsedModel.provider, parsedModel.modelId) : undefined;
+        await services.modelRuntime.reloadConfig();
+        const model = parsedModel ? services.modelRuntime.getModel(parsedModel.provider, parsedModel.modelId) : undefined;
         if (parsedModel && !model) {
             throw new Error(`Model not found: ${parsedModel.provider}/${parsedModel.modelId}`);
         }
@@ -250,7 +250,7 @@ export async function createPixRuntime(options, runtimeOptions = {}) {
             const scoped = parseScopedModelRef(modelRef);
             if (!scoped)
                 return [];
-            const scopedModel = services.modelRegistry.find(scoped.provider, scoped.modelId);
+            const scopedModel = services.modelRuntime.getModel(scoped.provider, scoped.modelId);
             if (!scopedModel)
                 return [];
             return [

package/dist/bundled-extensions/session-title/index.d.ts

@@ -1,5 +1,6 @@
 import { type ExtensionAPI, type ExtensionContext } from "@earendil-works/pi-coding-agent";
-export { fallbackSessionTitleFromInput, generateSessionTitle, sessionTitleModelRefs, sanitizeSessionTitle } from "./title-generation.js";
+export { generateSessionTitle } from "./title-generation-compat.js";
+export { fallbackSessionTitleFromInput, generateSessionTitleWithRuntime, sessionTitleModelRefs, sanitizeSessionTitle } from "./title-generation.js";
 export declare function firstUserMessageText(ctx: ExtensionContext): string | undefined;
 export declare function buildForkTitleInput(parentTitle: string | undefined, forkPrompt: string): string;
 export default function sessionTitle(pi: ExtensionAPI): void;

package/dist/bundled-extensions/session-title/index.js

@@ -1,8 +1,10 @@
 import { SessionManager } from "@earendil-works/pi-coding-agent";
 import { resolve } from "node:path";
 import { loadSessionTitleConfig } from "./config.js";
-import { fallbackSessionTitleFromInput, firstUserMessageText as firstUserMessageTextFromEntries, generateSessionTitle, sessionTitleModelRefs, } from "./title-generation.js";
-export { fallbackSessionTitleFromInput, generateSessionTitle, sessionTitleModelRefs, sanitizeSessionTitle } from "./title-generation.js";
+import { fallbackSessionTitleFromInput, firstUserMessageText as firstUserMessageTextFromEntries, sessionTitleModelRefs, } from "./title-generation.js";
+import { generateSessionTitle } from "./title-generation-compat.js";
+export { generateSessionTitle } from "./title-generation-compat.js";
+export { fallbackSessionTitleFromInput, generateSessionTitleWithRuntime, sessionTitleModelRefs, sanitizeSessionTitle } from "./title-generation.js";
 const DEFAULT_TERMINAL_TITLE = "pi";
 function isStaleExtensionContextError(error) {
     if (!(error instanceof Error))

package/dist/bundled-extensions/session-title/title-generation-compat.d.ts

@@ -0,0 +1,17 @@
+import type { Api, Model } from "@earendil-works/pi-ai";
+import type { SessionTitleConfig } from "./config.js";
+type TitleModelRegistry = {
+    find(provider: string, modelId: string): Model<Api> | undefined;
+    getApiKeyAndHeaders(model: Model<Api>): Promise<{
+        ok: true;
+        apiKey?: string;
+        headers?: Record<string, string>;
+        env?: Record<string, string>;
+    } | {
+        ok: false;
+        error: string;
+    }>;
+};
+/** Extension-side title generation through Pi's public ModelRegistry facade. */
+export declare function generateSessionTitle(input: string, modelRegistry: TitleModelRegistry, config: SessionTitleConfig, modelRef: string, signal: AbortSignal, onWarning?: (message: string) => void): Promise<string | undefined>;
+export {};

package/dist/bundled-extensions/session-title/title-generation-compat.js

@@ -0,0 +1,40 @@
+import { complete } from "@earendil-works/pi-ai/compat";
+import { buildTitlePrompt, parseTitleModelRef, sanitizeSessionTitle, TITLE_SYSTEM_PROMPT, titleResponseText, } from "./title-generation.js";
+/** Extension-side title generation through Pi's public ModelRegistry facade. */
+export async function generateSessionTitle(input, modelRegistry, config, modelRef, signal, onWarning) {
+    const parsedModel = parseTitleModelRef(modelRef);
+    if (!parsedModel) {
+        onWarning?.(`Invalid session-title model: ${modelRef}`);
+        return undefined;
+    }
+    const model = modelRegistry.find(parsedModel.provider, parsedModel.modelId);
+    if (!model) {
+        onWarning?.(`Session-title model not found: ${modelRef}`);
+        return undefined;
+    }
+    const auth = await modelRegistry.getApiKeyAndHeaders(model);
+    if (auth.ok === false) {
+        onWarning?.(auth.error);
+        return undefined;
+    }
+    const response = await complete(model, {
+        systemPrompt: TITLE_SYSTEM_PROMPT,
+        messages: [
+            {
+                role: "user",
+                content: [{ type: "text", text: buildTitlePrompt(input, config.maxTitleChars) }],
+                timestamp: Date.now(),
+            },
+        ],
+    }, {
+        ...(auth.apiKey === undefined ? {} : { apiKey: auth.apiKey }),
+        ...(auth.headers === undefined ? {} : { headers: auth.headers }),
+        ...(auth.env === undefined ? {} : { env: auth.env }),
+        cacheRetention: "none",
+        maxRetries: config.maxRetries,
+        maxTokens: config.maxTokens,
+        signal,
+        timeoutMs: config.timeoutMs,
+    });
+    return sanitizeSessionTitle(titleResponseText(response), config.maxTitleChars);
+}

package/dist/bundled-extensions/session-title/title-generation.d.ts

@@ -1,4 +1,4 @@
-import type { Api, Model } from "@earendil-works/pi-ai";
+import type { ModelRuntime } from "@earendil-works/pi-coding-agent";
 import type { SessionTitleConfig } from "./config.js";
 type SessionEntryLike = {
     type?: string;
@@ -7,20 +7,23 @@ type SessionEntryLike = {
         content?: unknown;
     };
 };
-type TitleModelRegistry = {
-    find(provider: string, modelId: string): Model<Api> | undefined;
-    getApiKeyAndHeaders(model: Model<Api>): Promise<{
-        ok: true;
-        apiKey?: string;
-        headers?: Record<string, string>;
-    } | {
-        ok: false;
-        error: string;
-    }>;
-};
+type TitleModelRuntime = Pick<ModelRuntime, "getModel" | "completeSimple">;
+export declare const TITLE_SYSTEM_PROMPT: string;
+export declare function parseTitleModelRef(modelRef: string): {
+    provider: string;
+    modelId: string;
+} | undefined;
 export declare function firstUserMessageText(entries: readonly SessionEntryLike[]): string | undefined;
 export declare function fallbackSessionTitleFromInput(input: string, maxTitleChars: number): string | undefined;
+export declare function buildTitlePrompt(input: string, maxTitleChars: number): string;
+export declare function titleResponseText(response: {
+    content: Array<{
+        type: string;
+        text?: string;
+    }>;
+}): string;
 export declare function sanitizeSessionTitle(raw: string, maxTitleChars: number): string | undefined;
 export declare function sessionTitleModelRefs(config: SessionTitleConfig): string[];
-export declare function generateSessionTitle(input: string, modelRegistry: TitleModelRegistry, config: SessionTitleConfig, modelRef: string, signal: AbortSignal, onWarning?: (message: string) => void): Promise<string | undefined>;
+/** Host-side title generation through the canonical SDK ModelRuntime. */
+export declare function generateSessionTitleWithRuntime(input: string, modelRuntime: TitleModelRuntime, config: SessionTitleConfig, modelRef: string, signal: AbortSignal, onWarning?: (message: string) => void): Promise<string | undefined>;
 export {};

package/dist/bundled-extensions/session-title/title-generation.js

@@ -1,11 +1,10 @@
-import { complete } from "@earendil-works/pi-ai/compat";
-const TITLE_SYSTEM_PROMPT = [
+export const TITLE_SYSTEM_PROMPT = [
     "You name Pi coding-agent sessions from the user's first request.",
     "Return only one concise title, without quotes, markdown, emoji, or explanations.",
     "Use the same language as the task when possible.",
     "Keep it specific, 3-7 words, and under the requested character limit.",
 ].join("\n");
-function parseModelRef(modelRef) {
+export function parseTitleModelRef(modelRef) {
     const trimmed = modelRef.trim();
     const slash = trimmed.indexOf("/");
     if (slash <= 0 || slash === trimmed.length - 1)
@@ -58,7 +57,7 @@ export function fallbackSessionTitleFromInput(input, maxTitleChars) {
     const candidate = selected.join(" ");
     return sanitizeSessionTitle(candidate || normalized, maxTitleChars);
 }
-function buildTitlePrompt(input, maxTitleChars) {
+export function buildTitlePrompt(input, maxTitleChars) {
     return [
         `Generate a session title under ${maxTitleChars} characters for this session.`,
         "If a parent session title is provided, use it only as context and focus on the new request.",
@@ -69,7 +68,7 @@ function buildTitlePrompt(input, maxTitleChars) {
         "</session_context>",
     ].join("\n");
 }
-function responseText(response) {
+export function titleResponseText(response) {
     return response.content
         .filter((block) => block.type === "text" && typeof block.text === "string")
         .map((block) => block.text)
@@ -103,23 +102,19 @@ export function sessionTitleModelRefs(config) {
         .filter(Boolean)
         .filter((modelRef, index, refs) => refs.indexOf(modelRef) === index);
 }
-export async function generateSessionTitle(input, modelRegistry, config, modelRef, signal, onWarning) {
-    const parsedModel = parseModelRef(modelRef);
+/** Host-side title generation through the canonical SDK ModelRuntime. */
+export async function generateSessionTitleWithRuntime(input, modelRuntime, config, modelRef, signal, onWarning) {
+    const parsedModel = parseTitleModelRef(modelRef);
     if (!parsedModel) {
         onWarning?.(`Invalid session-title model: ${modelRef}`);
         return undefined;
     }
-    const model = modelRegistry.find(parsedModel.provider, parsedModel.modelId);
+    const model = modelRuntime.getModel(parsedModel.provider, parsedModel.modelId);
     if (!model) {
         onWarning?.(`Session-title model not found: ${modelRef}`);
         return undefined;
     }
-    const auth = await modelRegistry.getApiKeyAndHeaders(model);
-    if (auth.ok === false) {
-        onWarning?.(auth.error);
-        return undefined;
-    }
-    const response = await complete(model, {
+    const response = await modelRuntime.completeSimple(model, {
         systemPrompt: TITLE_SYSTEM_PROMPT,
         messages: [
             {
@@ -129,13 +124,11 @@ export async function generateSessionTitle(input, modelRegistry, config, modelRe
             },
         ],
     }, {
-        ...(auth.apiKey === undefined ? {} : { apiKey: auth.apiKey }),
-        ...(auth.headers === undefined ? {} : { headers: auth.headers }),
         cacheRetention: "none",
         maxRetries: config.maxRetries,
         maxTokens: config.maxTokens,
         signal,
         timeoutMs: config.timeoutMs,
     });
-    return sanitizeSessionTitle(responseText(response), config.maxTitleChars);
+    return sanitizeSessionTitle(titleResponseText(response), config.maxTitleChars);
 }

package/external/pi-tools-suite/README.md

@@ -362,6 +362,22 @@ npm run test:async-subagents-selection-e2e
 npm run test:e2e
 ```
 
+### Prompt evaluations
+
+Prompt evaluations are opt-in because they call a real model. They cover model-facing behavior that deterministic tests cannot prove: tool selection for `todo` and `compress`, async-subagent delegation/lifecycle boundaries, default internal role routing, ultrawork classification, and DCP summary retention. They are intentionally excluded from `npm test`.
+
+```bash
+# Full prompt-eval suite
+npm run test:prompt-evals
+
+# Focused suites
+npm run test:prompt-evals:tool-selection
+npm run test:prompt-evals:async
+npm run test:prompt-evals:dcp
+```
+
+The default live model is `zai/glm-5-turbo`. Override it for the whole suite with `PI_TOOLS_SUITE_E2E_MODEL=provider/model`, or use the existing component variables such as `TOOL_SELECTION_E2E_MODEL`, `ASYNC_SUBAGENTS_MODEL`, `ASYNC_SUBAGENTS_ROUTING_E2E_MODEL`, and `DCP_SUMMARY_E2E_MODEL`. The normal deterministic coverage remains `npm test`; run prompt evals after changing tool descriptions, routing/classifier prompts, DCP summary prompts, or the default evaluation model.
+
 Supporting docs and historical standalone README content are kept in `docs/`; third-party license texts are kept in `licenses/`.
 
 ## SDK pin

package/external/pi-tools-suite/package.json

@@ -24,13 +24,18 @@
     "test": "bun test test",
     "test:async-subagents-e2e": "ASYNC_SUBAGENTS_E2E=1 ASYNC_SUBAGENTS_DEBUG_LOGS=1 ASYNC_SUBAGENTS_MODEL=zai/glm-5-turbo bun test --concurrent --max-concurrency=30 test/async-subagents",
     "test:async-subagents-selection-e2e": "ASYNC_SUBAGENTS_SELECTION_E2E=1 ASYNC_SUBAGENTS_MODEL=zai/glm-5-turbo bun test --concurrent --max-concurrency=30 test/async-subagents/selection-e2e.test.ts",
+    "test:prompt-evals:tool-selection": "PROMPT_EVAL_E2E=1 bun test --concurrent --max-concurrency=10 test/tool-selection-e2e.test.ts",
+    "test:prompt-evals:async": "PROMPT_EVAL_E2E=1 bun test --concurrent --max-concurrency=5 test/async-subagents/selection-e2e.test.ts test/prompt-evals/async-routing-e2e.test.ts",
+    "test:prompt-evals:dcp": "PROMPT_EVAL_E2E=1 bun test --concurrent --max-concurrency=5 test/prompt-evals/dcp-summary-e2e.test.ts",
+    "test:prompt-evals": "PROMPT_EVAL_E2E=1 bun test --concurrent --max-concurrency=5 test/tool-selection-e2e.test.ts test/async-subagents/selection-e2e.test.ts test/prompt-evals",
     "bench:locate": "PI_LOCATE_BENCH_ITERATIONS=5 PI_LOCATE_BENCH_FAKE_IDX=0 PI_LOCATE_BENCH_MODEL=zai/glm-5-turbo PI_LOCATE_BENCH_MODES=direct-read-grep,ast-structural,repo-search-hybrid,repo-discovery,subagent-search,unrestricted-suite node test/fixtures/hard-to-find-project/benchmark/run-locate-benchmark.mjs",
     "bench:locate:analyze": "node test/fixtures/hard-to-find-project/benchmark/analyze-locate-benchmark.mjs",
     "test:locate-benchmark-e2e": "PI_LOCATE_BENCH_E2E=1 PI_LOCATE_BENCH_MODEL=zai/glm-5-turbo bun test test/locate-benchmark-e2e.test.ts",
     "test:tool-selection-e2e": "TOOL_SELECTION_E2E=1 TOOL_SELECTION_E2E_MODEL=zai/glm-5-turbo bun test --concurrent --max-concurrency=30 test/tool-selection-e2e.test.ts",
     "test:e2e": "TOOL_SELECTION_E2E=1 TOOL_SELECTION_E2E_MODEL=zai/glm-5-turbo ASYNC_SUBAGENTS_E2E=1 ASYNC_SUBAGENTS_DEBUG_LOGS=1 ASYNC_SUBAGENTS_MODEL=zai/glm-5-turbo bun test --concurrent --max-concurrency=5 test/tool-selection-e2e.test.ts test/async-subagents/e2e.test.ts test/async-subagents/selection-e2e.test.ts test/todo-persistence-e2e.test.ts",
+    "typecheck": "tsc -p tsconfig.source.json",
     "typecheck:async-subagents": "bash scripts/typecheck-source.sh",
-    "check": "npm run smoke"
+    "check": "npm run typecheck && npm test && npm run smoke"
   },
   "dependencies": {
     "jsonc-parser": "^3.3.1",
@@ -38,9 +43,9 @@
     "vscode-languageserver-protocol": "^3.17.5"
   },
   "peerDependencies": {
-    "@earendil-works/pi-ai": "0.80.7",
-    "@earendil-works/pi-coding-agent": "0.80.7",
-    "@earendil-works/pi-tui": "0.80.7",
+    "@earendil-works/pi-ai": "0.80.9",
+    "@earendil-works/pi-coding-agent": "0.80.9",
+    "@earendil-works/pi-tui": "0.80.9",
     "typebox": "*"
   },
   "devDependencies": {

package/external/pi-tools-suite/src/async-subagents/core/routing.ts

@@ -1,5 +1,5 @@
-import { complete } from "@earendil-works/pi-ai/compat";
 import type { Api, Model } from "@earendil-works/pi-ai";
+import { completeWithModelRegistry, type ModelCompletionRegistry } from "../../model-completion.js";
 import type { AgentTask } from "./types.js";
 import {
 	currentModelRef,
@@ -11,10 +11,10 @@ import {
 
 export interface SubagentRoutingContext {
 	model?: unknown;
-	modelRegistry?: {
+	modelRegistry?: ModelCompletionRegistry & {
 		find(provider: string, modelId: string): Model<Api> | undefined;
 		getApiKeyAndHeaders(model: Model<Api>): Promise<
-			| { ok?: true; apiKey?: string; headers?: Record<string, string> }
+			| { ok?: true; apiKey?: string; headers?: Record<string, string>; env?: Record<string, string> }
 			| { ok: false; error: string }
 		>;
 	};
@@ -67,7 +67,8 @@ export async function routeSubagentTasks(
 		for (const candidate of candidates) {
 			if (signal?.aborted) throw new Error("Aborted");
 			try {
-				response = await complete(
+				response = await completeWithModelRegistry(
+					ctx.modelRegistry,
 					candidate.model,
 					{
 						systemPrompt: ROUTER_SYSTEM_PROMPT,
@@ -82,6 +83,7 @@ export async function routeSubagentTasks(
 					{
 						apiKey: candidate.apiKey,
 						headers: candidate.headers,
+						env: candidate.env,
 						cacheRetention: "none",
 						maxRetries: routing.maxRetries,
 						maxTokens: routing.maxTokens,
@@ -179,14 +181,16 @@ interface RoutingCandidate {
 	model: Model<Api>;
 	apiKey?: string;
 	headers?: Record<string, string>;
+	env?: Record<string, string>;
 }
 
-type RoutingResponse = Awaited<ReturnType<typeof complete>>;
+type RoutingResponse = Awaited<ReturnType<typeof completeWithModelRegistry>>;
 
 async function resolveModelRef(ctx: SubagentRoutingContext, modelRef: string): Promise<{
 	model: Model<Api>;
 	apiKey?: string;
 	headers?: Record<string, string>;
+	env?: Record<string, string>;
 } | undefined> {
 	const parsed = parseModelRef(modelRef);
 	if (!parsed || !ctx.modelRegistry) return undefined;
@@ -194,7 +198,7 @@ async function resolveModelRef(ctx: SubagentRoutingContext, modelRef: string): P
 	if (!model) return undefined;
 	const auth = await ctx.modelRegistry.getApiKeyAndHeaders(model);
 	if (auth.ok === false) return undefined;
-	return { model, apiKey: auth.apiKey, headers: auth.headers };
+	return { model, apiKey: auth.apiKey, headers: auth.headers, env: auth.env };
 }
 
 function parseModelRef(modelRef: string): { provider: string; modelId: string } | undefined {

package/external/pi-tools-suite/src/async-subagents/core/ultrawork-auto.ts

@@ -1,15 +1,15 @@
-import { complete } from "@earendil-works/pi-ai/compat";
 import type { Api, Model } from "@earendil-works/pi-ai";
+import { completeWithModelRegistry, type ModelCompletionRegistry } from "../../model-completion.js";
 import { currentModelRef, resolveSubagentRoutingConfig, type SubagentConfig } from "./config.js";
 
 export type UltraworkAutoDecision = "ultrawork" | "hint" | "none";
 
 export interface UltraworkAutoContext {
 	model?: unknown;
-	modelRegistry?: {
+	modelRegistry?: ModelCompletionRegistry & {
 		find(provider: string, modelId: string): Model<Api> | undefined;
 		getApiKeyAndHeaders(model: Model<Api>): Promise<
-			| { ok?: true; apiKey?: string; headers?: Record<string, string> }
+			| { ok?: true; apiKey?: string; headers?: Record<string, string>; env?: Record<string, string> }
 			| { ok: false; error: string }
 		>;
 	};
@@ -63,7 +63,8 @@ export async function decideUltraworkAuto(
 		const resolved = await resolveClassifierModel(ctx, routing.model);
 		if (!resolved) return "none";
 
-		const response = await complete(
+		const response = await completeWithModelRegistry(
+			ctx.modelRegistry,
 			resolved.model,
 			{
 				systemPrompt: CLASSIFIER_SYSTEM_PROMPT,
@@ -78,6 +79,7 @@ export async function decideUltraworkAuto(
 			{
 				apiKey: resolved.apiKey,
 				headers: resolved.headers,
+				env: resolved.env,
 				cacheRetention: "none",
 				maxRetries: routing.maxRetries,
 				maxTokens: Math.min(routing.maxTokens, 32),
@@ -130,6 +132,7 @@ async function resolveClassifierModel(ctx: UltraworkAutoContext, modelRef: strin
 	model: Model<Api>;
 	apiKey?: string;
 	headers?: Record<string, string>;
+	env?: Record<string, string>;
 } | undefined> {
 	const configured = await resolveModelRef(ctx, modelRef);
 	if (configured) return configured;
@@ -141,6 +144,7 @@ async function resolveModelRef(ctx: UltraworkAutoContext, modelRef: string): Pro
 	model: Model<Api>;
 	apiKey?: string;
 	headers?: Record<string, string>;
+	env?: Record<string, string>;
 } | undefined> {
 	const parsed = parseModelRef(modelRef);
 	if (!parsed || !ctx.modelRegistry) return undefined;
@@ -148,7 +152,7 @@ async function resolveModelRef(ctx: UltraworkAutoContext, modelRef: string): Pro
 	if (!model) return undefined;
 	const auth = await ctx.modelRegistry.getApiKeyAndHeaders(model);
 	if (auth.ok === false) return undefined;
-	return { model, apiKey: auth.apiKey, headers: auth.headers };
+	return { model, apiKey: auth.apiKey, headers: auth.headers, env: auth.env };
 }
 
 function parseModelRef(modelRef: string): { provider: string; modelId: string } | undefined {

package/external/pi-tools-suite/src/coding-discipline/index.ts

@@ -1,11 +1,11 @@
 import * as fs from "node:fs";
 import * as path from "node:path";
-import { complete } from "@earendil-works/pi-ai/compat";
 import type { Api, AssistantMessage, ImageContent, Model, TextContent } from "@earendil-works/pi-ai";
 import { Type } from "typebox";
 
 import { loadPiToolsSuiteConfig } from "../config.js";
 import { ignoreStaleExtensionContextError } from "../context-usage.js";
+import { completeWithModelRegistry, type ModelCompletionRegistry } from "../model-completion.js";
 
 type ExtensionAPI = any;
 
@@ -25,9 +25,11 @@ type LookupDetails = {
 };
 
 type ResolvedLookupModel = {
+	modelRegistry: ModelCompletionRegistry;
 	model: Model<Api>;
 	apiKey?: string;
 	headers?: Record<string, string>;
+	env?: Record<string, string>;
 };
 
 const SILENT_PROMPT_MARKER_START = "<glm_silent_mode>";
@@ -504,7 +506,8 @@ function createLookupTool() {
 			const promptText = buildLookupPrompt(params, recentContext, images.length, pathImages.warnings);
 
 			try {
-				const response = await complete(
+				const response = await completeWithModelRegistry(
+					resolved.modelRegistry,
 					resolved.model,
 					{
 						systemPrompt: LOOKUP_SYSTEM_PROMPT,
@@ -519,6 +522,7 @@ function createLookupTool() {
 					{
 						apiKey: resolved.apiKey,
 						headers: resolved.headers,
+						env: resolved.env,
 						cacheRetention: "none",
 						maxRetries: 1,
 						maxTokens: DEFAULT_LOOKUP_MAX_TOKENS,
@@ -765,9 +769,20 @@ async function resolveLookupModel(ctx: unknown, modelRef: string): Promise<Resol
 	if (!isRecord(registry) || typeof registry.find !== "function" || typeof registry.getApiKeyAndHeaders !== "function") return undefined;
 	const model = registry.find(parsed.provider, parsed.modelId) as Model<Api> | undefined;
 	if (!model) return undefined;
-	const auth = await registry.getApiKeyAndHeaders(model) as { ok?: true; apiKey?: string; headers?: Record<string, string> } | { ok: false; error: string };
+	const auth = await registry.getApiKeyAndHeaders(model) as {
+		ok?: true;
+		apiKey?: string;
+		headers?: Record<string, string>;
+		env?: Record<string, string>;
+	} | { ok: false; error: string };
 	if (auth.ok === false) return undefined;
-	return { model, apiKey: auth.apiKey, headers: auth.headers };
+	return {
+		modelRegistry: registry as ModelCompletionRegistry,
+		model,
+		apiKey: auth.apiKey,
+		headers: auth.headers,
+		env: auth.env,
+	};
 }
 
 function parseModelRef(modelRef: string): { provider: string; modelId: string } | undefined {

package/external/pi-tools-suite/src/dcp/auto-compress.ts

@@ -13,8 +13,8 @@
 // automatic fallback to the programmatic digest on any failure/timeout.
 // ---------------------------------------------------------------------------
 
-import { complete } from "@earendil-works/pi-ai/compat"
 import type { Model, Api } from "@earendil-works/pi-ai"
+import { completeWithModelRegistry, type ModelCompletionRegistry } from "../model-completion.js"
 import type { DcpState } from "./state.js"
 import type { DcpConfig } from "./config.js"
 import type { CompressionCandidate } from "./pruner-types.js"
@@ -116,7 +116,7 @@ export function buildProgrammaticSummary(
 	return lines.join("\n")
 }
 
-const SUMMARIZER_SYSTEM_PROMPT = `You summarize a slice of a coding agent's conversation so it can replace the raw messages in context. Produce a dense, continuation-focused summary: preserve user intent, decisions made, files/symbols changed or inspected, exact errors still actionable, verification status, and next steps. Drop full logs, repeated output, and incidental detail. Be concise (roughly 4-10 bullets). Output ONLY the summary text, no preamble.`
+const SUMMARIZER_SYSTEM_PROMPT = `You summarize a slice of a coding agent's conversation so it can replace the raw messages in context. Produce a dense, continuation-focused summary: preserve user intent, decisions made, files/symbols changed or inspected, exact errors still actionable, verification status, and next steps. Do not infer, invent, or add facts absent from the source; preserve uncertainty instead of filling gaps. Drop full logs, repeated output, and incidental detail. Be concise (roughly 4-10 bullets). Output ONLY the summary text, no preamble.`
 
 /** Outcome of one summarizer-model attempt, surfaced in DCP debug logs. */
 export interface ModelSummaryAttempt {
@@ -134,6 +134,14 @@ export interface ModelSummaryResult {
 	attempts: ModelSummaryAttempt[]
 }
 
+type ModelSummaryRegistry = ModelCompletionRegistry & {
+	find(provider: string, modelId: string): Model<Api> | undefined
+	getApiKeyAndHeaders(model: Model<Api>): Promise<
+		| { ok: true; apiKey?: string; headers?: Record<string, string>; env?: Record<string, string> }
+		| { ok: false; error: string }
+	>
+}
+
 /**
  * Try to produce a model-generated summary by calling each model in
  * `modelRefs` in order. On success returns `{ text, usedModelRef, attempts }`;
@@ -146,7 +154,7 @@ export interface ModelSummaryResult {
  */
 export async function generateModelSummary(
 	modelRefs: string[],
-	modelRegistry: any,
+	modelRegistry: ModelSummaryRegistry | undefined,
 	signal: AbortSignal | undefined,
 	topic: string,
 	messagesInRange: any[],
@@ -178,7 +186,7 @@ export async function generateModelSummary(
 			continue
 		}
 
-		let auth
+		let auth: Awaited<ReturnType<ModelSummaryRegistry["getApiKeyAndHeaders"]>>
 		try {
 			auth = await modelRegistry.getApiKeyAndHeaders(model)
 		} catch (error) {
@@ -186,7 +194,7 @@ export async function generateModelSummary(
 			attempts.push({ ref, outcome: "no-auth", error: error instanceof Error ? error.message : String(error) })
 			continue
 		}
-		if (!auth?.ok || !auth.apiKey) {
+		if (auth.ok === false) {
 			attempts.push({ ref, outcome: "no-auth" })
 			continue
 		}
@@ -202,7 +210,8 @@ export async function generateModelSummary(
 		}
 
 		try {
-			const result = await complete(
+			const result = await completeWithModelRegistry(
+				modelRegistry,
 				model,
 				{ systemPrompt: SUMMARIZER_SYSTEM_PROMPT, messages: [{ role: "user", content: userPrompt, timestamp: Date.now() }] },
 				{

package/external/pi-tools-suite/src/dcp/commands.ts

@@ -109,9 +109,10 @@ function collectNudgeStats(ctx: ExtensionCommandContext, state: DcpState): DcpNu
       else stats.upgraded++
       stats.byType[data.type]++
       const createdAt = typeof data.createdAt === "number" ? data.createdAt : undefined
-      const contextPercent = typeof data.contextPercent === "number" || data.contextPercent === null
-        ? data.contextPercent
-        : undefined
+      const rawContextPercent = data.contextPercent
+      let contextPercent: number | null | undefined
+      if (typeof rawContextPercent === "number") contextPercent = rawContextPercent
+      else if (rawContextPercent === null) contextPercent = null
       if (!stats.last || (createdAt ?? 0) >= (stats.last.createdAt ?? 0)) {
         stats.last = { type: data.type, event, createdAt, contextPercent }
       }

package/external/pi-tools-suite/src/dcp/prompts.ts

@@ -25,7 +25,7 @@ Do not compress active work, still-needed raw context, or material whose exact c
 
 DCP reminders: handle critical/high-context reminders promptly and compress any safe high-yield closed slice before more exploration; routine reminders mean compress only if a safe, closed, useful slice exists, otherwise continue the next atomic step and re-check later.
 
-Summaries must preserve only what is needed to continue: user intent and constraints, accepted decisions, files/symbols changed or inspected, actionable errors, verification status, and next steps. Drop incidental transcript detail, duplicate outputs, full logs, long code/JSON/diffs, and prose not needed later; include short literals only when required.
+Summaries must preserve only what is needed to continue: user intent and constraints, accepted decisions, files/symbols changed or inspected, actionable errors, verification status, and next steps. Do not infer, invent, or add facts that are not present in the selected messages. Drop incidental transcript detail, duplicate outputs, full logs, long code/JSON/diffs, and prose not needed later; include short literals only when required.
 `.trim()
 
 /**
@@ -55,6 +55,8 @@ If a \`<dcp-system-reminder>\` is present in context, treat it as a signal to ev
 THE SUMMARY
 Your summary must be COMPLETE FOR CONTINUATION, not a transcript rewrite. Preserve only information that will plausibly matter later: user intent, accepted constraints, decisions, files/symbols changed or inspected, exact errors that are still actionable, verification status, and next steps.
 
+Do not infer, invent, or add facts that are not present in the selected range. If the source is uncertain or incomplete, preserve that uncertainty instead of filling gaps.
+
 If active unfinished work exists, start with \`Active objective\` and \`Next step\`.
 
 Default to a compact structured summary (roughly 4-10 bullets for a normal completed work slice). Grow beyond that only when the compressed range contains multiple independent decisions, unresolved blockers, or precise state that is genuinely required to continue.

package/external/pi-tools-suite/src/model-completion.ts

@@ -0,0 +1,37 @@
+import type {
+	Api,
+	AssistantMessage,
+	AssistantMessageEventStream,
+	Context,
+	Model,
+	SimpleStreamOptions,
+} from "@earendil-works/pi-ai";
+import { completeSimple } from "@earendil-works/pi-ai/compat";
+
+type RegisteredProviderConfig = {
+	streamSimple?: (
+		model: Model<Api>,
+		context: Context,
+		options?: SimpleStreamOptions,
+	) => AssistantMessageEventStream;
+};
+
+export type ModelCompletionRegistry = {
+	getRegisteredProviderConfig?(providerId: string): RegisteredProviderConfig | undefined;
+};
+
+/**
+ * Complete through an extension provider's registered stream when available.
+ * Pi 0.80.8+ no longer copies extension streams into pi-ai's global compat
+ * registry, so falling back to compat is valid only for built-in APIs.
+ */
+export async function completeWithModelRegistry(
+	modelRegistry: ModelCompletionRegistry | undefined,
+	model: Model<Api>,
+	context: Context,
+	options?: SimpleStreamOptions,
+): Promise<AssistantMessage> {
+	const providerConfig = modelRegistry?.getRegisteredProviderConfig?.(model.provider);
+	if (providerConfig?.streamSimple) return providerConfig.streamSimple(model, context, options).result();
+	return completeSimple(model, context, options);
+}

package/external/pi-tools-suite/src/tool-descriptions.ts

@@ -220,6 +220,7 @@ export const TODO_TOOL_DESCRIPTION: ToolDescription = {
 	promptGuidelines: [
 		"Use `todo` for complex work with 3+ steps, explicit user task lists, or new non-trivial requirements. Skip single trivial tasks and purely conversational requests.",
 		"For multi-step implementation/debugging plans, include a final user-facing report todo in the initial plan with acceptance criteria for changed files/behavior, verification results, and remaining manual actions; close it immediately before the final response, never via compression.",
+		"When create or batch_create already sets the intended status, do not issue a redundant update with the same status; continue the work instead.",
 		"Resync before continuing when user/new findings change scope, requirements, safety, feasibility, approach, dependencies, or order; update tasks/blockers and defer obsolete work.",
 		"Update todos when starting, finishing, blocking, splitting, abandoning, or materially changing a step; before planned work mark exactly one in_progress with activeForm and complete it only after verification.",
 		"If partial, tests fail, or blocked, keep the task in_progress and add/update a blocker. Never use `clear`, `delete`, or batch deletion to hide unfinished/stale/forgotten todos; delete only on explicit request or creation mistake.",

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "pi-ui-extend",
-  "version": "0.1.60",
+  "version": "0.1.63",
   "private": false,
   "type": "module",
   "bin": {
@@ -65,9 +65,9 @@
     "prepublishOnly": "npm run check && npm run build:pix && npm run generate-schemas"
   },
   "dependencies": {
-    "@earendil-works/pi-ai": "0.80.7",
-    "@earendil-works/pi-coding-agent": "0.80.7",
-    "@earendil-works/pi-tui": "0.80.7",
+    "@earendil-works/pi-ai": "0.80.9",
+    "@earendil-works/pi-coding-agent": "0.80.9",
+    "@earendil-works/pi-tui": "0.80.9",
     "@mariozechner/clipboard": "^0.3.9",
     "jsonc-parser": "3.3.1",
     "typebox": "1.1.38",