pi-ui-extend 1.0.5 → 1.0.7

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

@@ -1,17 +1,6 @@
-import type { Api, Model } from "@earendil-works/pi-ai";
+import type { ModelRegistry } from "@earendil-works/pi-coding-agent";
 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;
-    }>;
-};
+type TitleModelRegistry = Pick<ModelRegistry, "complete" | "find">;
 /** 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

@@ -1,4 +1,3 @@
-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) {
@@ -12,12 +11,7 @@ export async function generateSessionTitle(input, modelRegistry, config, modelRe
         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 modelRegistry.complete(model, {
         systemPrompt: TITLE_SYSTEM_PROMPT,
         messages: [
             {
@@ -27,9 +21,6 @@ export async function generateSessionTitle(input, modelRegistry, config, modelRe
             },
         ],
     }, {
-        ...(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,

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

@@ -43,9 +43,9 @@
     "vscode-languageserver-protocol": "^3.17.5"
   },
   "peerDependencies": {
-    "@earendil-works/pi-ai": "0.83.0",
-    "@earendil-works/pi-coding-agent": "0.83.0",
-    "@earendil-works/pi-tui": "0.83.0",
+    "@earendil-works/pi-ai": "0.84.1",
+    "@earendil-works/pi-coding-agent": "0.84.1",
+    "@earendil-works/pi-tui": "0.84.1",
     "typebox": "*"
   },
   "devDependencies": {

package/external/pi-tools-suite/src/antigravity-auth/oauth.ts

@@ -221,7 +221,11 @@ export async function addAntigravityAccount(
 	};
 }
 
-async function refreshAccountToken(account: OpencodeAntigravityAccount, oauthClient?: GoogleOAuthClientCredentials): Promise<RefreshedAntigravityAccount> {
+async function refreshAccountToken(
+	account: OpencodeAntigravityAccount,
+	oauthClient?: GoogleOAuthClientCredentials,
+	signal?: AbortSignal,
+): Promise<RefreshedAntigravityAccount> {
 	const refreshToken = getAccountRefreshToken(account);
 	if (!refreshToken) throw new Error(`Missing refresh token for Antigravity account ${account.email ?? "<unknown>"}`);
 	const clientCredentials = getGoogleOAuthClientCredentials(account, oauthClient);
@@ -230,6 +234,7 @@ async function refreshAccountToken(account: OpencodeAntigravityAccount, oauthCli
 	const start = Date.now();
 	const response = await fetch("https://oauth2.googleapis.com/token", {
 		method: "POST",
+		signal,
 		headers: {
 			"Content-Type": "application/x-www-form-urlencoded;charset=UTF-8",
 			Accept: "*/*",
@@ -256,7 +261,7 @@ async function refreshAccountToken(account: OpencodeAntigravityAccount, oauthCli
 	};
 }
 
-export async function refreshAntigravityToken(credentials: OAuthCredentials): Promise<OAuthCredentials> {
+export async function refreshAntigravityToken(credentials: OAuthCredentials, signal: AbortSignal): Promise<OAuthCredentials> {
 	const credentialDetails = credentials as OAuthCredentials & PiAuthCredential;
 	const oauthClient = getGoogleOAuthClientCredentials(credentialDetails);
 	const storedAccounts = getStoredAccounts(credentialDetails);
@@ -272,6 +277,7 @@ export async function refreshAntigravityToken(credentials: OAuthCredentials): Pr
 			email: credentialDetails.email,
 		},
 		oauthClient,
+		signal,
 	);
 	return {
 		...refreshed.credentials,

package/external/pi-tools-suite/src/antigravity-auth/stream.ts

@@ -68,17 +68,21 @@ async function sendAntigravityRequest(
 	const headerStyle = getModelHeaderStyle(model);
 	const endpoints = headerStyle === "gemini-cli" ? [ENDPOINT_PROD] : STREAM_ENDPOINTS;
 	for (const endpoint of endpoints) {
+		const headers = new Headers({
+			Authorization: `Bearer ${apiKey}`,
+			"Content-Type": "application/json",
+			Accept: "text/event-stream",
+			...getAntigravityHeaders(headerStyle),
+			...requestHeaders,
+		});
+		for (const [name, value] of Object.entries(options?.headers ?? {})) {
+			if (value === null) headers.delete(name);
+			else headers.set(name, value);
+		}
 		response = await fetch(`${endpoint}/v1internal:streamGenerateContent?alt=sse`, {
 			method: "POST",
 			signal: options?.signal,
-			headers: {
-				Authorization: `Bearer ${apiKey}`,
-				"Content-Type": "application/json",
-				Accept: "text/event-stream",
-				...getAntigravityHeaders(headerStyle),
-				...requestHeaders,
-				...(options?.headers ?? {}),
-			},
+			headers,
 			body: JSON.stringify(payload),
 		});
 		await options?.onResponse?.({ status: response.status, headers: Object.fromEntries(response.headers.entries()) }, model);

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

@@ -1,4 +1,4 @@
-import type { Api, Model } from "@earendil-works/pi-ai";
+import type { Api, Model, ProviderHeaders } from "@earendil-works/pi-ai";
 import { completeWithModelRegistry, type ModelCompletionRegistry } from "../../model-completion.js";
 import type { AgentTask } from "./types.js";
 import {
@@ -14,7 +14,7 @@ export interface SubagentRoutingContext {
 	modelRegistry?: 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?: true; apiKey?: string; headers?: ProviderHeaders; baseUrl?: string; env?: Record<string, string> }
 			| { ok: false; error: string }
 		>;
 	};
@@ -180,7 +180,7 @@ async function resolveRoutingModels(
 interface RoutingCandidate {
 	model: Model<Api>;
 	apiKey?: string;
-	headers?: Record<string, string>;
+	headers?: ProviderHeaders;
 	env?: Record<string, string>;
 }
 
@@ -189,7 +189,7 @@ type RoutingResponse = Awaited<ReturnType<typeof completeWithModelRegistry>>;
 async function resolveModelRef(ctx: SubagentRoutingContext, modelRef: string): Promise<{
 	model: Model<Api>;
 	apiKey?: string;
-	headers?: Record<string, string>;
+	headers?: ProviderHeaders;
 	env?: Record<string, string>;
 } | undefined> {
 	const parsed = parseModelRef(modelRef);

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

@@ -1,4 +1,4 @@
-import type { Api, Model } from "@earendil-works/pi-ai";
+import type { Api, Model, ProviderHeaders } from "@earendil-works/pi-ai";
 import { completeWithModelRegistry, type ModelCompletionRegistry } from "../../model-completion.js";
 import { currentModelRef, resolveSubagentRoutingConfig, type SubagentConfig } from "./config.js";
 
@@ -9,7 +9,7 @@ export interface UltraworkAutoContext {
 	modelRegistry?: 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?: true; apiKey?: string; headers?: ProviderHeaders; baseUrl?: string; env?: Record<string, string> }
 			| { ok: false; error: string }
 		>;
 	};
@@ -131,7 +131,7 @@ function buildClassifierPrompt(userText: string): string {
 async function resolveClassifierModel(ctx: UltraworkAutoContext, modelRef: string): Promise<{
 	model: Model<Api>;
 	apiKey?: string;
-	headers?: Record<string, string>;
+	headers?: ProviderHeaders;
 	env?: Record<string, string>;
 } | undefined> {
 	const configured = await resolveModelRef(ctx, modelRef);
@@ -143,7 +143,7 @@ async function resolveClassifierModel(ctx: UltraworkAutoContext, modelRef: strin
 async function resolveModelRef(ctx: UltraworkAutoContext, modelRef: string): Promise<{
 	model: Model<Api>;
 	apiKey?: string;
-	headers?: Record<string, string>;
+	headers?: ProviderHeaders;
 	env?: Record<string, string>;
 } | undefined> {
 	const parsed = parseModelRef(modelRef);

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

@@ -1,9 +1,9 @@
 import * as fs from "node:fs";
 import * as path from "node:path";
-import type { Api, AssistantMessage, ImageContent, Model, TextContent } from "@earendil-works/pi-ai";
+import type { Api, AssistantMessage, ImageContent, Model, ProviderHeaders, TextContent } from "@earendil-works/pi-ai";
 import { Type } from "typebox";
 
-import { loadPiToolsSuiteConfig } from "../config.js";
+import { loadPiToolsSuiteConfig, DEFAULT_CODING_DISCIPLINE_STRICTNESS, type CodingDisciplineStrictness } from "../config.js";
 import { ignoreStaleExtensionContextError } from "../context-usage.js";
 import { completeWithModelRegistry, type ModelCompletionRegistry } from "../model-completion.js";
 
@@ -28,7 +28,7 @@ type ResolvedLookupModel = {
 	modelRegistry: ModelCompletionRegistry;
 	model: Model<Api>;
 	apiKey?: string;
-	headers?: Record<string, string>;
+	headers?: ProviderHeaders;
 	env?: Record<string, string>;
 };
 
@@ -43,6 +43,9 @@ const DEFAULT_LOOKUP_TIMEOUT_MS = 120_000;
 const MAX_IMAGE_BYTES = 16 * 1024 * 1024;
 const SILENCE_REMINDER_MIN_VIOLATION_GAP = 3;
 const SILENCE_REMINDER_MIN_MESSAGE_GAP = 12;
+// When the visible history shrinks by more than this many messages (compaction or
+// truncation), the chatter baseline is reset so it isn't measured against a stale peak.
+const SILENCE_REMINDER_COMPACTION_MARGIN = 8;
 const LOOKUP_TOOL_NAME = "lookup";
 
 const LOOKUP_TOOL_PARAMS = Type.Object(
@@ -60,31 +63,46 @@ const LOOKUP_TOOL_PARAMS = Type.Object(
 	{ additionalProperties: false },
 );
 
-const QUALITY_DISCIPLINE_LINES = [
-	"TOOL-ONLY CODING AGENT CONTRACT.",
-	"",
-	"Follow this output-channel contract literally for coding tasks.",
-	"WORKING STATE: from the first assistant turn until complete, final, or blocked, emit exactly one tool call with empty text.",
-	"No prose, markdown, acknowledgments, plans, progress, summaries, explanations, or commentary before/after tool calls.",
-	"Keep reasoning, planning, hypotheses, and retry decisions internal; if a safe next local action exists, call the tool silently.",
-	"",
-	"After every tool result, choose exactly one transition:",
-	"- WORKING: exactly one next tool call with empty text;",
-	"- FINAL: one final answer after completion or practical verification;",
-	"- BLOCKED: one concise question only when no safe/useful tool action can continue.",
-	"No transition permits commentary between tool calls.",
-	"",
-	"PRIORITY: this overrides default assistant friendliness and conversational behavior.",
-	"",
-	"Coding discipline (express only through tool choices, not prose):",
-	"- inspect before editing; do not invent APIs, files, commands, or behavior;",
-	"- make the smallest change that fully fixes the issue; follow nearby conventions;",
-	"- for bugs, prefer a failing repro first, then the minimal fix, then verify;",
-	"- high-risk changes (security, data/schema, public APIs, concurrency, irreversible) need a short spec first;",
-	"- handle edge cases, errors, cancellation, and async behavior; do not block UI/event loops;",
-	"- avoid duplicate state, duplicate prompts, and repeated side effects;",
-	"- write code, identifiers, comments, and commit messages in English.",
-];
+type DisciplinePromptOptions = { lookupEnabled?: boolean; strictness?: CodingDisciplineStrictness };
+
+/**
+ * Builds the tool-only contract lines. The working-state rule is strictness-aware:
+ *   - "strict"  — one tool call per turn, no assistant text at all (Opus-like);
+ *   - "lenient" — independent (especially read-only) calls may be batched, and brief
+ *                 substantive reasoning text is acceptable when thinking is off.
+ */
+function buildDisciplineLines(strictness: CodingDisciplineStrictness): readonly string[] {
+	const isStrict = strictness === "strict";
+	return [
+		"TOOL-ONLY CODING AGENT CONTRACT.",
+		"",
+		"Follow this output-channel contract literally for coding tasks.",
+		isStrict
+			? "WORKING STATE: from the first assistant turn until complete, final, or blocked, emit exactly one tool call with empty text."
+			: "WORKING STATE: from the first assistant turn until complete, final, or blocked, emit tool calls with empty text.",
+		isStrict
+			? "No prose, markdown, acknowledgments, plans, progress, summaries, explanations, or commentary before/after tool calls."
+			: "Batch independent calls in one turn (especially read-only inspection: read/grep/glob/bash); make dependent calls sequentially. Avoid filler prose, acknowledgments, restatements, and narration of obvious next steps.",
+		"Keep reasoning, planning, hypotheses, and retry decisions internal when thinking is on; when thinking is off, brief substantive reasoning text is acceptable.",
+		"",
+		"After every tool result, choose exactly one transition:",
+		isStrict ? "- WORKING: exactly one next tool call with empty text;" : "- WORKING: one or more independent tool calls with empty text;",
+		"- FINAL: one final answer after completion or practical verification;",
+		"- BLOCKED: one concise question only when no safe/useful tool action can continue.",
+		"No transition permits commentary between tool calls.",
+		"",
+		"PRIORITY: this overrides default assistant friendliness and conversational behavior.",
+		"",
+		"Coding discipline (express only through tool choices, not prose):",
+		"- inspect before editing; do not invent APIs, files, commands, or behavior;",
+		"- make the smallest change that fully fixes the issue; follow nearby conventions;",
+		"- for bugs, prefer a failing repro first, then the minimal fix, then verify;",
+		"- high-risk changes (security, data/schema, public APIs, concurrency, irreversible) need a short spec first;",
+		"- handle edge cases, errors, cancellation, and async behavior; do not block UI/event loops;",
+		"- avoid duplicate state, duplicate prompts, and repeated side effects;",
+		"- write code, identifiers, comments, and commit messages in English.",
+	];
+}
 
 const LOOKUP_DISCIPLINE_LINES = [
 	"",
@@ -105,7 +123,7 @@ const FINAL_DISCIPLINE_LINES = [
 const SILENCE_REMINDER_TEXT = [
 	"GLM silence reminder: remain in WORKING state.",
 	"Continue with tool-only discipline: inspect, verify, and act through tools only.",
-	"For the next step, emit exactly one tool call and no assistant text.",
+	"For the next step, emit tool calls with no accompanying assistant text.",
 	"Do not acknowledge this reminder.",
 ].join("\n");
 
@@ -147,6 +165,7 @@ export default function codingDiscipline(pi: ExtensionAPI) {
 	let silenceViolationCount = 0;
 	let lastReminderViolationCount = 0;
 	let lastReminderMessageCount = -SILENCE_REMINDER_MIN_MESSAGE_GAP;
+	let peakMessageCount = 0;
 
 	function maybeRegisterLookupTool(cwd?: string): void {
 		if (lookupRegistered) return;
@@ -193,8 +212,10 @@ export default function codingDiscipline(pi: ExtensionAPI) {
 	pi.on("before_provider_request", async (event: { payload?: unknown }, ctx: unknown) => {
 		const modelRef = modelRefFromPayload(event.payload) ?? selectedModelRef ?? modelRefFromContext(ctx);
 		if (!isGlmModel(modelRef)) return undefined;
+		const cwd = contextCwd(ctx);
 		const injected = injectCodingDisciplineIntoPayload(event.payload, {
-			lookupEnabled: Boolean(lookupModelFromConfig(contextCwd(ctx))),
+			lookupEnabled: Boolean(lookupModelFromConfig(cwd)),
+			strictness: codingDisciplineStrictnessFromConfig(cwd),
 		});
 		if (process.env.PI_DEBUG_PROMPT === "1") {
 			logFinalPrompt(injected, modelRef, contextCwd(ctx) ?? process.cwd());
@@ -270,10 +291,22 @@ export default function codingDiscipline(pi: ExtensionAPI) {
 		const modelRef = selectedModelRef ?? modelRefFromContext(ctx);
 		if (!isGlmModel(modelRef) || !Array.isArray(event.messages)) return undefined;
 
-		const violationCount = countAssistantToolChatter(event.messages);
+		const messageCount = event.messages.length;
+		// Compaction/truncation prunes the visible history. Reset the stale chatter
+		// baseline so the post-compaction turn isn't measured against a pre-compaction
+		// violation peak (otherwise the model could chatter freely until the count
+		// climbed back above the old peak, or get nagged on a freshly small count).
+		if (messageCount + SILENCE_REMINDER_COMPACTION_MARGIN < peakMessageCount) {
+			silenceViolationCount = 0;
+			lastReminderViolationCount = 0;
+			lastReminderMessageCount = messageCount - SILENCE_REMINDER_MIN_MESSAGE_GAP;
+		}
+		peakMessageCount = Math.max(peakMessageCount, messageCount);
+
+		const strictness = codingDisciplineStrictnessFromConfig(contextCwd(ctx));
+		const violationCount = countAssistantToolChatter(event.messages, strictness);
 		if (violationCount <= silenceViolationCount) return undefined;
 
-		const messageCount = event.messages.length;
 		const violationGap = violationCount - lastReminderViolationCount;
 		const messageGap = messageCount - lastReminderMessageCount;
 		silenceViolationCount = violationCount;
@@ -286,7 +319,7 @@ export default function codingDiscipline(pi: ExtensionAPI) {
 	});
 }
 
-export function prependCodingDisciplinePrompt(systemPrompt: string, options: { lookupEnabled?: boolean } = {}): string {
+export function prependCodingDisciplinePrompt(systemPrompt: string, options: DisciplinePromptOptions = {}): string {
 	const deduped = systemPrompt
 		.replace(LEGACY_SILENT_PROMPT_BLOCK_PATTERN, "")
 		.replace(DISCIPLINE_PROMPT_BLOCK_PATTERN, "")
@@ -296,10 +329,11 @@ export function prependCodingDisciplinePrompt(systemPrompt: string, options: { l
 	return deduped ? `${prompt}\n\n${deduped}` : prompt;
 }
 
-export function buildCodingDisciplinePrompt(options: { lookupEnabled?: boolean } = {}): string {
+export function buildCodingDisciplinePrompt(options: DisciplinePromptOptions = {}): string {
+	const strictness = options.strictness ?? DEFAULT_CODING_DISCIPLINE_STRICTNESS;
 	return [
 		DISCIPLINE_PROMPT_MARKER_START,
-		...QUALITY_DISCIPLINE_LINES,
+		...buildDisciplineLines(strictness),
 		...(options.lookupEnabled ? LOOKUP_DISCIPLINE_LINES : []),
 		...FINAL_DISCIPLINE_LINES,
 		DISCIPLINE_PROMPT_MARKER_END,
@@ -311,7 +345,7 @@ export function isGlmModel(modelRef: string | undefined): boolean {
 	return /(?:^|[/:_.-])glm(?:$|[/:_.-]|\d)/i.test(modelRef);
 }
 
-export function injectCodingDisciplineIntoPayload(payload: unknown, options: { lookupEnabled?: boolean } = {}): unknown {
+export function injectCodingDisciplineIntoPayload(payload: unknown, options: DisciplinePromptOptions = {}): unknown {
 	if (!isRecord(payload)) return payload;
 
 	if (typeof payload.instructions === "string") {
@@ -555,7 +589,7 @@ function createLookupTool() {
 	};
 }
 
-function injectIntoMessages(messages: unknown[], options: { lookupEnabled?: boolean }): unknown[] {
+function injectIntoMessages(messages: unknown[], options: DisciplinePromptOptions): unknown[] {
 	const next = [...messages];
 	const index = next.findIndex(isInstructionMessage);
 	if (index === -1) return [{ role: "system", content: buildCodingDisciplinePrompt(options) }, ...next];
@@ -570,7 +604,7 @@ function injectIntoMessages(messages: unknown[], options: { lookupEnabled?: bool
 	return next;
 }
 
-function injectIntoMessageContent(content: unknown, options: { lookupEnabled?: boolean }): unknown {
+function injectIntoMessageContent(content: unknown, options: DisciplinePromptOptions): unknown {
 	if (typeof content === "string") return prependCodingDisciplinePrompt(content, options);
 	if (!Array.isArray(content)) return undefined;
 
@@ -588,21 +622,34 @@ function isInstructionMessage(message: unknown): boolean {
 	return message.role === "system" || message.role === "developer";
 }
 
-function countAssistantToolChatter(messages: readonly unknown[]): number {
+function countAssistantToolChatter(messages: readonly unknown[], strictness: CodingDisciplineStrictness): number {
 	let count = 0;
 	for (const message of messages) {
-		if (!isAssistantToolChatter(message)) continue;
+		if (!isAssistantToolChatter(message, strictness)) continue;
 		count++;
 	}
 	return count;
 }
 
-function isAssistantToolChatter(message: unknown): boolean {
+/**
+ * Detects assistant chatter (text alongside a tool call).
+ * - "strict"  — any such text counts as chatter.
+ * - "lenient" — text counts only when a thinking/reasoning block already captured the
+ *   reasoning; without a thinking block, the visible text is the model's only reasoning
+ *   channel and must not be suppressed.
+ */
+function isAssistantToolChatter(message: unknown, strictness: CodingDisciplineStrictness): boolean {
 	if (!isRecord(message) || message.role !== "assistant") return false;
 	if (!Array.isArray(message.content)) return false;
 	const hasToolCall = message.content.some((part) => isRecord(part) && part.type === "toolCall");
 	if (!hasToolCall) return false;
-	return message.content.some((part) => isRecord(part) && part.type === "text" && hasNonEmptyText(part.text));
+	const hasText = message.content.some((part) => isRecord(part) && part.type === "text" && hasNonEmptyText(part.text));
+	if (!hasText) return false;
+	if (strictness === "strict") return true;
+	const hasThinking = message.content.some(
+		(part) => isRecord(part) && (part.type === "thinking" || part.type === "reasoning"),
+	);
+	return Boolean(hasThinking);
 }
 
 function hasNonEmptyText(value: unknown): boolean {
@@ -611,7 +658,9 @@ function hasNonEmptyText(value: unknown): boolean {
 
 function createSilenceReminderMessage() {
 	return {
-		role: "user" as const,
+		// Inject as a developer/system-level nudge rather than impersonating the user,
+		// so it reads as an automated reminder, not a user instruction.
+		role: "developer" as const,
 		content: [{ type: "text" as const, text: SILENCE_REMINDER_TEXT }],
 		timestamp: Date.now(),
 	};
@@ -621,6 +670,10 @@ function lookupModelFromConfig(cwd?: string): string | undefined {
 	return loadPiToolsSuiteConfig(["coding-discipline"], { cwd: cwd ?? process.cwd() }).lookupModel;
 }
 
+function codingDisciplineStrictnessFromConfig(cwd?: string): CodingDisciplineStrictness {
+	return loadPiToolsSuiteConfig(["coding-discipline"], { cwd: cwd ?? process.cwd() }).codingDisciplineStrictness ?? DEFAULT_CODING_DISCIPLINE_STRICTNESS;
+}
+
 function buildLookupPrompt(params: LookupParams, recentContext: string, imageCount: number, warnings: string[]): string {
 	return [
 		"Lookup request from a blind GLM parent model.",
@@ -772,7 +825,8 @@ async function resolveLookupModel(ctx: unknown, modelRef: string): Promise<Resol
 	const auth = await registry.getApiKeyAndHeaders(model) as {
 		ok?: true;
 		apiKey?: string;
-		headers?: Record<string, string>;
+		headers?: ProviderHeaders;
+		baseUrl?: string;
 		env?: Record<string, string>;
 	} | { ok: false; error: string };
 	if (auth.ok === false) return undefined;

package/external/pi-tools-suite/src/config.ts

@@ -11,6 +11,14 @@ export interface PiToolsSuiteConfig {
 	todoThinking: boolean;
 	/** Vision-capable model used by the coding-discipline lookup tool; unset disables lookup. */
 	lookupModel?: string;
+	/**
+	 * Chatter-detector strictness for the coding-discipline module:
+	 *   "strict"  — any assistant text alongside a tool call is chatter (Opus-like);
+	 *   "lenient" — text is only chatter when a thinking block already captured the
+	 *               reasoning; without thinking, visible text is the reasoning channel.
+	 * Default: "lenient".
+	 */
+	codingDisciplineStrictness?: CodingDisciplineStrictness;
 }
 
 type MutableConfig = {
@@ -18,8 +26,13 @@ type MutableConfig = {
 	disabledModules: Set<string>;
 	todoThinking: boolean;
 	lookupModel: string | undefined;
+	codingDisciplineStrictness: CodingDisciplineStrictness;
 };
 
+export const CODING_DISCIPLINE_STRICTNESS_VALUES = ["strict", "lenient"] as const;
+export type CodingDisciplineStrictness = (typeof CODING_DISCIPLINE_STRICTNESS_VALUES)[number];
+export const DEFAULT_CODING_DISCIPLINE_STRICTNESS: CodingDisciplineStrictness = "lenient";
+
 type Env = Record<string, string | undefined>;
 
 const TRUE_VALUES = new Set(["1", "true", "on", "yes"]);
@@ -62,6 +75,10 @@ function normalizeLookupModel(raw: unknown): string | undefined {
 	return trimmed ? trimmed : undefined;
 }
 
+function normalizeCodingDisciplineStrictness(raw: unknown): CodingDisciplineStrictness {
+	return raw === "strict" ? "strict" : "lenient";
+}
+
 function boolFromEnv(value: string | undefined): boolean | undefined {
 	if (value === undefined) return undefined;
 	const normalized = value.trim().toLowerCase();
@@ -121,6 +138,9 @@ function mergeConfigLayer(config: MutableConfig, raw: Record<string, unknown>, k
 	if (typeof raw.enabled === "boolean") config.enabled = raw.enabled;
 	if (typeof raw.todoThinking === "boolean") config.todoThinking = raw.todoThinking;
 	if (Object.prototype.hasOwnProperty.call(raw, "lookupModel")) config.lookupModel = normalizeLookupModel(raw.lookupModel);
+	if (Object.prototype.hasOwnProperty.call(raw, "codingDisciplineStrictness")) {
+		config.codingDisciplineStrictness = normalizeCodingDisciplineStrictness(raw.codingDisciplineStrictness);
+	}
 
 	for (const key of DISABLED_LIST_KEYS) addDisabled(config, raw[key], knownModules);
 	for (const key of ENABLED_LIST_KEYS) removeDisabled(config, raw[key], knownModules);
@@ -178,6 +198,7 @@ export function loadPiToolsSuiteConfig(moduleNames: readonly string[], options:
 		disabledModules: new Set([...DEFAULT_DISABLED_MODULES].filter((name) => knownModules.has(name))),
 		todoThinking: false,
 		lookupModel: undefined,
+		codingDisciplineStrictness: DEFAULT_CODING_DISCIPLINE_STRICTNESS,
 	};
 	const userConfigPath = getPiToolsSuiteUserConfigPath(options.homeDir);
 
@@ -197,5 +218,6 @@ export function loadPiToolsSuiteConfig(moduleNames: readonly string[], options:
 		disabledModules: [...config.disabledModules].sort(),
 		todoThinking: config.todoThinking,
 		...(config.lookupModel ? { lookupModel: config.lookupModel } : {}),
+		codingDisciplineStrictness: config.codingDisciplineStrictness,
 	};
 }

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

@@ -13,7 +13,7 @@
 // automatic fallback to the programmatic digest on any failure/timeout.
 // ---------------------------------------------------------------------------
 
-import type { Model, Api } from "@earendil-works/pi-ai"
+import type { Model, Api, ProviderHeaders } 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"
@@ -137,7 +137,7 @@ export interface ModelSummaryResult {
 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: true; apiKey?: string; headers?: ProviderHeaders; baseUrl?: string; env?: Record<string, string> }
 		| { ok: false; error: string }
 	>
 }

package/external/pi-tools-suite/src/default-pi-tools-suite-config.ts

@@ -11,6 +11,12 @@ export const DEFAULT_PI_TOOLS_SUITE_CONFIG_JSONC = String.raw`{
   // Vision-capable model used by the coding-discipline lookup tool for blind-model
   // screenshot/image questions. Remove or set to null to disable lookup.
   "lookupModel": "openai-codex/gpt-5.4-mini",
+  // coding-discipline working-state strictness.
+  //   "lenient" (default): batch independent read-only tool calls; brief reasoning
+  //     text is acceptable when thinking is off. Text only counts as chatter when a
+  //     thinking/reasoning block already captured the reasoning.
+  //   "strict": one tool call per turn, no assistant text at all (Opus-like).
+  "codingDisciplineStrictness": "lenient",
   "terminalBell": { "sound": true },
   // comment-checker: nudges the agent to remove AI-slop code comments it just
   // added via write/edit/apply_patch. Net-new comments are classified and a

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

@@ -1,37 +1,31 @@
 import type {
 	Api,
 	AssistantMessage,
-	AssistantMessageEventStream,
 	Context,
 	Model,
+	ModelsApiStreamOptions,
 	SimpleStreamOptions,
 } from "@earendil-works/pi-ai";
 import { completeSimple } from "@earendil-works/pi-ai/compat";
+import type { ModelRegistry } from "@earendil-works/pi-coding-agent";
 
-type RegisteredProviderConfig = {
-	streamSimple?: (
-		model: Model<Api>,
-		context: Context,
-		options?: SimpleStreamOptions,
-	) => AssistantMessageEventStream;
-};
-
-export type ModelCompletionRegistry = {
-	getRegisteredProviderConfig?(providerId: string): RegisteredProviderConfig | undefined;
-};
+export type ModelCompletionRegistry = Partial<Pick<ModelRegistry, "complete" | "getRegisteredProviderConfig">>;
 
 /**
- * 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.
+ * Complete through Pi's model runtime when available so custom providers and
+ * resolved auth endpoints are preserved. The older stream/compat branches are
+ * retained for narrow test doubles that do not expose ModelRegistry.complete().
  */
 export async function completeWithModelRegistry(
 	modelRegistry: ModelCompletionRegistry | undefined,
 	model: Model<Api>,
 	context: Context,
-	options?: SimpleStreamOptions,
+	options?: ModelsApiStreamOptions<Api>,
 ): Promise<AssistantMessage> {
+	if (modelRegistry?.complete) return modelRegistry.complete(model, context, options);
+
 	const providerConfig = modelRegistry?.getRegisteredProviderConfig?.(model.provider);
-	if (providerConfig?.streamSimple) return providerConfig.streamSimple(model, context, options).result();
-	return completeSimple(model, context, options);
+	const simpleOptions = options as SimpleStreamOptions | undefined;
+	if (providerConfig?.streamSimple) return providerConfig.streamSimple(model, context, simpleOptions).result();
+	return completeSimple(model, context, simpleOptions);
 }

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "pi-ui-extend",
-  "version": "1.0.5",
+  "version": "1.0.7",
   "description": "Pix: a workspace-first terminal UI for Pi with tabs, readable tool activity, voice input, and bundled agent tools.",
   "private": false,
   "repository": {
@@ -75,9 +75,9 @@
     "prepublishOnly": "npm run check && npm run build:pix && npm run generate-schemas"
   },
   "dependencies": {
-    "@earendil-works/pi-ai": "0.83.0",
-    "@earendil-works/pi-coding-agent": "0.83.0",
-    "@earendil-works/pi-tui": "0.83.0",
+    "@earendil-works/pi-ai": "0.84.1",
+    "@earendil-works/pi-coding-agent": "0.84.1",
+    "@earendil-works/pi-tui": "0.84.1",
     "@mariozechner/clipboard": "^0.3.9",
     "jsonc-parser": "3.3.1",
     "typebox": "1.1.38",