pi-ui-extend 1.0.39 → 1.0.40

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

@@ -3,7 +3,7 @@ import * as path from "node:path";
 import type { Api, AssistantMessage, ImageContent, Model, ProviderHeaders, TextContent } from "@earendil-works/pi-ai";
 import { Type } from "typebox";
 
-import { loadPiToolsSuiteConfig, DEFAULT_CODING_DISCIPLINE_STRICTNESS, type CodingDisciplineStrictness } from "../config.js";
+import { loadPiToolsSuiteConfig } from "../config.js";
 import { ignoreStaleExtensionContextError } from "../context-usage.js";
 import { completeWithModelRegistry, type ModelCompletionRegistry } from "../model-completion.js";
 
@@ -41,13 +41,6 @@ const DEFAULT_LOOKUP_MAX_IMAGES = 6;
 const DEFAULT_LOOKUP_MAX_TOKENS = 1_600;
 const DEFAULT_LOOKUP_TIMEOUT_MS = 120_000;
 const MAX_IMAGE_BYTES = 16 * 1024 * 1024;
-// Keep recovery nudges sparse: newer GLM coding models are substantially better at
-// long-horizon tool use, and repeated developer reminders can become prompt noise.
-const SILENCE_REMINDER_MIN_VIOLATION_GAP = 5;
-const SILENCE_REMINDER_MIN_MESSAGE_GAP = 20;
-// 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(
@@ -65,48 +58,46 @@ const LOOKUP_TOOL_PARAMS = Type.Object(
 	{ additionalProperties: false },
 );
 
-type DisciplinePromptOptions = { lookupEnabled?: boolean; strictness?: CodingDisciplineStrictness };
+type DisciplinePromptOptions = { lookupEnabled?: boolean };
 
-/**
- * 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.",
-		isStrict
-			? "No transition permits commentary between tool calls."
-			: "When a thinking/reasoning channel is available, do not duplicate that reasoning as visible 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 QUALITY_DISCIPLINE_LINES = [
+	"GLM CODING QUALITY CONTRACT.",
+	"",
+	"Optimize for correct, evidence-backed changes rather than fast patching or conversational polish.",
+	"",
+	"Evidence and hypothesis discipline:",
+	"- inspect before editing; do not invent APIs, files, commands, behavior, or repository conventions;",
+	"- separate observed facts from hypotheses; do not commit to the first plausible root cause;",
+	"- for non-obvious bugs, keep at least two plausible hypotheses until evidence distinguishes them, and choose the next inspection/test for information gain;",
+	"- do not use edits as diagnostic probes when read/search/repro/test evidence can answer the question first;",
+	"- when behavior crosses modules, trace the relevant execution/data flow before editing instead of serially opening plausible files; use indexed architecture/search/dependency tools when available;",
+	"- before changing an API, state contract, persistence format, or async lifecycle, inspect relevant callers, consumers, tests, persistence boundaries, and cancellation/error paths.",
+	"",
+	"Implementation discipline:",
+	"- make the smallest change that fully fixes the issue and follow nearby conventions;",
+	"- for bugs, reproduce the failure or its defining condition when practical before changing code;",
+	"- high-risk changes (security, data/schema, public APIs, concurrency, irreversible behavior) need a short evidence-backed spec before implementation;",
+	"- handle relevant edge cases, errors, cancellation, stale state, and async behavior; do not block UI/event loops;",
+	"- avoid duplicate state, duplicate prompts, repeated side effects, and speculative refactors unrelated to the verified cause;",
+	"- write code, identifiers, comments, and commit messages in English.",
+	"",
+	"Failure and retry discipline:",
+	"- treat failed verification as new evidence; do not repeat the same approach with cosmetic variations;",
+	"- after two materially failed fix/verification attempts, stop patching: re-inspect evidence, revise the root-cause hypothesis, or escalate for an independent opinion.",
+	"",
+	"Verification discipline:",
+	"- verification must cover the behavioral claim that justified the change; typecheck/build/lint prove structural validity, not behavioral correctness;",
+	"- for bug fixes, verify the old failure or its defining condition no longer occurs whenever practical;",
+	"- for async/stateful changes, verify the success path plus the relevant cancellation, error, retry, duplicate-event, or stale-state path;",
+	"- before finalizing a non-trivial change, actively inspect or test one likely counterexample, regression path, or missed edge case;",
+	"- never claim a test/check passed unless it actually ran and its result was observed.",
+	"",
+	"Escalation discipline:",
+	"- when async subagents are available, use an independent `oracle` (preferably Sol through the configured oracle role) when focused inspection still leaves conflicting evidence, multiple incompatible root causes, a subtle cross-module invariant you cannot verify locally, or a high-risk security/schema/public-API/concurrency/irreversible decision;",
+	"- also escalate after two materially different fixes fail for the same unresolved cause;",
+	"- do not escalate routine implementation, mechanical edits, straightforward test failures, or facts that local tools can verify directly;",
+	"- treat oracle output as a second opinion, not authority: reconcile it with repository evidence before acting.",
+];
 
 const LOOKUP_DISCIPLINE_LINES = [
 	"",
@@ -120,17 +111,9 @@ const LOOKUP_DISCIPLINE_LINES = [
 const FINAL_DISCIPLINE_LINES = [
 	"",
 	"When uncertain, test or inspect instead of assuming; otherwise proceed with grounded best effort.",
-	"Verify every non-trivial change. Never claim tests passed unless actually run.",
 	"Final report: what changed, what was verified, what was not verified, and any risks.",
 ];
 
-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 tool calls with no accompanying assistant text.",
-	"Do not acknowledge this reminder.",
-].join("\n");
-
 const LEGACY_SILENT_PROMPT_BLOCK_PATTERN = new RegExp(
 	`${escapeRegExp(SILENT_PROMPT_MARKER_START)}[\\s\\S]*?${escapeRegExp(SILENT_PROMPT_MARKER_END)}\\s*`,
 	"g",
@@ -167,10 +150,6 @@ const LOOKUP_SYSTEM_PROMPT = [
 export default function codingDiscipline(pi: ExtensionAPI) {
 	let selectedModelRef: string | undefined;
 	let lookupRegistered = false;
-	let silenceViolationCount = 0;
-	let lastReminderViolationCount = 0;
-	let lastReminderMessageCount = -SILENCE_REMINDER_MIN_MESSAGE_GAP;
-	let peakMessageCount = 0;
 
 	function maybeRegisterLookupTool(cwd?: string): void {
 		if (lookupRegistered) return;
@@ -220,7 +199,6 @@ export default function codingDiscipline(pi: ExtensionAPI) {
 		const cwd = contextCwd(ctx);
 		const injected = injectCodingDisciplineIntoPayload(event.payload, {
 			lookupEnabled: Boolean(lookupModelFromConfig(cwd)) && isBlindGlmModel(modelRef),
-			strictness: codingDisciplineStrictnessFromConfig(cwd),
 		});
 		if (process.env.PI_DEBUG_PROMPT === "1") {
 			logFinalPrompt(injected, modelRef, contextCwd(ctx) ?? process.cwd());
@@ -292,36 +270,6 @@ export default function codingDiscipline(pi: ExtensionAPI) {
 		return undefined;
 	});
 
-	pi.on("context", async (event: { messages?: unknown[] }, ctx: unknown) => {
-		const modelRef = selectedModelRef ?? modelRefFromContext(ctx);
-		if (!isGlmModel(modelRef) || !Array.isArray(event.messages)) return undefined;
-
-		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 violationGap = violationCount - lastReminderViolationCount;
-		const messageGap = messageCount - lastReminderMessageCount;
-		silenceViolationCount = violationCount;
-
-		if (violationGap < SILENCE_REMINDER_MIN_VIOLATION_GAP && messageGap < SILENCE_REMINDER_MIN_MESSAGE_GAP) return undefined;
-
-		lastReminderViolationCount = violationCount;
-		lastReminderMessageCount = messageCount;
-		return { messages: [...event.messages, createSilenceReminderMessage()] };
-	});
 }
 
 export function prependCodingDisciplinePrompt(systemPrompt: string, options: DisciplinePromptOptions = {}): string {
@@ -335,10 +283,9 @@ export function prependCodingDisciplinePrompt(systemPrompt: string, options: Dis
 }
 
 export function buildCodingDisciplinePrompt(options: DisciplinePromptOptions = {}): string {
-	const strictness = options.strictness ?? DEFAULT_CODING_DISCIPLINE_STRICTNESS;
 	return [
 		DISCIPLINE_PROMPT_MARKER_START,
-		...buildDisciplineLines(strictness),
+		...QUALITY_DISCIPLINE_LINES,
 		...(options.lookupEnabled ? LOOKUP_DISCIPLINE_LINES : []),
 		...FINAL_DISCIPLINE_LINES,
 		DISCIPLINE_PROMPT_MARKER_END,
@@ -637,58 +584,10 @@ function isInstructionMessage(message: unknown): boolean {
 	return message.role === "system" || message.role === "developer";
 }
 
-function countAssistantToolChatter(messages: readonly unknown[], strictness: CodingDisciplineStrictness): number {
-	let count = 0;
-	for (const message of messages) {
-		if (!isAssistantToolChatter(message, strictness)) continue;
-		count++;
-	}
-	return count;
-}
-
-/**
- * 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;
-	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 {
-	return typeof value === "string" && value.trim().length > 0;
-}
-
-function createSilenceReminderMessage() {
-	return {
-		// 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(),
-	};
-}
-
 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 text-only GLM parent model.",

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

@@ -12,14 +12,6 @@ export interface PiToolsSuiteConfig {
 	todoThinkingOverrides: Record<string, TodoThinkingLevel>;
 	/** 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 = {
@@ -28,12 +20,8 @@ type MutableConfig = {
 	todoThinking: boolean;
 	todoThinkingOverrides: Map<string, TodoThinkingLevel>;
 	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";
 const TODO_THINKING_OVERRIDE_LEVELS = ["off", "minimal", "low", "medium", "high", "xhigh", "max"] as const;
 export type TodoThinkingLevel = (typeof TODO_THINKING_OVERRIDE_LEVELS)[number];
 
@@ -80,10 +68,6 @@ function normalizeLookupModel(raw: unknown): string | undefined {
 	return trimmed ? trimmed : undefined;
 }
 
-function normalizeCodingDisciplineStrictness(raw: unknown): CodingDisciplineStrictness {
-	return raw === "strict" ? "strict" : "lenient";
-}
-
 function isTodoThinkingLevel(raw: unknown): raw is TodoThinkingLevel {
 	return TODO_THINKING_OVERRIDE_LEVELS.includes(raw as TodoThinkingLevel);
 }
@@ -161,9 +145,6 @@ function mergeConfigLayer(config: MutableConfig, raw: Record<string, unknown>, k
 	if (typeof raw.todoThinking === "boolean") config.todoThinking = raw.todoThinking;
 	mergeTodoThinkingOverrides(config, raw.todoThinkingOverrides);
 	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);
@@ -222,7 +203,6 @@ export function loadPiToolsSuiteConfig(moduleNames: readonly string[], options:
 		todoThinking: false,
 		todoThinkingOverrides: new Map(DEFAULT_TODO_THINKING_OVERRIDES),
 		lookupModel: undefined,
-		codingDisciplineStrictness: DEFAULT_CODING_DISCIPLINE_STRICTNESS,
 	};
 	const userConfigPath = getPiToolsSuiteUserConfigPath(options.homeDir);
 
@@ -243,6 +223,5 @@ export function loadPiToolsSuiteConfig(moduleNames: readonly string[], options:
 		todoThinking: config.todoThinking,
 		todoThinkingOverrides: Object.fromEntries(config.todoThinkingOverrides),
 		...(config.lookupModel ? { lookupModel: config.lookupModel } : {}),
-		codingDisciplineStrictness: config.codingDisciplineStrictness,
 	};
 }

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

@@ -24,12 +24,6 @@ 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": "zai/glm-5.3-flash",
-  // 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
@@ -398,6 +392,29 @@ export const DEFAULT_PI_TOOLS_SUITE_CONFIG_JSONC = String.raw`{
       }
       // ,
       // {
+      //   "id": "svelte",
+      //   "include": [
+      //     "**/*.svelte"
+      //   ],
+      //   "exclude": [
+      //     "**/node_modules/**"
+      //   ],
+      //   "rootMarkers": [
+      //     "svelte.config.js",
+      //     "package.json"
+      //   ],
+      //   "bin": "svelteserver",
+      //   "args": [
+      //     "--stdio"
+      //   ],
+      //   "startupTimeoutMs": 30000,
+      //   "diagnosticsWaitMs": 8000,
+      //   "languageIdByExtension": {
+      //     ".svelte": "svelte"
+      //   }
+      // }
+      // ,
+      // {
       //   "id": "python",
       //   "include": [
       //     "**/*.py",

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

@@ -39,10 +39,11 @@ export function astGrepToolDescriptions(maxLines: number, maxBytesLabel: string)
 		astGrep: {
 			name: "ast_grep",
 			label: "ast-grep",
-			description: `Read-only AST structural search/scan. Use for language-aware patterns, sgconfig/rule scans, JSON matches, and rewrite previews. Use text search for plain strings and ast_apply for mutations. Output truncates at ${maxLines} lines or ${maxBytesLabel} with full output saved to a temp file.`,
-			promptSnippet: "Use ast_grep for AST/structural code search, not plain text search. It previews rewrites only; use ast_apply to mutate files.",
+			description: `Read-only AST structural search/scan. MANDATORY ROUTING: for structural, syntax-aware, AST, language-aware, or code-shape matching, ast_grep must be the FIRST tool call. Never preflight with Glob, Grep, Read, or shell even when files are unknown; omit paths to scan the current project. Use for sgconfig/rule scans, JSON matches, and rewrite previews. Use Grep for exact literals/regex, Glob only for filename/path-only requests, and ast_apply for mutations. Output truncates at ${maxLines} lines or ${maxBytesLabel} with full output saved to a temp file.`,
+			promptSnippet: "MANDATORY: structural/syntax-aware/AST/code-shape matching => call ast_grep FIRST. Do not call Glob, Grep, Read, or shell before it; unknown files are not a reason to pre-search because ast_grep scans the project by default. Exact literal/regex only => Grep; filename/path only => Glob. Use ast_apply for mutations.",
 			promptGuidelines: [
-				"Use ast_grep when syntax/AST structure matters; use text search for exact strings/regex. Keep paths/globs narrow and set lang for ambiguous snippets.",
+				"The first tool for syntax relationships or code-shape matching must be ast_grep, even when the file or language is unknown. Do not make a preliminary Glob/Grep/Read/shell call; start at the current project and narrow within ast_grep when needed.",
+				"Use Grep/Read directly for exact literal or regex lookups and Glob only for filename/path-only discovery; those text-only tasks must not trigger ast_grep.",
 				"ast_grep is read-only: use rewrite to preview only; use ast_apply for mutations or command=scan fixes.",
 			],
 		},

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "pi-ui-extend",
-  "version": "1.0.39",
+  "version": "1.0.40",
   "description": "Pix: a workspace-first terminal UI for Pi with tabs, readable tool activity, voice input, and bundled agent tools.",
   "private": false,
   "repository": {
@@ -83,10 +83,9 @@
     "prepublishOnly": "npm run check && npm run build:pix && npm run generate-schemas"
   },
   "dependencies": {
-    "@earendil-works/pi-ai": "0.85.0",
-    "@earendil-works/pi-coding-agent": "0.85.0",
-    "@earendil-works/pi-server": "0.85.0",
-    "@earendil-works/pi-tui": "0.85.0",
+    "@earendil-works/pi-ai": "0.85.1",
+    "@earendil-works/pi-coding-agent": "0.85.1",
+    "@earendil-works/pi-tui": "0.85.1",
     "@mariozechner/clipboard": "^0.3.9",
     "grok-mermaid": "0.2.2",
     "jsonc-parser": "3.3.1",