@cleocode/cleo 2026.9.7 → 2026.9.9

package/studio-dist/server/chunks/llm-extraction-DVCu4heR.js

@@ -1,408 +0,0 @@
-import { s as object, z as array, u as string, v as number, _ as _enum } from './paths-CV7S5-Zm.js';
-import './index-Dtt_h58T.js';
-import 'node:fs';
-import 'node:path';
-import 'node:crypto';
-import 'node:fs/promises';
-import './index-61916rmV.js';
-import { a as resolveAnthropicForRole } from './role-resolver-D9QaDT0e.js';
-import { c as checkHashDedup, a as verifyAndStore } from './extraction-gate2-D77EZBQ6.js';
-import 'node:os';
-import 'node:events';
-import 'node:diagnostics_channel';
-import 'fs';
-import 'events';
-import 'util';
-import 'path';
-import 'assert';
-import 'worker_threads';
-import 'module';
-import 'node:url';
-import 'url';
-import 'buffer';
-import 'node:module';
-import 'node:child_process';
-import 'node:process';
-import 'node:async_hooks';
-import 'constants';
-import 'stream';
-import './config-DHj9CcFs.js';
-import './json-Bfixam3g.js';
-import './lock-CA7d6lj7.js';
-import './credential-pool-vF6NdtT9.js';
-import './brain-embedding-Djw9ekXu.js';
-import './graph-auto-populate2-Cz0qeebM.js';
-import './sqlite-Bm5LzbPs.js';
-import './dual-scope-db-DKnWu1KM.js';
-import './indexes-DZ52grq6.js';
-import './utils-BPSZyajG.js';
-import './query-promise-BkPmGPJy.js';
-import './nexus-sqlite2-BBfY2xnY.js';
-import './nexus-schema2-DP0Qd7CA.js';
-import './chunk-DDP5xfzh.js';
-import './logger-DQIoyZbq.js';
-import 'node:sqlite';
-import './unique-constraint-BZrrwgwc.js';
-import './brain-similarity-DI4yzm9e.js';
-import './memory-accessor2-B981tHDR.js';
-import './typed-query-l1D_QElt.js';
-
-//#region ../core/dist/memory/llm-extraction.js
-/**
-* LLM-driven memory extraction gate.
-*
-* Replaces the legacy keyword regex in `auto-extract.ts`. Takes a session
-* transcript and uses an Anthropic model to extract typed, structured
-* memories (decisions, patterns, learnings, constraints, corrections)
-* with justification, importance score, and referenced entities.
-*
-* Design:
-*   - Runs only when `brain.llmExtraction.enabled` is true AND an Anthropic
-*     API key is available (from ANTHROPIC_API_KEY env var OR Claude Code
-*     credentials at ~/.claude/.credentials.json). Otherwise returns empty.
-*   - Uses structured output (JSON schema) for reliable parsing — no regex
-*     post-processing.
-*   - Never throws — all errors are caught and logged. The pipeline must
-*     never block session end.
-*   - Store routing: each extracted memory is sent through the existing
-*     verify-and-store pipeline. Decisions → brain_decisions, patterns →
-*     brain_patterns, learnings/constraints/corrections → brain_learnings.
-*   - Source is tagged `agent-llm-extracted` so downstream dedup, quality
-*     scoring, and consolidation can distinguish LLM output from other
-*     pipelines.
-*
-* Research basis: `.cleo/agent-outputs/R-llm-memory-systems-research.md`
-* (Mem0, Hindsight, Letta, Mastra Observational Memory).
-*/
-var ExtractedMemorySchema = object({
-	type: _enum([
-		"decision",
-		"pattern",
-		"learning",
-		"constraint",
-		"correction"
-	]),
-	content: string().min(1).max(500),
-	importance: number().min(0).max(1),
-	entities: array(string()).max(20),
-	justification: string().min(1).max(200)
-});
-var ExtractionResponseSchema = object({ memories: array(ExtractedMemorySchema).max(20) });
-var SYSTEM_PROMPT = `You are extracting durable knowledge from a coding session transcript.
-Extract ONLY high-value items. Reject noise (greetings, meta-commentary, transient state).
-
-For each extraction provide:
-- type: decision|pattern|learning|constraint|correction
-- content: the knowledge itself (<=500 chars, declarative form)
-- importance: 0.0-1.0 (only >=0.6 will be stored)
-- entities: code symbols, files, concepts mentioned
-- justification: why this is worth remembering (<=200 chars)
-
-Type definitions:
-- decision: architectural or design choice with rationale ("We chose X because Y")
-- pattern: recurring approach that works ("When X, do Y because Z")
-- learning: factual knowledge gained ("X depends on Y via Z")
-- constraint: rule/limitation discovered ("X must always Y")
-- correction: anti-pattern to avoid ("Avoid X because Y; use Z instead")
-
-Rules:
-- Capture the WHY, not just the WHAT
-- If nothing of durable value exists in the transcript, return an empty array
-- Prefer fewer high-quality extractions over many low-quality ones
-- Ignore: tool output noise, routine status messages, file reads without a finding
-- Maximum 7 extractions per call`;
-function buildUserPrompt(sessionId, transcript, maxExtractions) {
-	return `Session ID: ${sessionId}
-
-Transcript:
-${transcript}
-
-Extract up to ${maxExtractions} high-value memories. Return empty array if nothing valuable.`;
-}
-/**
-* Trim a transcript to a maximum character budget, preserving the head and tail
-* (the ends typically contain the request and the final outcome).
-*/
-function clipTranscript(transcript, maxChars) {
-	if (transcript.length <= maxChars) return transcript;
-	const half = Math.floor(maxChars / 2) - 64;
-	const head = transcript.slice(0, half);
-	const tail = transcript.slice(-half);
-	return `${head}\n\n[... ${transcript.length - maxChars} chars omitted ...]\n\n${tail}`;
-}
-/**
-* Store a single LLM-extracted memory, routing it through the full dedup gate.
-*
-* T736: All memory types now pass through the hash-dedup gate before reaching
-* the underlying store function. For decisions (which have domain-specific fields
-* that `verifyAndStore` cannot set), we run `checkHashDedup` first and call
-* `storeDecision` only when no duplicate is found. For all other types, we
-* build a `MemoryCandidate` and delegate to `verifyAndStore` which runs both
-* hash-dedup (Check A) and the full embedding + confidence gate (Checks B, C).
-*
-* Routing:
-*   - decision    → checkHashDedup(brain_decisions) → storeDecision (domain fields preserved)
-*   - pattern     → verifyAndStore (procedural candidate → brain_patterns)
-*   - learning    → verifyAndStore (semantic candidate → brain_learnings)
-*   - constraint  → verifyAndStore (semantic candidate, higher confidence → brain_learnings)
-*   - correction  → verifyAndStore (procedural candidate → brain_patterns)
-*/
-async function storeExtracted(projectRoot, sessionId, memory) {
-	if (memory.type === "decision") {
-		const dedupResult = await checkHashDedup(projectRoot, memory.content, "brain_decisions");
-		if (dedupResult.matched) {
-			const { getBrainNativeDb, getBrainDb } = await import('./memory-sqlite2-B7dq8doO.js');
-			getBrainDb(projectRoot).then(() => {
-				getBrainNativeDb(projectRoot)?.prepare("UPDATE brain_decisions SET citation_count = citation_count + 1 WHERE id = ?").run(dedupResult.id);
-			}).catch(() => void 0);
-			return "merged";
-		}
-		const { storeDecision } = await import('./decisions2-XQsSYdxa.js');
-		try {
-			const { decisionText, rationale } = splitDecisionContent(memory.content);
-			await storeDecision(projectRoot, {
-				type: "technical",
-				decision: decisionText,
-				rationale,
-				confidence: mapImportanceToConfidence(memory.importance)
-			});
-			return "stored";
-		} catch {
-			return "rejected";
-		}
-	}
-	const confidence = memory.type === "constraint" ? Math.max(memory.importance, .8) : memory.importance;
-	let candidate;
-	if (memory.type === "correction" || memory.type === "pattern") candidate = {
-		text: memory.content,
-		title: `${memory.type === "correction" ? "Correction" : "Pattern"}: ${memory.content.slice(0, 80)}`,
-		memoryType: "procedural",
-		tier: "medium",
-		confidence,
-		source: "transcript",
-		sourceSessionId: sessionId,
-		sourceConfidence: "agent"
-	};
-	else candidate = {
-		text: memory.content,
-		title: `${memory.type === "constraint" ? "Constraint" : "Learning"}: ${memory.content.slice(0, 80)}`,
-		memoryType: "semantic",
-		tier: "medium",
-		confidence,
-		source: "transcript",
-		sourceSessionId: sessionId,
-		sourceConfidence: "agent"
-	};
-	try {
-		const result = await verifyAndStore(projectRoot, candidate);
-		if (result.action === "stored") return "stored";
-		if (result.action === "merged") return "merged";
-		return "rejected";
-	} catch {
-		return "rejected";
-	}
-}
-/**
-* Split a decision extraction into decision + rationale at the first " because "
-* marker. When no marker exists, the full content becomes the decision and the
-* justification becomes the rationale.
-*/
-function splitDecisionContent(content) {
-	const idx = content.toLowerCase().indexOf(" because ");
-	if (idx === -1) return {
-		decisionText: content.trim(),
-		rationale: "See justification"
-	};
-	return {
-		decisionText: content.slice(0, idx).trim(),
-		rationale: content.slice(idx + 1).trim()
-	};
-}
-/**
-* Map an importance score to a BrainDecisionRow confidence bucket.
-* Importance is continuous 0–1; confidence is `low` | `medium` | `high`.
-*/
-function mapImportanceToConfidence(importance) {
-	if (importance >= .8) return "high";
-	if (importance >= .6) return "medium";
-	return "low";
-}
-/**
-* Resolve provider+model+credentials via `resolveAnthropicForRole('extraction')`
-* (T-LLM-CRED Phase 2 DRY review P2-1) and return the wired SDK client +
-* selected model name.
-*
-* Honors `cred.authType` so OAuth tokens (Claude Code) are routed through
-* `Authorization: Bearer` instead of `x-api-key` — the registry's
-* `clientForModelConfig` handles this internally via the `extraHeaders`
-* channel set by the resolver.
-* Returns null when no Anthropic credential is available (helper returns
-* null on both missing-credential and wrong-provider paths).
-*/
-async function buildExtractionClient(projectRoot) {
-	const llm = await resolveAnthropicForRole("extraction", { projectRoot });
-	if (!llm) return null;
-	return {
-		client: llm.client,
-		model: llm.model
-	};
-}
-/**
-* Build the zodOutputFormat helper asynchronously (dynamic import matches
-* SDK's subpath export style). Returns null when the SDK cannot be loaded.
-*/
-async function buildZodFormat(schema) {
-	try {
-		return (await import('./zod-DsAeKNLm.js')).zodOutputFormat(schema);
-	} catch {
-		return null;
-	}
-}
-/**
-* Extract and persist structured memories from a session transcript using
-* an Anthropic model.
-*
-* This is the sole entry point replacing the legacy keyword-regex extractor.
-* Callers (e.g. the session-end hook) invoke this and the function handles:
-*   1. Config lookup (brain.llmExtraction.enabled, model, thresholds)
-*   2. API key detection (skips extraction if missing)
-*   3. Transcript clipping to stay within token budget
-*   4. Structured-output LLM call with the research-approved prompt
-*   5. Importance filtering (below minImportance → dropped)
-*   6. Storage routing through the existing verify-and-store pipeline
-*
-* Never throws. Returns an ExtractionReport describing what was stored.
-*
-* @example
-* ```ts
-* await extractFromTranscript({
-*   projectRoot: '/path/to/project',
-*   sessionId: 'ses_20260413192026_519188',
-*   transcript: rawTranscript,
-* });
-* ```
-*/
-async function extractFromTranscript(options) {
-	const report = {
-		extractedCount: 0,
-		storedCount: 0,
-		mergedCount: 0,
-		rejectedCount: 0,
-		warnings: []
-	};
-	const { projectRoot, sessionId, transcript } = options;
-	if (typeof transcript !== "string" || transcript.trim().length === 0) return report;
-	const { loadConfig } = await import('./config2-F6PPIn7D.js');
-	let config;
-	try {
-		config = await loadConfig(projectRoot);
-	} catch {
-		report.warnings.push("config load failed, using defaults");
-		config = void 0;
-	}
-	const llmCfg = config?.brain?.llmExtraction;
-	if (!(llmCfg?.enabled ?? true)) {
-		report.warnings.push("llmExtraction.enabled=false");
-		return report;
-	}
-	const minImportance = llmCfg?.minImportance ?? .6;
-	const maxExtractions = llmCfg?.maxExtractions ?? 7;
-	const maxTranscriptChars = llmCfg?.maxTranscriptChars ?? 6e4;
-	const resolved = options.client ? null : await buildExtractionClient(projectRoot);
-	const client = options.client ?? resolved?.client ?? null;
-	if (!client) {
-		report.warnings.push("No Anthropic API key found (checked ANTHROPIC_API_KEY env and ~/.claude/.credentials.json) — extraction skipped");
-		return report;
-	}
-	const model = llmCfg?.model ?? resolved?.model ?? "claude-haiku-4-5-20251001";
-	const userPrompt = buildUserPrompt(sessionId, clipTranscript(transcript, maxTranscriptChars), maxExtractions);
-	let extracted = [];
-	try {
-		const format = await buildZodFormat(ExtractionResponseSchema);
-		if (format) extracted = (await client.messages.parse({
-			model,
-			max_tokens: 4096,
-			system: SYSTEM_PROMPT,
-			messages: [{
-				role: "user",
-				content: userPrompt
-			}],
-			output_config: { format }
-		})).parsed_output?.memories ?? [];
-		else extracted = await extractViaPlainCall(client, model, userPrompt);
-	} catch (err) {
-		const message = err instanceof Error ? err.message : String(err);
-		report.warnings.push(`extraction call failed: ${message}`);
-		return report;
-	}
-	report.extractedCount = extracted.length;
-	for (const memory of extracted) {
-		if (memory.importance < minImportance) {
-			report.rejectedCount += 1;
-			continue;
-		}
-		try {
-			const outcome = await storeExtracted(projectRoot, sessionId, memory);
-			if (outcome === "stored") report.storedCount += 1;
-			else if (outcome === "merged") report.mergedCount += 1;
-			else report.rejectedCount += 1;
-		} catch (err) {
-			const message = err instanceof Error ? err.message : String(err);
-			report.warnings.push(`store failed: ${message}`);
-			report.rejectedCount += 1;
-		}
-	}
-	return report;
-}
-/**
-* Fallback extraction path: calls messages.create and parses the response
-* content manually. Used when the zod helper cannot be loaded (e.g. SDK
-* version mismatch) but the messages.create API is still available.
-*/
-async function extractViaPlainCall(client, model, userPrompt) {
-	const text = extractTextFromResponse(await client.messages.create({
-		model,
-		max_tokens: 4096,
-		system: `${SYSTEM_PROMPT}\n\nReturn ONLY a JSON object of shape {"memories": [...]} with no prose.`,
-		messages: [{
-			role: "user",
-			content: userPrompt
-		}]
-	}));
-	if (!text) return [];
-	const parsed = safeJsonParse(text);
-	if (!parsed) return [];
-	const validated = ExtractionResponseSchema.safeParse(parsed);
-	if (!validated.success) return [];
-	return validated.data.memories;
-}
-/**
-* Pull plain text from a messages.create response. Anthropic returns a
-* content array; we concatenate all text blocks.
-*/
-function extractTextFromResponse(response) {
-	const content = response?.content;
-	if (!Array.isArray(content)) return "";
-	return content.filter((block) => block?.type === "text" && typeof block.text === "string").map((block) => block.text ?? "").join("");
-}
-/**
-* Parse JSON, extracting the first {...} object even if the model wrapped
-* it in markdown fences or added prose. Returns null on any failure.
-*/
-function safeJsonParse(text) {
-	const trimmed = text.trim();
-	try {
-		return JSON.parse(trimmed);
-	} catch {}
-	const first = trimmed.indexOf("{");
-	const last = trimmed.lastIndexOf("}");
-	if (first === -1 || last === -1 || last < first) return null;
-	try {
-		return JSON.parse(trimmed.slice(first, last + 1));
-	} catch {
-		return null;
-	}
-}
-
-export { extractFromTranscript };
-//# sourceMappingURL=llm-extraction-DVCu4heR.js.map

package/studio-dist/server/chunks/loadSharedConfigFiles-umvW7kuz.js

@@ -1,73 +0,0 @@
-import { I as IniSectionType } from './role-executor-CC8BU54L.js';
-import { g as getHomeDir, r as readFile$1 } from './readFile-CxwXhymG.js';
-import { join } from 'node:path';
-
-//#region ../../node_modules/.pnpm/@smithy+core@3.24.3/node_modules/@smithy/core/dist-es/submodules/config/shared-ini-file-loader/getConfigData.js
-var getConfigData = (data) => Object.entries(data).filter(([key]) => {
-	const indexOfSeparator = key.indexOf(".");
-	if (indexOfSeparator === -1) return false;
-	return Object.values(IniSectionType).includes(key.substring(0, indexOfSeparator));
-}).reduce((acc, [key, value]) => {
-	const indexOfSeparator = key.indexOf(".");
-	const updatedKey = key.substring(0, indexOfSeparator) === IniSectionType.PROFILE ? key.substring(indexOfSeparator + 1) : key;
-	acc[updatedKey] = value;
-	return acc;
-}, { ...data.default && { default: data.default } });
-var getConfigFilepath = () => process.env["AWS_CONFIG_FILE"] || join(getHomeDir(), ".aws", "config");
-var getCredentialsFilepath = () => process.env["AWS_SHARED_CREDENTIALS_FILE"] || join(getHomeDir(), ".aws", "credentials");
-//#endregion
-//#region ../../node_modules/.pnpm/@smithy+core@3.24.3/node_modules/@smithy/core/dist-es/submodules/config/shared-ini-file-loader/parseIni.js
-var prefixKeyRegex = /^([\w-]+)\s(["'])?([\w-@\+\.%:/]+)\2$/;
-var profileNameBlockList = ["__proto__", "profile __proto__"];
-var parseIni = (iniData) => {
-	const map = {};
-	let currentSection;
-	let currentSubSection;
-	for (const iniLine of iniData.split(/\r?\n/)) {
-		const trimmedLine = iniLine.split(/(^|\s)[;#]/)[0].trim();
-		if (trimmedLine[0] === "[" && trimmedLine[trimmedLine.length - 1] === "]") {
-			currentSection = void 0;
-			currentSubSection = void 0;
-			const sectionName = trimmedLine.substring(1, trimmedLine.length - 1);
-			const matches = prefixKeyRegex.exec(sectionName);
-			if (matches) {
-				const [, prefix, , name] = matches;
-				if (Object.values(IniSectionType).includes(prefix)) currentSection = [prefix, name].join(".");
-			} else currentSection = sectionName;
-			if (profileNameBlockList.includes(sectionName)) throw new Error(`Found invalid profile name "${sectionName}"`);
-		} else if (currentSection) {
-			const indexOfEqualsSign = trimmedLine.indexOf("=");
-			if (![0, -1].includes(indexOfEqualsSign)) {
-				const [name, value] = [trimmedLine.substring(0, indexOfEqualsSign).trim(), trimmedLine.substring(indexOfEqualsSign + 1).trim()];
-				if (value === "") currentSubSection = name;
-				else {
-					if (currentSubSection && iniLine.trimStart() === iniLine) currentSubSection = void 0;
-					map[currentSection] = map[currentSection] || {};
-					const key = currentSubSection ? [currentSubSection, name].join(".") : name;
-					map[currentSection][key] = value;
-				}
-			}
-		}
-	}
-	return map;
-};
-//#endregion
-//#region ../../node_modules/.pnpm/@smithy+core@3.24.3/node_modules/@smithy/core/dist-es/submodules/config/shared-ini-file-loader/loadSharedConfigFiles.js
-var swallowError = () => ({});
-var loadSharedConfigFiles = async (init = {}) => {
-	const { filepath = getCredentialsFilepath(), configFilepath = getConfigFilepath() } = init;
-	const homeDir = getHomeDir();
-	const relativeHomeDirPrefix = "~/";
-	let resolvedFilepath = filepath;
-	if (filepath.startsWith(relativeHomeDirPrefix)) resolvedFilepath = join(homeDir, filepath.slice(2));
-	let resolvedConfigFilepath = configFilepath;
-	if (configFilepath.startsWith(relativeHomeDirPrefix)) resolvedConfigFilepath = join(homeDir, configFilepath.slice(2));
-	const parsedFiles = await Promise.all([readFile$1(resolvedConfigFilepath, { ignoreCache: init.ignoreCache }).then(parseIni).then(getConfigData).catch(swallowError), readFile$1(resolvedFilepath, { ignoreCache: init.ignoreCache }).then(parseIni).catch(swallowError)]);
-	return {
-		configFile: parsedFiles[0],
-		credentialsFile: parsedFiles[1]
-	};
-};
-
-export { getConfigFilepath as g, loadSharedConfigFiles as l, parseIni as p };
-//# sourceMappingURL=loadSharedConfigFiles-umvW7kuz.js.map