llm-kb 0.4.0 → 0.4.2

package/plan.md

@@ -1,300 +0,0 @@
-# llm-kb — Phase 1 Build Plan
-
-> Emergent design. Each slice is a thin vertical slice that works end-to-end, is demoable, and informs the next step. Decisions are made at the last responsible moment.
-
-## Key Learnings
-- **PDF is the only adapter we build.** Everything else (Excel, Word, PPT, CSV, images) handled dynamically by Pi SDK agent at query time.
-- **`@llamaindex/liteparse`** proven (from parser-study). Extracts text + bounding boxes locally.
-- **Two-output pattern**: `.md` (spatial text) + `.json` (bounding boxes for citations).
-- **OCR off by default.** Most PDFs have native text. Enable via `OCR_SERVER_URL` or `OCR_ENABLED=true`.
-- **Pi SDK `createAgentSession()`** with defaults — no auth/model config needed. Uses Pi's existing auth.
-- **AGENTS.md injected via `agentsFilesOverride`** — user's folder stays clean.
-- **NODE_PATH** set so agent's bash scripts can use bundled libraries (exceljs, mammoth, officeparser).
-- **Config file skipped** — nothing reads it yet. Add when Phase 2/3 needs it.
-
----
-
-## Slice 1: "Hello World" CLI ✅
-Commander CLI with `run <folder>`. Scans folder, lists files by extension.
-
-## Slice 2: PDF → markdown + bounding boxes ✅
-LiteParse parses PDFs → `.md` + `.json` in `.llm-kb/wiki/sources/`. Tested on 9 real PDFs (1000+ pages).
-
-## Slice 3: Scanned PDF handling (OCR) ✅
-LiteParse has Tesseract.js built-in. `ocrEnabled` + `ocrServerUrl` config. OCR off by default. Azure OCR bridge tested on 16 legal PDFs (3000+ pages).
-
-## Slice 4: Progress + error handling ✅
-Inline progress. Stderr suppression. Corrupt file skip + warning. Mtime check — re-runs instant.
-
-## Slice 5: Indexer (Pi SDK) ✅
-`createAgentSession` with cwd = user's folder. AGENTS.md injected. Agent reads sources, writes `index.md` with summary table.
-
-## Slice 6: File watcher ✅
-chokidar watches folder. New/changed PDFs → parse → re-index. 2s debounce for batch drops.
-
-## Slice 7: Config + polish → Skipped
-Config file has no readers yet. Deferred to Phase 2/3. README updated instead.
-
----
-
-## Phase 1 Complete ✅
-
-**What ships:**
-- `llm-kb run ./folder` — scan, parse PDFs, build index, watch for new files
-- Pre-bundled libraries for agent to handle Excel, Word, PowerPoint at query time
-- OCR via env var (local Tesseract or remote Azure bridge)
-- Auth via Pi SDK (zero config)
-
-**Phase 2 complete ✅:**
-- `llm-kb query "question"` — auto-detects KB, streams cited answers
-- `--save` flag — research mode, saves to `outputs/`, re-indexes
-- Query mode is read-only (read tool only). Research mode adds bash + write.
-
----
-
-## Phase 3: Auth Fix + Eval Loop + LLM Config
-
-> Full spec: `PHASE3_SPEC.md`
-> **Trigger:** 182 LinkedIn saves — people coming back to try `npx llm-kb run` this week.
-
-### Why now
-- **Auth fix is urgent.** Users hit a wall if Pi SDK isn't installed. `ANTHROPIC_API_KEY` must work as fallback.
-- **Eval loop is the differentiator.** Nobody else building llm-wiki scripts has traces + citation checking.
-- **Model config cuts cost.** Haiku for indexing (10x cheaper than Sonnet). Users shouldn't pay Sonnet prices for one-line summaries.
-
-### Build order
-
-## Slice 1: Auth fix 🔴 DO FIRST
-
-Check auth before creating any session. Support two paths:
-
-```
-npx llm-kb run ./docs
-  ├─ ~/.pi/agent/auth.json exists? → Pi SDK auth. Done.
-  ├─ ANTHROPIC_API_KEY set? → Configure Pi SDK programmatically. Done.
-  └─ Neither? → Clear error:
-       No LLM authentication found.
-       Option 1: npm install -g @mariozechner/pi-coding-agent && pi
-       Option 2: export ANTHROPIC_API_KEY=sk-ant-...
-```
-
-Definition of done:
-- [ ] `ANTHROPIC_API_KEY=sk-... npx llm-kb run ./docs` works without Pi installed
-- [ ] Pi SDK auth works as before (no regression)
-- [ ] Clear error when neither is available
-- [ ] README updated with both auth options
-
-## Slice 2: LLM config
-
-Auto-generate `.llm-kb/config.json` on first run:
-
-```json
-{
-  "indexModel": "claude-haiku-3-5",
-  "queryModel": "claude-sonnet-4-20250514",
-  "provider": "anthropic"
-}
-```
-
-Env var overrides: `LLM_KB_INDEX_MODEL`, `LLM_KB_QUERY_MODEL`
-Priority: env var → config file → defaults
-
-Definition of done:
-- [ ] Config auto-generated on first run
-- [ ] Haiku for indexing, Sonnet for query by default
-- [ ] Env vars override config
-- [ ] `llm-kb status` shows current model config
-
-## Slice 3: Trace logging
-
-Every query logs JSON to `.llm-kb/traces/`:
-
-```json
-{
-  "id": "2026-04-06T14-30-00-query",
-  "timestamp": "2026-04-06T14:30:00Z",
-  "question": "what are the reserve requirements?",
-  "mode": "query",
-  "filesRead": ["index.md", "reserve-policy.md"],
-  "filesSkipped": ["board-deck.md", "pipeline.md"],
-  "answer": "Reserve requirements are defined in...",
-  "citations": [
-    { "file": "reserve-policy.md", "location": "p.3", "claim": "Minimum reserve ratio of 12%" }
-  ],
-  "durationMs": 4200
-}
-```
-
-Capture via `session.subscribe()` — intercept tool calls to track `filesRead`.
-
-Definition of done:
-- [ ] Every query writes a trace JSON to `.llm-kb/traces/`
-- [ ] Trace includes: question, mode, filesRead, filesSkipped, answer, citations, durationMs
-
-## Slice 4: `llm-kb status`
-
-```
-Knowledge Base: ./research/.llm-kb/
-  Sources: 12 files (8 PDF, 2 XLSX, 1 DOCX, 1 TXT)
-  Index: 12 entries, last updated 2 min ago
-  Outputs: 3 saved research answers
-  Traces: 47 queries logged
-  Model: claude-sonnet-4 (query), claude-haiku-3-5 (index)
-  Auth: Pi SDK
-```
-
-Definition of done:
-- [ ] `llm-kb status` prints KB stats, auth method, model config, trace count
-
-## Slice 5: `llm-kb eval`
-
-Pi SDK session (read-only) that:
-1. Reads trace files from `.llm-kb/traces/`
-2. For each trace checks: citation validity, missing sources, answer consistency
-3. Writes report to `.llm-kb/wiki/outputs/eval-report.md`
-4. Watcher detects report, re-indexes
-
-```bash
-llm-kb eval --folder ./research
-llm-kb eval --folder ./research --last 20
-```
-
-Definition of done:
-- [ ] `llm-kb eval` reads traces and writes eval-report.md
-- [ ] Flags: invalid citations, skipped-but-relevant files, answer contradictions
-- [ ] Report gets re-indexed by watcher
-
-## Slice 6: Blog Part 4
-
-- Write AFTER `llm-kb eval` runs on real data
-- Show actual eval-report.md output
-- Title: "How llm-kb Knows When It Got It Wrong"
-
----
-
-## Phase 3 Definition of Done
-
-- [ ] `ANTHROPIC_API_KEY` works without Pi SDK installed
-- [ ] Clear error when no auth found
-- [ ] Config file with model selection (index vs query model)
-- [ ] Every query logs a trace JSON to `.llm-kb/traces/`
-- [ ] `llm-kb eval` checks citations and writes report
-- [ ] `llm-kb status` shows KB stats + config + auth
-- [ ] README updated with auth options + eval command
-- [ ] Blog Part 4 written with real eval output
-
----
-
-## Phase 4: Wiki Compilation (The Farzapedia Pattern)
-
-> **Insight:** Farza's implementation is what Karpathy called the best version. The key isn't just indexing sources — it's compiling them into a **concept-organized wiki** with backlinked articles. The agent navigates concepts, not raw files.
->
-> Current llm-kb: agent reads source summaries → picks source files → answers from raw docs.
-> Wiki pattern: agent reads concept articles → drills into specific articles → answers from synthesized knowledge.
-
-### The Structural Shift
-
-**Current (source-organized):**
-```
-sources/lease-castlelake.md    ─┐
-sources/lease-genesis.md       ─┤→ index.md (flat table, one row per source)
-sources/maintenance-manual.md  ─┘
-```
-
-**Wiki (concept-organized):**
-```
-sources/lease-castlelake.md    ─┐
-sources/lease-genesis.md       ─┤→ articles/reserve-requirements.md
-sources/maintenance-manual.md  ─┘   articles/key-parties.md
-                                      articles/engine-types.md
-                                      articles/maintenance-obligations.md
-                                      index.md (catalog of ARTICLES with backlinks)
-```
-
-Farza's quote: *"The structure of the wiki files and how it's all backlinked is very easily crawlable by any agent. Starting at index.md, the agent does a really good job at drilling into the specific pages it needs."*
-
-Karpathy's quote: *"I thought I had to reach for fancy RAG, but the LLM has been pretty good about auto-maintaining index files and brief summaries of all the documents and it reads all the important related data fairly easily at this ~small scale."*
-
-### New Command
-
-```bash
-llm-kb compile ./folder
-```
-
-Or auto-triggered after indexing when `--wiki` flag is passed:
-
-```bash
-llm-kb run ./folder --wiki   # parse + index + compile wiki
-```
-
-### Wiki Directory Structure
-
-```
-.llm-kb/
-  wiki/
-    sources/          # raw parsed files (unchanged)
-    articles/         # NEW: concept-organized articles
-      concepts/       # thematic articles (reserve-requirements.md, etc)
-      entities/       # people, companies, assets
-      timeline.md     # chronological view of events
-    index.md          # NOW points to articles, not sources
-    source-index.md   # OLD flat source table (kept for reference)
-```
-
-### How Compile Works
-
-1. LLM reads ALL source files (via existing sources/ dir)
-2. LLM identifies key concepts, entities, and themes
-3. LLM writes one markdown article per concept — like a Wikipedia entry
-4. Each article has:
-   - Description (synthesized from multiple sources)
-   - Backlinks to related articles: `[[reserve-requirements]]`, `[[key-parties]]`
-   - Source citations: `(Source: lease-castlelake.md, p.3)`
-5. Writes `articles/index.md` — catalog of all articles with one-line descriptions
-
-### Incremental Updates (The Compounding Part)
-
-When a new source is added (watcher detects it):
-1. Parse it → sources/
-2. LLM reads new source + `articles/index.md`
-3. LLM updates 2-3 most relevant existing articles where new content fits
-4. OR creates a new article if the topic is genuinely new
-5. Updates `articles/index.md` catalog
-
-Farza: *"The most magical thing now is as I add new things, the system updates 2-3 different articles where it feels the context belongs, or just creates a new article. Like a super genius librarian."*
-
-### Query Change
-
-With wiki mode active, query agent reads `articles/index.md` (concept catalog) instead of `source-index.md` (source table). Agent drills into concept articles, gets synthesized cross-referenced answers.
-
-Fallback: if no articles compiled yet, falls back to source-index behavior.
-
-### Blog Post
-
-Part 5 or 6: *"Why Your Knowledge Base Needs a Librarian, Not a Search Engine"*
-- Show the before/after: flat source index vs concept wiki
-- Show Farza's insight: built for agents, not humans
-- Show the compounding: add one doc, 3 articles updated automatically
-- Live demo: `llm-kb compile` on a real document set
-
-### Why This Matters for llm-kb
-
-- Karpathy publicly called this the best implementation of his pattern
-- Farza has **no public code** — llm-kb ships it as `npx llm-kb compile`
-- Farzapedia got **920K views** on Farza's tweet, **Karpathy quote-tweeted** it at **920K views**
-- First open-source implementation of the exact pattern Karpathy validated
-- Blog post can reference both tweets legitimately ("inspired by Farza's Farzapedia")
-
-### Definition of Done (Phase 4)
-
-- [ ] `llm-kb compile ./folder` reads sources and writes articles/ directory
-- [ ] Each article is a proper markdown file with backlinks and source citations
-- [ ] `articles/index.md` is a concept catalog (not a source table)
-- [ ] Watcher triggers incremental article updates when new source added
-- [ ] Query uses article index when available, falls back to source index
-- [ ] `llm-kb status` shows: X sources, Y articles, last compiled
-- [ ] Blog Part 5 written: concept wiki vs flat index, Farzapedia reference
-
----
-
-*Phase 4 added April 6, 2026 — inspired by Farzapedia (@FarzaTV), highlighted by Karpathy as best implementation of the LLM wiki pattern. 920K views on Farza's tweet, Karpathy quote-tweet at 920K views.*

package/src/auth.ts

@@ -1,55 +0,0 @@
-import { existsSync } from "node:fs";
-import { join } from "node:path";
-import { homedir } from "node:os";
-import { AuthStorage } from "@mariozechner/pi-coding-agent";
-import chalk from "chalk";
-
-export type AuthMethod = "pi-sdk" | "api-key" | "none";
-
-export interface AuthResult {
-  ok: true;
-  method: AuthMethod;
-  authStorage?: AuthStorage; // undefined = use Pi SDK default (file-based)
-}
-
-export interface AuthFailure {
-  ok: false;
-}
-
-/**
- * Detect available auth and return an AuthStorage if needed.
- *
- * Priority:
- * 1. Pi SDK auth.json  → use default file-based storage (no override needed)
- * 2. ANTHROPIC_API_KEY → create in-memory storage with the key
- * 3. Neither           → return failure
- */
-export function checkAuth(): AuthResult | AuthFailure {
-  const piAuthPath = join(homedir(), ".pi", "agent", "auth.json");
-
-  if (existsSync(piAuthPath)) {
-    return { ok: true, method: "pi-sdk" };
-  }
-
-  if (process.env.ANTHROPIC_API_KEY) {
-    const authStorage = AuthStorage.inMemory({
-      anthropic: { type: "api_key", key: process.env.ANTHROPIC_API_KEY },
-    });
-    return { ok: true, method: "api-key", authStorage };
-  }
-
-  return { ok: false };
-}
-
-/**
- * Print a friendly error when no auth is configured and exit.
- */
-export function exitWithAuthError(): never {
-  console.error(chalk.red("\n  No LLM authentication found.\n"));
-  console.error(`  ${chalk.bold("Option 1:")} Install Pi SDK ${chalk.dim("(recommended)")}`);
-  console.error(chalk.dim("    npm install -g @mariozechner/pi-coding-agent"));
-  console.error(chalk.dim("    pi\n"));
-  console.error(`  ${chalk.bold("Option 2:")} Set your Anthropic API key`);
-  console.error(chalk.dim("    export ANTHROPIC_API_KEY=sk-ant-...\n"));
-  process.exit(1);
-}

package/src/cli.ts

@@ -1,257 +0,0 @@
-#!/usr/bin/env node
-
-import { Command } from "commander";
-import { scan, summarize } from "./scan.js";
-import { parsePDF } from "./pdf.js";
-import { buildIndex } from "./indexer.js";
-import { startWatcher } from "./watcher.js";
-import { startSessionWatcher } from "./session-watcher.js";
-import { query, createChat } from "./query.js";
-import { runEval } from "./eval.js";
-import { ChatDisplay } from "./tui-display.js";
-import { resolveKnowledgeBase } from "./resolve-kb.js";
-import { checkAuth, exitWithAuthError } from "./auth.js";
-import { ensureConfig, loadConfig } from "./config.js";
-import { existsSync } from "node:fs";
-import { mkdir, readdir, stat } from "node:fs/promises";
-import { resolve, join } from "node:path";
-import chalk from "chalk";
-
-const program = new Command();
-
-program
-  .name("llm-kb")
-  .description("Drop files into a folder. Get a knowledge base you can query.")
-  .version("0.4.0");
-
-program
-  .command("run")
-  .description("Scan, parse, index, and watch a folder")
-  .argument("<folder>", "Path to your documents folder")
-  .action(async (folder: string) => {
-    console.log(`\n${chalk.bold("llm-kb")} v0.4.0\n`);
-
-    const auth = checkAuth();
-    if (!auth.ok) exitWithAuthError();
-
-    if (!existsSync(folder)) {
-      console.error(chalk.red(`Error: Folder not found: ${folder}`));
-      process.exit(1);
-    }
-
-    const root = resolve(folder);
-    const config = await ensureConfig(root);
-
-    console.log(`Scanning ${folder}...`);
-
-    const files = await scan(folder);
-
-    if (files.length === 0) {
-      console.log(chalk.yellow("  No supported files found."));
-      return;
-    }
-
-    const pdfs = files.filter((f) => f.ext === ".pdf");
-    console.log(`  Found ${chalk.bold(files.length.toString())} files (${summarize(files)})`);
-    if (pdfs.length === 0) return;
-
-    const sourcesDir = join(root, ".llm-kb", "wiki", "sources");
-    await mkdir(sourcesDir, { recursive: true });
-
-    // Parse PDFs
-    let parsed = 0;
-    let skipped = 0;
-    let failed = 0;
-    const errors: { name: string; message: string }[] = [];
-
-    for (let i = 0; i < pdfs.length; i++) {
-      const pdf = pdfs[i];
-      const progress = `  Parsing... ${i + 1}/${pdfs.length} \u2014 ${pdf.name}`;
-      process.stdout.write(`\r${progress.padEnd(process.stdout.columns || 80)}`);
-      try {
-        const result = await parsePDF(join(root, pdf.path), sourcesDir);
-        if (result.skipped) skipped++; else parsed++;
-      } catch (err: any) {
-        failed++;
-        errors.push({ name: pdf.name, message: err.message });
-      }
-    }
-
-    process.stdout.write(`\r${"".padEnd(process.stdout.columns || 80)}\r`);
-
-    const parts: string[] = [];
-    if (parsed > 0) parts.push(chalk.green(`${parsed} parsed`));
-    if (skipped > 0) parts.push(chalk.dim(`${skipped} skipped (up to date)`));
-    if (failed > 0) parts.push(chalk.red(`${failed} failed`));
-    console.log(`  ${parts.join(", ")}`);
-    for (const err of errors) console.log(chalk.red(`    \u2717 ${err.name} \u2014 ${err.message}`));
-
-    // Build index — skip if up to date
-    const indexFile = join(root, ".llm-kb", "wiki", "index.md");
-    let indexUpToDate = false;
-    if (parsed === 0 && existsSync(indexFile)) {
-      try {
-        const indexMtime = (await stat(indexFile)).mtimeMs;
-        const sourceFiles = await readdir(sourcesDir);
-        const mtimes = await Promise.all(sourceFiles.map((f) => stat(join(sourcesDir, f)).then((s) => s.mtimeMs)));
-        indexUpToDate = mtimes.every((mt) => indexMtime >= mt);
-      } catch {}
-    }
-
-    if (indexUpToDate) {
-      console.log(chalk.dim(`\n  Index up to date.`));
-    } else {
-      console.log(`\n  Building index... ${chalk.dim(`(${config.indexModel})`)}`);
-      try {
-        await buildIndex(root, sourcesDir, undefined, auth.authStorage, config.indexModel);
-        console.log(chalk.green(`  Index built: .llm-kb/wiki/index.md`));
-      } catch (err: any) {
-        console.error(chalk.red(`  Index failed: ${err.message}`));
-      }
-    }
-
-    console.log(`\n  ${chalk.dim("Output:")} ${sourcesDir}`);
-
-    // Start watchers
-    startWatcher({ folder: root, sourcesDir, authStorage: auth.authStorage, indexModel: config.indexModel });
-    startSessionWatcher(root);
-
-    // TUI chat
-    const chatUI = new ChatDisplay();
-    const { session, display } = await createChat(root, {
-      authStorage: auth.authStorage,
-      modelId: config.queryModel,
-      tuiDisplay: chatUI,
-    });
-
-    chatUI.onSubmit = (text) => {
-      display.setQuestion(text);
-      // Fire-and-forget — don't await, TUI must stay responsive
-      session.prompt(text).catch(() => {});
-    };
-
-    chatUI.onExit = () => {
-      display.flush().then(() => {
-        session.dispose();
-        process.exit(0);
-      });
-    };
-
-    console.log(`\n${chalk.bold("Ready.")} Ask a question or drop files in to re-index.\n`);
-    chatUI.start();
-  });
-
-program
-  .command("query")
-  .description("Ask a single question (non-interactive, stdout)")
-  .argument("<question>", "Your question")
-  .option("--folder <path>", "Path to document folder (auto-detects if omitted)")
-  .option("--save", "Save the answer to wiki/outputs/ (research mode)")
-  .action(async (question: string, options: { folder?: string; save?: boolean }) => {
-    const auth = checkAuth();
-    if (!auth.ok) exitWithAuthError();
-
-    const root = resolveKnowledgeBase(options.folder || process.cwd());
-    if (!root) {
-      console.error(chalk.red("No knowledge base found. Run 'llm-kb run <folder>' first."));
-      process.exit(1);
-    }
-
-    const config = await loadConfig(root);
-    try {
-      await query(root, question, {
-        save: options.save,
-        authStorage: auth.authStorage,
-        modelId: config.queryModel,
-      });
-    } catch (err: any) {
-      console.error(chalk.red(err.message));
-      process.exit(1);
-    }
-  });
-
-program
-  .command("eval")
-  .description("Analyze sessions for quality issues, wiki gaps, and performance")
-  .option("--folder <path>", "Path to document folder (auto-detects if omitted)")
-  .option("--last <n>", "Only check last N sessions", parseInt)
-  .action(async (options: { folder?: string; last?: number }) => {
-    const auth = checkAuth();
-    if (!auth.ok) exitWithAuthError();
-
-    const root = resolveKnowledgeBase(options.folder || process.cwd());
-    if (!root) {
-      console.error(chalk.red("No knowledge base found. Run 'llm-kb run <folder>' first."));
-      process.exit(1);
-    }
-
-    console.log(`\n${chalk.bold("llm-kb eval")}\n`);
-
-    const result = await runEval(root, {
-      authStorage: auth.authStorage,
-      last: options.last,
-      onProgress: (msg) => console.log(chalk.dim(`  ${msg}`)),
-    });
-
-    const { metrics, issues, wikiGaps } = result;
-    const errors = issues.filter((i) => i.severity === "error").length;
-    const warnings = issues.filter((i) => i.severity === "warning").length;
-
-    console.log();
-    console.log(`  ${chalk.bold("Results:")}`);
-    console.log(`  Queries analyzed:  ${metrics.totalQAs}`);
-    console.log(`  Wiki hit rate:     ${metrics.totalQAs > 0 ? Math.round(metrics.wikiHits / metrics.totalQAs * 100) : 0}%`);
-    console.log(`  Wasted reads:      ${metrics.wastedReads}`);
-    console.log(`  Issues:            ${errors > 0 ? chalk.red(`${errors} errors`) : chalk.green("0 errors")}  ${warnings > 0 ? chalk.yellow(`${warnings} warnings`) : chalk.dim("0 warnings")}`);
-    console.log(`  Wiki gaps:         ${wikiGaps.length > 0 ? chalk.yellow(String(wikiGaps.length)) : chalk.green("0")}`);
-    console.log();
-    console.log(chalk.green(`  Report: .llm-kb/wiki/outputs/eval-report.md`));
-    console.log();
-  });
-
-program
-  .command("status")
-  .description("Show knowledge base stats and current config")
-  .option("--folder <path>", "Path to document folder (auto-detects if omitted)")
-  .action(async (options: { folder?: string }) => {
-    const root = resolveKnowledgeBase(options.folder || process.cwd());
-    if (!root) {
-      console.error(chalk.red("No knowledge base found. Run 'llm-kb run <folder>' first."));
-      process.exit(1);
-    }
-
-    const auth = checkAuth();
-    const config = await loadConfig(root);
-
-    const sourcesDir  = join(root, ".llm-kb", "wiki", "sources");
-    const indexFile   = join(root, ".llm-kb", "wiki", "index.md");
-    const articlesDir = join(root, ".llm-kb", "wiki", "articles");
-    const outputsDir  = join(root, ".llm-kb", "wiki", "outputs");
-
-    let sourceCount = 0;
-    try { sourceCount = (await readdir(sourcesDir)).filter((f) => f.endsWith(".md")).length; } catch {}
-
-    let indexAge = "not built yet";
-    try {
-      const diffMin = Math.round((Date.now() - (await stat(indexFile)).mtimeMs) / 60000);
-      indexAge = diffMin < 1 ? "just now" : diffMin < 60 ? `${diffMin} min ago` : `${Math.round(diffMin / 60)} hr ago`;
-    } catch {}
-
-    let outputCount = 0;
-    try { outputCount = (await readdir(outputsDir)).filter((f) => f.endsWith(".md")).length; } catch {}
-
-    console.log(`\n${chalk.bold("Knowledge Base Status")}`);
-    console.log(`  ${chalk.dim("Folder:")}  ${root}`);
-    console.log(`  ${chalk.dim("Sources:")} ${sourceCount > 0 ? `${sourceCount} parsed source${sourceCount !== 1 ? "s" : ""}` : chalk.yellow("none yet")}`);
-    console.log(`  ${chalk.dim("Index:")}   ${indexAge}`);
-    let articleCount = 0;
-    try { articleCount = (await readdir(articlesDir)).filter((f) => f.endsWith(".md") && f !== "index.md").length; } catch {}
-
-    if (articleCount > 0) console.log(`  ${chalk.dim("Articles:")} ${articleCount} compiled`);
-    if (outputCount > 0) console.log(`  ${chalk.dim("Outputs:")} ${outputCount} saved answer${outputCount !== 1 ? "s" : ""}`);
-    console.log(`  ${chalk.dim("Models:")}  ${chalk.cyan(config.queryModel)} ${chalk.dim("(query)")}  ${chalk.cyan(config.indexModel)} ${chalk.dim("(index)")}`);
-    console.log(`  ${chalk.dim("Auth:")}    ${auth.ok ? (auth.method === "pi-sdk" ? "Pi SDK" : "ANTHROPIC_API_KEY") : chalk.red("not configured")}`);
-    console.log();
-  });
-
-program.parse();

package/src/config.ts

@@ -1,61 +0,0 @@
-import { existsSync } from "node:fs";
-import { readFile, writeFile, mkdir } from "node:fs/promises";
-import { join } from "node:path";
-
-export const DEFAULT_INDEX_MODEL = "claude-haiku-4-5";
-export const DEFAULT_QUERY_MODEL = "claude-sonnet-4-6";
-
-export interface KBConfig {
-  indexModel: string;
-  queryModel: string;
-}
-
-const DEFAULTS: KBConfig = {
-  indexModel: DEFAULT_INDEX_MODEL,
-  queryModel: DEFAULT_QUERY_MODEL,
-};
-
-function configPath(kbRoot: string): string {
-  return join(kbRoot, ".llm-kb", "config.json");
-}
-
-/**
- * Load config from .llm-kb/config.json, applying env var overrides.
- *
- * Priority: env var > config file > defaults
- */
-export async function loadConfig(kbRoot: string): Promise<KBConfig> {
-  let base: KBConfig = { ...DEFAULTS };
-
-  const path = configPath(kbRoot);
-  if (existsSync(path)) {
-    try {
-      const raw = await readFile(path, "utf-8");
-      const parsed = JSON.parse(raw);
-      if (parsed.indexModel) base.indexModel = parsed.indexModel;
-      if (parsed.queryModel) base.queryModel = parsed.queryModel;
-    } catch {
-      // Ignore malformed config — fall back to defaults
-    }
-  }
-
-  if (process.env.LLM_KB_INDEX_MODEL) base.indexModel = process.env.LLM_KB_INDEX_MODEL;
-  if (process.env.LLM_KB_QUERY_MODEL) base.queryModel = process.env.LLM_KB_QUERY_MODEL;
-
-  return base;
-}
-
-/**
- * Create .llm-kb/config.json with defaults if it doesn't exist yet.
- */
-export async function ensureConfig(kbRoot: string): Promise<KBConfig> {
-  const path = configPath(kbRoot);
-
-  if (!existsSync(path)) {
-    await mkdir(join(kbRoot, ".llm-kb"), { recursive: true });
-    await writeFile(path, JSON.stringify(DEFAULTS, null, 2) + "\n", "utf-8");
-    return { ...DEFAULTS };
-  }
-
-  return loadConfig(kbRoot);
-}