nlm-memory 0.5.0 → 0.5.2

package/migrations/010_sources_opencode.sql

@@ -1,30 +0,0 @@
--- Migration 010: add 'opencode' to the sources.kind CHECK constraint.
---
--- SQLite does not support ALTER COLUMN to modify CHECK constraints in place.
--- Standard approach: rename → recreate → copy → drop old.
-
-PRAGMA foreign_keys = OFF;
-
-CREATE TABLE sources_new (
-  id            INTEGER PRIMARY KEY AUTOINCREMENT,
-  kind          TEXT    NOT NULL CHECK (kind IN ('claude-code', 'hermes', 'opencode', 'pi', 'jsonl-generic', 'webhook')),
-  name          TEXT    NOT NULL UNIQUE,
-  path_or_url   TEXT,
-  runtime_label TEXT    NOT NULL,
-  parse_config  TEXT    NOT NULL DEFAULT '{}',
-  enabled       INTEGER NOT NULL DEFAULT 1,
-  token         TEXT,
-  created_at    TEXT    NOT NULL DEFAULT (datetime('now')),
-  updated_at    TEXT    NOT NULL DEFAULT (datetime('now'))
-);
-
-INSERT INTO sources_new SELECT id, kind, name, path_or_url, runtime_label, parse_config, enabled, token, created_at, updated_at FROM sources;
-
-DROP TABLE sources;
-ALTER TABLE sources_new RENAME TO sources;
-
-CREATE INDEX IF NOT EXISTS idx_sources_enabled ON sources(enabled) WHERE enabled = 1;
-
-PRAGMA foreign_keys = ON;
-
-INSERT OR IGNORE INTO schema_migrations (version, name) VALUES (10, '010_sources_opencode');

package/migrations/011_sources_hermes_agent.sql

@@ -1,30 +0,0 @@
--- Migration 011: add 'hermes-agent' to the sources.kind CHECK constraint.
---
--- SQLite does not support ALTER COLUMN to modify CHECK constraints in place.
--- Standard approach: rename → recreate → copy → drop old.
-
-PRAGMA foreign_keys = OFF;
-
-CREATE TABLE sources_new (
-  id            INTEGER PRIMARY KEY AUTOINCREMENT,
-  kind          TEXT    NOT NULL CHECK (kind IN ('claude-code', 'hermes', 'hermes-agent', 'opencode', 'pi', 'jsonl-generic', 'webhook')),
-  name          TEXT    NOT NULL UNIQUE,
-  path_or_url   TEXT,
-  runtime_label TEXT    NOT NULL,
-  parse_config  TEXT    NOT NULL DEFAULT '{}',
-  enabled       INTEGER NOT NULL DEFAULT 1,
-  token         TEXT,
-  created_at    TEXT    NOT NULL DEFAULT (datetime('now')),
-  updated_at    TEXT    NOT NULL DEFAULT (datetime('now'))
-);
-
-INSERT INTO sources_new SELECT id, kind, name, path_or_url, runtime_label, parse_config, enabled, token, created_at, updated_at FROM sources;
-
-DROP TABLE sources;
-ALTER TABLE sources_new RENAME TO sources;
-
-CREATE INDEX IF NOT EXISTS idx_sources_enabled ON sources(enabled) WHERE enabled = 1;
-
-PRAGMA foreign_keys = ON;
-
-INSERT OR IGNORE INTO schema_migrations (version, name) VALUES (11, '011_sources_hermes_agent');

package/migrations/012_sources_aider.sql

@@ -1,30 +0,0 @@
--- Migration 012: add 'aider' to the sources.kind CHECK constraint.
---
--- SQLite does not support ALTER COLUMN to modify CHECK constraints in place.
--- Standard approach: rename → recreate → copy → drop old.
-
-PRAGMA foreign_keys = OFF;
-
-CREATE TABLE sources_new (
-  id            INTEGER PRIMARY KEY AUTOINCREMENT,
-  kind          TEXT    NOT NULL CHECK (kind IN ('claude-code', 'hermes', 'hermes-agent', 'aider', 'opencode', 'pi', 'jsonl-generic', 'webhook')),
-  name          TEXT    NOT NULL UNIQUE,
-  path_or_url   TEXT,
-  runtime_label TEXT    NOT NULL,
-  parse_config  TEXT    NOT NULL DEFAULT '{}',
-  enabled       INTEGER NOT NULL DEFAULT 1,
-  token         TEXT,
-  created_at    TEXT    NOT NULL DEFAULT (datetime('now')),
-  updated_at    TEXT    NOT NULL DEFAULT (datetime('now'))
-);
-
-INSERT INTO sources_new SELECT id, kind, name, path_or_url, runtime_label, parse_config, enabled, token, created_at, updated_at FROM sources;
-
-DROP TABLE sources;
-ALTER TABLE sources_new RENAME TO sources;
-
-CREATE INDEX IF NOT EXISTS idx_sources_enabled ON sources(enabled) WHERE enabled = 1;
-
-PRAGMA foreign_keys = ON;
-
-INSERT OR IGNORE INTO schema_migrations (version, name) VALUES (12, '012_sources_aider');

package/migrations/013_adapter_state_failure_count.sql

@@ -1,12 +0,0 @@
--- Migration 013: add failure_count to adapter_state.
---
--- Tracks consecutive classify/storage failures per source file. The scheduler
--- skips files whose failure_count has reached the backoff ceiling (default 3)
--- and whose file_size has not changed since the last attempt. When the file
--- grows (new content appended), failure_count resets to 0 and the file is
--- retried. This prevents a single permanently-broken or over-large transcript
--- from flooding daemon-err.log on every 30-minute tick.
-
-ALTER TABLE adapter_state ADD COLUMN failure_count INTEGER NOT NULL DEFAULT 0;
-
-INSERT OR IGNORE INTO schema_migrations (version, name) VALUES (13, '013_adapter_state_failure_count');

package/migrations/014_sources_cursor.sql

@@ -1,30 +0,0 @@
--- Migration 014: add 'cursor' to the sources.kind CHECK constraint.
---
--- SQLite does not support ALTER COLUMN to modify CHECK constraints in place.
--- Standard approach: rename → recreate → copy → drop old.
-
-PRAGMA foreign_keys = OFF;
-
-CREATE TABLE sources_new (
-  id            INTEGER PRIMARY KEY AUTOINCREMENT,
-  kind          TEXT    NOT NULL CHECK (kind IN ('claude-code', 'hermes', 'hermes-agent', 'aider', 'cursor', 'opencode', 'pi', 'jsonl-generic', 'webhook')),
-  name          TEXT    NOT NULL UNIQUE,
-  path_or_url   TEXT,
-  runtime_label TEXT    NOT NULL,
-  parse_config  TEXT    NOT NULL DEFAULT '{}',
-  enabled       INTEGER NOT NULL DEFAULT 1,
-  token         TEXT,
-  created_at    TEXT    NOT NULL DEFAULT (datetime('now')),
-  updated_at    TEXT    NOT NULL DEFAULT (datetime('now'))
-);
-
-INSERT INTO sources_new SELECT id, kind, name, path_or_url, runtime_label, parse_config, enabled, token, created_at, updated_at FROM sources;
-
-DROP TABLE sources;
-ALTER TABLE sources_new RENAME TO sources;
-
-CREATE INDEX IF NOT EXISTS idx_sources_enabled ON sources(enabled) WHERE enabled = 1;
-
-PRAGMA foreign_keys = ON;
-
-INSERT OR IGNORE INTO schema_migrations (version, name) VALUES (14, '014_sources_cursor');

package/migrations/015_sources_windsurf.sql

@@ -1,30 +0,0 @@
--- Migration 015: add 'windsurf' to the sources.kind CHECK constraint.
---
--- SQLite does not support ALTER COLUMN to modify CHECK constraints in place.
--- Standard approach: rename → recreate → copy → drop old.
-
-PRAGMA foreign_keys = OFF;
-
-CREATE TABLE sources_new (
-  id            INTEGER PRIMARY KEY AUTOINCREMENT,
-  kind          TEXT    NOT NULL CHECK (kind IN ('claude-code', 'hermes', 'hermes-agent', 'aider', 'cursor', 'windsurf', 'opencode', 'pi', 'jsonl-generic', 'webhook')),
-  name          TEXT    NOT NULL UNIQUE,
-  path_or_url   TEXT,
-  runtime_label TEXT    NOT NULL,
-  parse_config  TEXT    NOT NULL DEFAULT '{}',
-  enabled       INTEGER NOT NULL DEFAULT 1,
-  token         TEXT,
-  created_at    TEXT    NOT NULL DEFAULT (datetime('now')),
-  updated_at    TEXT    NOT NULL DEFAULT (datetime('now'))
-);
-
-INSERT INTO sources_new SELECT id, kind, name, path_or_url, runtime_label, parse_config, enabled, token, created_at, updated_at FROM sources;
-
-DROP TABLE sources;
-ALTER TABLE sources_new RENAME TO sources;
-
-CREATE INDEX IF NOT EXISTS idx_sources_enabled ON sources(enabled) WHERE enabled = 1;
-
-PRAGMA foreign_keys = ON;
-
-INSERT OR IGNORE INTO schema_migrations (version, name) VALUES (15, '015_sources_windsurf');

package/plugin-hermes-agent/README.md

@@ -1,49 +0,0 @@
-# nlm-memory Hermes Agent plugin
-
-This directory is the [NousResearch Hermes Agent](https://github.com/NousResearch/hermes-agent) plugin distribution surface for nlm-memory.
-
-## Install
-
-Prerequisite: `npm install -g nlm-memory` (puts `nlm` on PATH; the MCP server spawns `nlm mcp`).
-
-```bash
-# Recommended — uses nlm connect to copy the plugin and enable it
-nlm connect hermes-agent
-
-# Or manually
-cp -r plugin-hermes-agent ~/.hermes/plugins/nlm-memory
-hermes plugins enable nlm-memory
-nlm connect hermes          # also writes the MCP server entry to ~/.hermes/config.yaml
-```
-
-## What this ships
-
-- **`plugin.yaml`** — Hermes plugin manifest (name: `nlm-memory`, kind: `memory`)
-- **`__init__.py`** — Python shim that registers 6 lifecycle hooks; all delegate to the local nlm daemon at `http://localhost:3940`
-
-## How it works
-
-| Hermes hook | Action |
-|---|---|
-| `pre_llm_call` | Keyword-recalls up to 3 relevant prior sessions; injects them as context above the user message |
-| `post_llm_call` | Scans the assistant response for cited session IDs and logs prose citation events |
-| `on_session_start` | No-op (memo is created lazily on first recall hit) |
-| `on_session_end` | Clears the per-session surfaced-ID memo |
-| `on_session_finalize` | Clears the per-session surfaced-ID memo |
-| `on_session_reset` | Clears the per-session surfaced-ID memo |
-
-## Configuration
-
-| Env var | Default | Purpose |
-|---|---|---|
-| `NLM_DAEMON_PORT` | `3940` | Port the nlm daemon listens on |
-
-## Daemon required
-
-The plugin delegates entirely to the nlm daemon. Start it with `nlm start` or `nlm install` (macOS LaunchAgent). If the daemon is unreachable, every hook silently returns `None` — the agent loop is never blocked.
-
-## Uninstall
-
-```bash
-nlm disconnect hermes-agent
-```

package/plugin-hermes-agent/__init__.py

@@ -1,75 +0,0 @@
-"""
-nlm-memory Hermes Agent plugin.
-
-Wires 6 Hermes lifecycle hooks to the local nlm-memory daemon
-(http://localhost:3940 by default). All calls are fire-and-forget
-except pre_llm_call, which returns a context string for injection.
-
-The daemon must be running for hooks to have effect. If it is unreachable,
-every hook silently returns None so the agent loop is never blocked.
-"""
-
-import json
-import os
-import urllib.error
-import urllib.request
-from typing import Any
-
-
-def _base_url() -> str:
-    port = os.environ.get("NLM_DAEMON_PORT", "3940")
-    return f"http://localhost:{port}"
-
-
-def _post(path: str, payload: dict[str, Any]) -> dict[str, Any]:
-    url = f"{_base_url()}/api/{path}"
-    data = json.dumps(payload).encode()
-    req = urllib.request.Request(
-        url,
-        data=data,
-        headers={"Content-Type": "application/json"},
-        method="POST",
-    )
-    try:
-        with urllib.request.urlopen(req, timeout=5) as resp:
-            return json.loads(resp.read())
-    except (urllib.error.URLError, OSError, json.JSONDecodeError):
-        return {}
-
-
-def register(ctx: Any) -> None:
-    def pre_llm_call(
-        session_id: str,
-        user_message: str,
-        is_first_turn: bool = False,
-        **kwargs: Any,
-    ) -> str | None:
-        result = _post(
-            "hook/hermes-agent/pre-turn",
-            {"session_id": session_id, "user_message": user_message, "is_first_turn": is_first_turn},
-        )
-        context = result.get("context")
-        return context if isinstance(context, str) and context else None
-
-    def post_llm_call(
-        session_id: str,
-        assistant_response: str,
-        **kwargs: Any,
-    ) -> None:
-        _post(
-            "hook/hermes-agent/post-turn",
-            {"session_id": session_id, "assistant_response": assistant_response},
-        )
-
-    def _session_event(event: str, session_id: str | None, **kwargs: Any) -> None:
-        _post(
-            "hook/hermes-agent/session-lifecycle",
-            {"event": event, "session_id": session_id},
-        )
-
-    ctx.register_hook("pre_llm_call", pre_llm_call)
-    ctx.register_hook("post_llm_call", post_llm_call)
-    ctx.register_hook("on_session_start", lambda session_id, **kw: _session_event("start", session_id, **kw))
-    ctx.register_hook("on_session_end", lambda session_id, **kw: _session_event("end", session_id, **kw))
-    ctx.register_hook("on_session_finalize", lambda session_id=None, **kw: _session_event("finalize", session_id, **kw))
-    ctx.register_hook("on_session_reset", lambda session_id, **kw: _session_event("reset", session_id, **kw))

package/plugin-hermes-agent/plugin.yaml

@@ -1,15 +0,0 @@
-name: nlm-memory
-version: 0.1.0
-description: Non-linear memory for Hermes — surfaces relevant prior sessions at each turn via keyword recall
-kind: memory
-author: pbmagnet4
-provides_hooks:
-  - pre_llm_call
-  - post_llm_call
-  - on_session_start
-  - on_session_end
-  - on_session_finalize
-  - on_session_reset
-requires_env:
-  - name: NLM_DAEMON_PORT
-    description: "Port the nlm-memory daemon listens on (default: 3940)"

package/scripts/build-codex-plugin.mjs

@@ -1,61 +0,0 @@
-#!/usr/bin/env node
-/**
- * Bundles each hook entry under src/hook/ as a single-file .mjs into
- * plugin/scripts/, ready for Codex to invoke via ${CLAUDE_PLUGIN_ROOT}.
- *
- * Why single-file bundles: Codex plugins are content-addressed by file hash
- * for the trust system. Shipping a tree of imports would force the user to
- * re-trust every release across dozens of files. One file per hook keeps
- * the trust footprint minimal.
- *
- * Why esbuild not tsc: tsc emits a tree of files with relative imports; the
- * plugin would have to ship `dist/` plus its dependency tree. esbuild flattens
- * to one .mjs per entry with all internal deps inlined. node builtins are
- * left as-is (node:*) and zero runtime deps are bundled because the hook
- * scripts only talk to the daemon over HTTP — no native bindings required.
- */
-
-import { build } from "esbuild";
-import { resolve, dirname } from "node:path";
-import { fileURLToPath } from "node:url";
-import { mkdirSync } from "node:fs";
-
-const __dirname = dirname(fileURLToPath(import.meta.url));
-const REPO_ROOT = resolve(__dirname, "..");
-const OUT_DIR = resolve(REPO_ROOT, "plugin/scripts");
-
-const ENTRIES = [
-  "src/hook/prompt-recall-hook.ts",
-  "src/hook/stop-hook.ts",
-];
-
-mkdirSync(OUT_DIR, { recursive: true });
-
-const results = await Promise.all(
-  ENTRIES.map(async (entry) => {
-    const entryPath = resolve(REPO_ROOT, entry);
-    const outName = entry.split("/").pop().replace(/\.ts$/, ".mjs");
-    const outFile = resolve(OUT_DIR, outName);
-    const result = await build({
-      entryPoints: [entryPath],
-      outfile: outFile,
-      bundle: true,
-      platform: "node",
-      format: "esm",
-      target: "node20",
-      minify: false,
-      sourcemap: false,
-      legalComments: "none",
-      logLevel: "warning",
-      banner: { js: "#!/usr/bin/env node" },
-      tsconfig: resolve(REPO_ROOT, "tsconfig.json"),
-    });
-    return { entry, outFile, warnings: result.warnings.length };
-  }),
-);
-
-for (const r of results) {
-  process.stdout.write(`bundled ${r.entry} -> ${r.outFile.replace(REPO_ROOT + "/", "")}`);
-  if (r.warnings > 0) process.stdout.write(` (${r.warnings} warnings)`);
-  process.stdout.write("\n");
-}

package/scripts/deepseek-probe.mjs

@@ -1,67 +0,0 @@
-#!/usr/bin/env node
-// Direct DeepSeek probe — calls the API with the same prompt + a known-failing
-// session body and dumps everything: raw choice content, finish_reason, usage.
-
-import { readFileSync } from "node:fs";
-import Database from "better-sqlite3";
-
-const envContent = readFileSync(`${process.env.HOME}/.nlm/.env`, "utf8");
-const apiKey = envContent.split("\n").find((l) => l.startsWith("DEEPSEEK_API_KEY="))?.split("=")[1]?.trim();
-if (!apiKey) {
-  console.error("no DEEPSEEK_API_KEY in ~/.nlm/.env");
-  process.exit(1);
-}
-
-const db = new Database(`${process.env.HOME}/.nlm/canonical.sqlite`, { readonly: true });
-const targets = process.argv.slice(2);
-if (targets.length === 0) {
-  console.error("usage: deepseek-probe.mjs <session-id> [more ids...]");
-  process.exit(1);
-}
-
-// Pull the actual classifier prompt from the build
-const promptModule = await import("/Users/echalupa/Documents/Coding Projects/nlm-memory-ts/src/core/classifier/prompt.ts");
-
-for (const sid of targets) {
-  const row = db.prepare("SELECT body FROM sessions WHERE id = ?").get(sid);
-  if (!row?.body) {
-    console.log(`\n=== ${sid}: no body ===`);
-    continue;
-  }
-  const userPrompt = promptModule.buildUserPrompt(row.body, "");
-  console.log(`\n=== ${sid} (body ${row.body.length} chars, prompt ${userPrompt.length} chars) ===`);
-
-  const res = await fetch("https://api.deepseek.com/v1/chat/completions", {
-    method: "POST",
-    headers: { "Content-Type": "application/json", Authorization: `Bearer ${apiKey}` },
-    body: JSON.stringify({
-      model: "deepseek-v4-flash",
-      messages: [
-        { role: "system", content: promptModule.CLASSIFIER_SYSTEM_PROMPT },
-        { role: "user", content: userPrompt },
-      ],
-      response_format: { type: "json_object" },
-      temperature: 0.1,
-      max_tokens: 8192,
-      stream: false,
-    }),
-  });
-  if (!res.ok) {
-    console.log(`HTTP ${res.status}: ${(await res.text()).slice(0, 500)}`);
-    continue;
-  }
-  const data = await res.json();
-  const choice = data.choices?.[0];
-  console.log(`finish_reason: ${choice?.finish_reason}`);
-  console.log(`usage: ${JSON.stringify(data.usage)}`);
-  const content = choice?.message?.content ?? "";
-  console.log(`content (${content.length} chars):`);
-  console.log(content);
-  console.log("--- end content ---");
-  try {
-    JSON.parse(content.replace(/^```(?:json)?\s*|\s*```$/gm, "").trim());
-    console.log("JSON.parse: OK");
-  } catch (e) {
-    console.log(`JSON.parse: FAILED — ${e.message}`);
-  }
-}

package/scripts/extract-triples.mjs

@@ -1,207 +0,0 @@
-#!/usr/bin/env node
-/**
- * Extract (query, surfaced_id, surfaced_body, label, weight, source) triples
- * from the NLM telemetry logs for reranker training data.
- *
- * Data sources:
- *   ~/.nlm/hook-log.jsonl  — one row per UserPromptSubmit/SessionStart fire,
- *                            fields: ts, conversationId, promptPreview, wouldInject[]
- *   ~/.nlm/citation-log.jsonl — one row per detected citation,
- *                            fields: ts, conversation_id, cited_id, kind
- *   ~/.nlm/canonical.sqlite — sessions table for surfaced_body lookup
- *
- * Algorithm:
- *   1. Index citations by (conversationId, citedId, kind).
- *   2. Identify "gold conversations": conversations that produced at least one
- *      tool_use citation. These are the only conversations with confirmed
- *      positive signal — surfaced-but-not-cited sessions in them are genuine
- *      hard negatives.
- *   3. For each hook-log entry with wouldInject.length > 0:
- *      - If the conversation is a gold conversation:
- *        * tool_use cited sessions → weight 1.0, source "tool_use" (gold positive)
- *        * NOT cited sessions → weight 0.0, source "hard_negative"
- *      - Prose-only conversations are skipped (signal too noisy).
- *   4. Fetch surfaced_body from SQLite for each (query, surfaced_id) pair.
- *   5. Write JSONL to --output path (or stdout).
- *
- * Dedup: one triple per (query, surfaced_id, source) key — duplicate hook
- * fires for the same (conversationId, id) pair are collapsed.
- *
- * Usage:
- *   node scripts/extract-triples.mjs                    # stdout, last 30d
- *   node scripts/extract-triples.mjs --days 7           # last 7 days
- *   node scripts/extract-triples.mjs --output triples.jsonl
- *   node scripts/extract-triples.mjs --stats            # summary only, no output
- */
-
-import { createReadStream, existsSync, writeFileSync } from "node:fs";
-import { createInterface } from "node:readline";
-import { homedir } from "node:os";
-import { join } from "node:path";
-import { createWriteStream } from "node:fs";
-import Database from "better-sqlite3";
-
-// ── CLI args ─────────────────────────────────────────────────────────────────
-
-const args = process.argv.slice(2);
-const days = (() => {
-  const idx = args.indexOf("--days");
-  return idx !== -1 ? Number.parseInt(args[idx + 1], 10) || 30 : 30;
-})();
-const outputPath = (() => {
-  const idx = args.indexOf("--output");
-  return idx !== -1 ? args[idx + 1] : null;
-})();
-const statsOnly = args.includes("--stats");
-
-const hookLogPath = process.env["NLM_HOOK_LOG"] ?? join(homedir(), ".nlm", "hook-log.jsonl");
-const citationLogPath = process.env["NLM_CITATION_LOG"] ?? join(homedir(), ".nlm", "citation-log.jsonl");
-const dbPath = process.env["NLM_DB_PATH"] ?? join(homedir(), ".nlm", "canonical.sqlite");
-
-const cutoff = Date.now() - days * 24 * 60 * 60 * 1000;
-
-// ── Helpers ───────────────────────────────────────────────────────────────────
-
-async function readJsonl(path) {
-  if (!existsSync(path)) return [];
-  const rows = [];
-  const rl = createInterface({ input: createReadStream(path), crlfDelay: Infinity });
-  for await (const line of rl) {
-    const trimmed = line.trim();
-    if (!trimmed) continue;
-    try {
-      rows.push(JSON.parse(trimmed));
-    } catch {
-      // skip malformed lines
-    }
-  }
-  return rows;
-}
-
-function tsMs(entry, field = "ts") {
-  const v = entry[field];
-  if (typeof v !== "string") return 0;
-  return Date.parse(v);
-}
-
-// ── Step 1: Load citations ────────────────────────────────────────────────────
-
-const citationRows = await readJsonl(citationLogPath);
-
-// Map: conversationId → Set of tool_use cited IDs
-const toolUseCitations = new Map();
-// Set of gold conversations (those with ≥1 tool_use citation)
-const goldConversations = new Set();
-
-for (const row of citationRows) {
-  if (row.kind !== "tool_use") continue;
-  const convId = row.conversation_id;
-  const citedId = row.cited_id;
-  if (!convId || !citedId) continue;
-  if (!toolUseCitations.has(convId)) toolUseCitations.set(convId, new Set());
-  toolUseCitations.get(convId).add(citedId);
-  goldConversations.add(convId);
-}
-
-// ── Step 2: Load hook-log entries in window ───────────────────────────────────
-
-const hookRows = await readJsonl(hookLogPath);
-
-// Collect: for each (query, surfaced_id), record (conversationId, weight, source)
-// Dedup key: `${conversationId}::${query}::${surfaced_id}::${source}`
-const seen = new Set();
-const rawTriples = [];
-
-for (const row of hookRows) {
-  // Skip stop-hook entries
-  if (typeof row.kind === "string") continue;
-  const ts = tsMs(row);
-  if (!ts || ts < cutoff) continue;
-  const conversationId = row.conversationId;
-  const query = row.promptPreview;
-  const wouldInject = Array.isArray(row.wouldInject) ? row.wouldInject : [];
-  if (!conversationId || !query || wouldInject.length === 0) continue;
-  // Only process gold conversations
-  if (!goldConversations.has(conversationId)) continue;
-
-  const cited = toolUseCitations.get(conversationId) ?? new Set();
-
-  for (const surfacedId of wouldInject) {
-    if (typeof surfacedId !== "string") continue;
-    const isCited = cited.has(surfacedId);
-    const source = isCited ? "tool_use" : "hard_negative";
-    const weight = isCited ? 1.0 : 0.0;
-    const key = `${conversationId}::${query}::${surfacedId}::${source}`;
-    if (seen.has(key)) continue;
-    seen.add(key);
-    rawTriples.push({ query, surfaced_id: surfacedId, weight, source });
-  }
-}
-
-if (rawTriples.length === 0) {
-  console.error(
-    `extract-triples: no triples found. ` +
-    `Checked ${goldConversations.size} gold conversations in the last ${days}d. ` +
-    `Run nlm useful-scan and ensure tool_use citations exist in citation-log.jsonl.`,
-  );
-  process.exit(0);
-}
-
-// ── Step 3: Fetch surfaced_body from SQLite ────────────────────────────────────
-
-const uniqueIds = [...new Set(rawTriples.map((t) => t.surfaced_id))];
-
-let bodyById = new Map();
-if (existsSync(dbPath)) {
-  try {
-    const db = new Database(dbPath, { readonly: true });
-    const placeholders = uniqueIds.map(() => "?").join(",");
-    const rows = db.prepare(
-      `SELECT id, body FROM sessions WHERE id IN (${placeholders})`,
-    ).all(...uniqueIds);
-    for (const row of rows) {
-      bodyById.set(row.id, typeof row.body === "string" ? row.body : "");
-    }
-    db.close();
-  } catch (err) {
-    console.error(`extract-triples: sqlite lookup failed — ${err.message}. Bodies will be empty.`);
-  }
-} else {
-  console.error(`extract-triples: db not found at ${dbPath}. Bodies will be empty.`);
-}
-
-// ── Step 4: Assemble final triples ────────────────────────────────────────────
-
-const triples = rawTriples.map((t) => ({
-  query: t.query,
-  surfaced_id: t.surfaced_id,
-  surfaced_body: bodyById.get(t.surfaced_id) ?? "",
-  label: t.weight === 1.0 ? "positive" : "negative",
-  weight: t.weight,
-  source: t.source,
-}));
-
-// ── Stats ─────────────────────────────────────────────────────────────────────
-
-const positives = triples.filter((t) => t.weight === 1.0).length;
-const negatives = triples.filter((t) => t.weight === 0.0).length;
-const withBody = triples.filter((t) => t.surfaced_body.length > 0).length;
-
-console.error(
-  `extract-triples: ${triples.length} triples ` +
-  `(${positives} positive, ${negatives} hard-negative, ${withBody}/${triples.length} with body) ` +
-  `from ${goldConversations.size} gold conversations over last ${days}d`,
-);
-
-if (statsOnly) process.exit(0);
-
-// ── Step 5: Write output ──────────────────────────────────────────────────────
-
-const lines = triples.map((t) => JSON.stringify(t)).join("\n") + "\n";
-
-if (outputPath) {
-  writeFileSync(outputPath, lines, "utf8");
-  console.error(`extract-triples: wrote ${triples.length} rows to ${outputPath}`);
-} else {
-  process.stdout.write(lines);
-}