Files
joelhooks__joelclaw/scripts/backfill-run-typesense.ts
2026-07-29 19:10:22 -07:00

346 lines
13 KiB
TypeScript

#!/usr/bin/env bun
/**
* Emergency restore for the retired ADR-0243 Typesense Run projections.
*
* Source of truth is ~/.joelclaw/runs-dev/<user>/<yyyy-mm>/*.metadata.json +
* *.jsonl. Live session search uses sessions.db. This script requires an
* explicit retired-projection flag so routine recovery cannot recreate the
* collections.
*/
import { createHash } from "node:crypto";
import { existsSync, readdirSync, readFileSync, statSync } from "node:fs";
import { homedir } from "node:os";
import { join } from "node:path";
import { embed } from "../packages/inference-router/src/embeddings";
import {
type Chunk,
chunkTurns,
detectFormat,
embeddingModelTag,
extractTurns,
parseJsonl,
RUN_CHUNKS_COLLECTION,
RUNS_COLLECTION,
type Run,
runChunksSchema,
runsSchema,
} from "../packages/memory/src/index";
const TYPESENSE_URL = process.env.TYPESENSE_URL ?? "http://localhost:8108";
const TYPESENSE_API_KEY = process.env.TYPESENSE_API_KEY;
const EMBED_DIMS = 768;
interface Args {
user: string;
month: string;
since: number;
until: number;
machine?: string;
runtime?: string;
limit: number;
dryRun: boolean;
sleepMs: number;
allowRetiredProjection: boolean;
}
interface RunMetadata {
run_id?: string;
user_id?: string;
machine_id?: string;
agent_runtime?: Run["agent_runtime"];
parent_run_id?: string | null;
conversation_id?: string | null;
tags?: string[];
started_at?: number;
captured_at?: number;
jsonl_path?: string;
}
function usage(): never {
console.error(`Usage: TYPESENSE_API_KEY=... bun scripts/backfill-run-typesense.ts --allow-retired-projection [options]\n\nDANGER: this recreates retired runs_dev/run_chunks_dev collections. Live search uses sessions.db.\n\nOptions:\n --allow-retired-projection Explicitly permit emergency projection restore\n --user <id> User partition (default: joel)\n --month <yyyy-mm> Run store month (default: current UTC month)\n --since <iso|ms> Include runs at/after this start time\n --until <iso|ms> Include runs before this start time (default: now)\n --machine <id> Filter machine_id\n --runtime <runtime> Filter agent_runtime\n --limit <n> Max missing runs to index; 0 = all (default: 100)\n --sleep-ms <n> Delay after each run to avoid starving live embeds (default: 100)\n --dry-run Print planned work only\n`);
process.exit(2);
}
function parseTime(value: string | undefined, fallback: number): number {
if (!value) return fallback;
if (/^\d+$/.test(value)) return Number(value);
const parsed = Date.parse(value);
if (Number.isNaN(parsed)) throw new Error(`invalid timestamp: ${value}`);
return parsed;
}
function defaultMonth(): string {
const now = new Date();
return `${now.getUTCFullYear()}-${String(now.getUTCMonth() + 1).padStart(2, "0")}`;
}
function parseArgs(argv: string[]): Args {
const args: Args = {
user: "joel",
month: defaultMonth(),
since: 0,
until: Date.now() + 1,
limit: 100,
dryRun: false,
sleepMs: 100,
allowRetiredProjection: false,
};
for (let i = 0; i < argv.length; i += 1) {
const arg = argv[i]!;
const next = () => argv[++i] ?? usage();
if (arg === "--user") args.user = next();
else if (arg === "--month") args.month = next();
else if (arg === "--since") args.since = parseTime(next(), args.since);
else if (arg === "--until") args.until = parseTime(next(), args.until);
else if (arg === "--machine") args.machine = next();
else if (arg === "--runtime") args.runtime = next();
else if (arg === "--limit") args.limit = Number(next());
else if (arg === "--sleep-ms") args.sleepMs = Number(next());
else if (arg === "--dry-run") args.dryRun = true;
else if (arg === "--allow-retired-projection") args.allowRetiredProjection = true;
else if (arg === "--help" || arg === "-h") usage();
else usage();
}
if (!Number.isFinite(args.limit) || args.limit < 0) usage();
if (!Number.isFinite(args.sleepMs) || args.sleepMs < 0) usage();
return args;
}
function sleep(ms: number): Promise<void> {
return new Promise((resolve) => setTimeout(resolve, ms));
}
async function typesenseRequest(
path: string,
init: RequestInit = {}
): Promise<Response> {
if (!TYPESENSE_API_KEY) throw new Error("TYPESENSE_API_KEY not set");
return fetch(`${TYPESENSE_URL}${path}`, {
...init,
headers: {
...init.headers,
"X-TYPESENSE-API-KEY": TYPESENSE_API_KEY,
"Content-Type": "application/json",
},
});
}
async function ensureCollections(): Promise<void> {
for (const [name, schema] of [
[RUN_CHUNKS_COLLECTION, runChunksSchema(RUN_CHUNKS_COLLECTION, EMBED_DIMS)],
[RUNS_COLLECTION, runsSchema(RUNS_COLLECTION)],
] as const) {
const existing = await typesenseRequest(`/collections/${name}`);
if (existing.status === 200) continue;
if (existing.status !== 404) {
throw new Error(`typesense status ${existing.status} for ${name}: ${await existing.text()}`);
}
const res = await typesenseRequest("/collections", {
method: "POST",
body: JSON.stringify(schema),
});
if (!res.ok) throw new Error(`create ${name} failed: ${res.status} ${await res.text()}`);
}
}
async function loadExistingRunIds(): Promise<Set<string>> {
const res = await typesenseRequest(`/collections/${RUNS_COLLECTION}/documents/export?include_fields=id`);
if (!res.ok) throw new Error(`export ${RUNS_COLLECTION} failed: ${res.status} ${await res.text()}`);
const ids = new Set<string>();
for (const line of (await res.text()).split("\n")) {
if (!line.trim()) continue;
try {
const doc = JSON.parse(line) as { id?: string };
if (doc.id) ids.add(doc.id);
} catch {
// ignore malformed export lines
}
}
return ids;
}
function runStoreDir(user: string, month: string): string {
return join(process.env.MEMORY_RUN_STORE ?? join(homedir(), ".joelclaw", "runs-dev"), user, month);
}
function loadCandidates(args: Args, existingIds: Set<string>): Array<{ metadata: RunMetadata; metadataPath: string; jsonlPath: string }> {
const dir = runStoreDir(args.user, args.month);
if (!existsSync(dir)) throw new Error(`run store directory not found: ${dir}`);
const candidates: Array<{ metadata: RunMetadata; metadataPath: string; jsonlPath: string }> = [];
for (const name of readdirSync(dir)) {
if (!name.endsWith(".metadata.json")) continue;
const metadataPath = join(dir, name);
const metadata = JSON.parse(readFileSync(metadataPath, "utf8")) as RunMetadata;
const runId = metadata.run_id ?? name.replace(/\.metadata\.json$/, "");
if (existingIds.has(runId)) continue;
const startedAt = Number(metadata.started_at ?? 0);
if (startedAt < args.since || startedAt >= args.until) continue;
if (args.machine && metadata.machine_id !== args.machine) continue;
if (args.runtime && metadata.agent_runtime !== args.runtime) continue;
const jsonlPath = metadata.jsonl_path ?? join(dir, `${runId}.jsonl`);
if (!existsSync(jsonlPath)) continue;
candidates.push({ metadata: { ...metadata, run_id: runId }, metadataPath, jsonlPath });
}
candidates.sort((a, b) => Number(a.metadata.started_at ?? 0) - Number(b.metadata.started_at ?? 0));
return args.limit === 0 ? candidates : candidates.slice(0, args.limit);
}
async function importNdjson(collection: string, docs: unknown[]): Promise<void> {
if (docs.length === 0) return;
const res = await typesenseRequest(`/collections/${collection}/documents/import?action=upsert`, {
method: "POST",
headers: { "Content-Type": "text/plain" },
body: docs.map((doc) => JSON.stringify(doc)).join("\n"),
});
if (!res.ok) throw new Error(`${collection} import failed: ${res.status} ${await res.text()}`);
const body = await res.text();
const failures = body
.split("\n")
.filter((line) => line.trim())
.filter((line) => {
try {
return (JSON.parse(line) as { success?: boolean }).success === false;
} catch {
return true;
}
});
if (failures.length > 0) {
throw new Error(`${collection} import had ${failures.length}/${docs.length} failures: ${failures[0]!.slice(0, 300)}`);
}
}
async function indexRun(entry: { metadata: RunMetadata; jsonlPath: string }): Promise<{ chunks: number; bytes: number }> {
const metadata = entry.metadata;
const runId = metadata.run_id!;
const userId = metadata.user_id ?? "joel";
const machineId = metadata.machine_id ?? "unknown";
const agentRuntime = metadata.agent_runtime ?? "other";
const startedAt = Number(metadata.started_at ?? statSync(entry.jsonlPath).mtimeMs);
const jsonl = readFileSync(entry.jsonlPath, "utf8");
const jsonlBytes = Buffer.byteLength(jsonl, "utf8");
const jsonlSha256 = createHash("sha256").update(jsonl).digest("hex");
const entries = parseJsonl(jsonl);
const format = detectFormat(entries);
const turns = extractTurns(entries, format);
const candidates = chunkTurns(turns);
const modelTag = embeddingModelTag();
const chunks: Chunk[] = [];
for (const cand of candidates) {
const result = await embed(cand.text, {
priority: "ingest-bulk",
dimensions: EMBED_DIMS,
});
chunks.push({
id: `${runId}:${cand.chunk_idx}`,
run_id: runId,
chunk_idx: cand.chunk_idx,
role: cand.role,
text: cand.text,
embedding: result.embedding,
embedding_model: modelTag,
token_count: cand.token_count,
started_at: cand.started_at,
user_id: userId,
readable_by: [userId],
root_run_id: metadata.parent_run_id ?? null,
agent_runtime: agentRuntime,
conversation_id: metadata.conversation_id ?? null,
tags: metadata.tags ?? [],
machine_id: machineId,
});
}
await importNdjson(RUN_CHUNKS_COLLECTION, chunks);
const run: Partial<Run> & { id: string } = {
id: runId,
user_id: userId,
machine_id: machineId,
agent_runtime: agentRuntime,
agent_version: "",
model: "",
parent_run_id: metadata.parent_run_id ?? null,
root_run_id: metadata.parent_run_id ?? null,
conversation_id: metadata.conversation_id ?? null,
tags: metadata.tags ?? [],
readable_by: [userId],
intent: turns.find((turn) => turn.role === "user")?.text.slice(0, 500) ?? "",
started_at: startedAt,
ended_at: turns[turns.length - 1]?.started_at ?? startedAt,
duration_ms: (turns[turns.length - 1]?.started_at ?? startedAt) - startedAt,
turn_count: turns.length,
user_turn_count: turns.filter((turn) => turn.role === "user").length,
assistant_turn_count: turns.filter((turn) => turn.role === "assistant").length,
tool_turn_count: turns.filter((turn) => turn.role === "tool").length,
token_total: turns.reduce((sum, turn) => sum + turn.token_estimate, 0),
tool_call_count: turns.filter((turn) => turn.role === "tool").length,
files_touched: [],
skills_invoked: [],
entities_mentioned: [],
enriched_at: null,
enrichment_model: null,
status: "active",
full_text: turns.map((turn) => turn.text).join("\n"),
jsonl_path: entry.jsonlPath,
jsonl_bytes: jsonlBytes,
jsonl_sha256: jsonlSha256,
};
await importNdjson(RUNS_COLLECTION, [run]);
return { chunks: chunks.length, bytes: jsonlBytes };
}
async function main() {
const args = parseArgs(Bun.argv.slice(2));
if (!args.allowRetiredProjection) {
throw new Error(
"runs_dev/run_chunks_dev are retired; use scripts/backfill-session-index.ts or pass --allow-retired-projection for an emergency restore",
);
}
await ensureCollections();
const existingIds = await loadExistingRunIds();
const candidates = loadCandidates(args, existingIds);
console.log(JSON.stringify({ phase: "plan", candidates: candidates.length, args }));
if (args.dryRun) {
for (const c of candidates.slice(0, 20)) {
console.log(JSON.stringify({ run_id: c.metadata.run_id, started_at: new Date(Number(c.metadata.started_at ?? 0)).toISOString(), machine_id: c.metadata.machine_id, agent_runtime: c.metadata.agent_runtime, jsonl_path: c.jsonlPath }));
}
return;
}
let indexed = 0;
let chunks = 0;
let errors = 0;
const t0 = performance.now();
for (const candidate of candidates) {
const started = performance.now();
try {
const result = await indexRun(candidate);
indexed += 1;
chunks += result.chunks;
console.log(JSON.stringify({ phase: "indexed", run_id: candidate.metadata.run_id, machine_id: candidate.metadata.machine_id, chunks: result.chunks, bytes: result.bytes, duration_ms: Math.round(performance.now() - started) }));
} catch (err) {
errors += 1;
console.error(JSON.stringify({ phase: "error", run_id: candidate.metadata.run_id, error: String(err) }));
}
if (args.sleepMs > 0) await sleep(args.sleepMs);
}
console.log(JSON.stringify({ phase: "done", indexed, chunks, errors, duration_ms: Math.round(performance.now() - t0) }));
if (errors > 0) process.exitCode = 1;
}
main().catch((err) => {
console.error(err);
process.exit(1);
});