Longrun

longrun

Write scripts that are concurrent inside, streaming out, and async from conversation (tmux).

Trigger: About to write a script that loops over items, calls APIs, processes files, or will take >10s.

The Pattern

1. Script Structure (Deno + dax + std)

#!/usr/bin/env -S deno run --allow-all
/** One-line description of what this does. */
import $ from "jsr:@david/dax@0.44.2";
import { pooledMap } from "jsr:@std/async";

const CONCURRENCY = 10;
const OUTPUT = "/tmp/TASKNAME.jsonl";

type Result = { item: string; status: "ok" | "error"; data?: unknown; error?: string };

async function process(item: string): Promise<Result> {
  try {
    const resp = await fetch(`https://api.example.com/${item}`);
    const data = await resp.json();
    return { item, status: "ok", data };
  } catch (e) {
    return { item, status: "error", error: String(e) };
  }
}

// Resume: skip already-processed items
const done = new Set<string>();
try {
  for (const line of (await Deno.readTextFile(OUTPUT)).split("\n").filter(Boolean)) {
    done.add(JSON.parse(line).item);
  }
} catch { /* file doesn't exist yet */ }

const items = ALL_ITEMS.filter((i) => !done.has(i));
if (!items.length) { console.error("All done."); Deno.exit(0); }
console.error(`Processing ${items.length} items (${done.size} already done)`);

// Concurrent pool → stream to disk
const file = await Deno.open(OUTPUT, { append: true, create: true });
const enc = new TextEncoder();

for await (const result of pooledMap(CONCURRENCY, items, process)) {
  const line = JSON.stringify(result);
  file.writeSync(enc.encode(line + "\n"));
  console.log(line);
}

file.close();

2. Run in tmux

tmux has-session -t pi 2>/dev/null || tmux new-session -d -s pi
tmux new-window -d -t pi -n TASKNAME 'deno run --allow-all /tmp/TASKNAME.ts'

3. Monitor

# Progress count
wc -l /tmp/TASKNAME.jsonl

# Last few results
tail -3 /tmp/TASKNAME.jsonl | jq .

# Error count
grep -c '"status":"error"' /tmp/TASKNAME.jsonl

# Sanity check first result
head -1 /tmp/TASKNAME.jsonl | jq .

# Live follow
tmux capture-pane -t pi:TASKNAME -p -S -10

4. Cleanup

tmux kill-window -t pi:TASKNAME 2>/dev/null

Key Rules

• Always .jsonl — one JSON object per line. Parseable, appendable, streamable.
• Always writeSync — results hit disk immediately. No buffering.
• Always append mode — enables resume after crash.
• Always try/catch per item — one failure must not kill the batch.
• pooledMap for concurrency — don't blast APIs. Start with 10, adjust.
• Status field in every result — "ok" or "error", makes filtering trivial.
• Progress to stderr — counts, summaries, warnings go to stderr. Data goes to stdout/file.
• Output to /tmp/ — disposable. Move to .git/reports/ if worth keeping.
• File extension .ts — all scripts are TypeScript, run with deno run --allow-all.

Variations

Simple (no concurrency needed)

#!/usr/bin/env -S deno run --allow-all
for (const item of items) {
  const result = await process(item);
  console.log(JSON.stringify(result));
}

Shell commands with dax

#!/usr/bin/env -S deno run --allow-all
import $ from "jsr:@david/dax@0.44.2";
import { pooledMap } from "jsr:@std/async";

for await (const result of pooledMap(5, files, async (file) => {
  const output = await $`ffprobe -v quiet -print_format json -show_format ${file}`.json();
  return { file, status: "ok", data: output };
})) {
  console.log(JSON.stringify(result));
}

Fetch with retry

#!/usr/bin/env -S deno run --allow-all
import { pooledMap, retry } from "jsr:@std/async";

for await (const result of pooledMap(5, urls, async (url) => {
  const data = await retry(async () => {
    const r = await fetch(url);
    if (!r.ok) throw new Error(`${r.status}`);
    return r.json();
  }, { maxAttempts: 3, minTimeout: 1000 });
  return { url, status: "ok", data };
})) {
  console.log(JSON.stringify(result));
}

LLM vs Algorithm — Decision Point

Before writing fuzzy matching, scoring functions, or multi-iteration string comparison — STOP.

These tasks are judgment, not computation. Use ask():

• Entity matching: "Is hotel A the same as hotel B?" — $0.001, 100% accuracy
• Classification: "Is this review positive?" — cheaper than sentiment libraries
• Extraction: "Extract the room type from this string" — no regex needed
• Validation: "Does this result make sense?" — catches edge cases you'd never code for

Tokens are unlimited (Max subscription). Don't optimize for cost — optimize for result quality. Writing 10 iterations of fuzzy matching = hours wasted + bugs + false matches. ask() gets it right first time.

Rule of thumb: if you're on iteration 2+ of fixing edge cases in string matching, you already wasted time — should have used ask() from the start.

When NOT to use LLM

• Exact lookups, numeric comparisons, deterministic transforms — code is faster
• Real-time hot paths — ask() adds 200-500ms per call

pi-llm for Smart Batch Jobs

import { ask, run } from "~/.pi/agent/lib/pi-llm.ts";

// Judgment per item — cheap and accurate
for (const item of remaining) {
  const match = await ask(
    `Is "${item.nameA}" the same entity as "${item.nameB}"? ` +
    `Context: ${item.city}, ${item.country}. Answer: yes/no`,
    { model: "claude-haiku-4-5" }  // $0.001/call
  );
  // ...
}

// Research per item — when agent needs tools
for (const item of remaining) {
  const result = await run(`Research ${item.name}`, { tools: "full", maxTurns: 8 });
  file.writeSync(enc.encode(JSON.stringify({ id: item.id, text: result.text, cost: result.cost }) + "\n"));
}

ask() = text→text, no tools, instant. run() = bounded agent with bash/search/read. Tokens are unlimited — use freely.

Anti-patterns (DON'T)

• ❌ Accumulate results in a list, dump at the end
• ❌ Run inline and block the conversation
• ❌ No error handling per item
• ❌ No flush — buffered output defeats the purpose
• ❌ CSV/custom formats — jsonl is universal
• ❌ Overwrite mode — kills resume capability
• ❌ Manual semaphore spin-wait — use pooledMap
• ❌ Writing fuzzy matching / scoring algorithms for entity resolution — use ask() for judgment
• ❌ Iterating on string heuristics (identity keywords, compound tokenizers, etc.) — if iteration 2+ of edge case fixes, switch to LLM