The Orchestra Learned To Celebrate

📺 Same Data, Different Show

The orchestra runner already exposes everything you'd want to see while a run is happening. telemetry.jsonl ticks every spawn, every result, every score. state.json tells you which phase is running. scoreboard.json is fresh every two seconds.

The original tree viewer rendered all of that as a green-orange-red dot graph. Functional. Sterile. You could read it, but you wouldn't want to watch it.

🎯 The Hybrid Scoring Model

The trick was to keep quality as the floor while letting drama sit on top of it. Inventing points for things that don't matter would have rewarded the wrong behaviours. So scoring is a sandwich.

The base score comes from the existing judge dimensions — correctness, robustness, completeness. Achievements fire from telemetry events. Combos multiply.

Quality drives the score. There are no points for cutting corners. Speed only counts when the output is correct.

🎉 Achievements That Fire Live

Each achievement is a single object in achievements.js — an ID, a condition function, a points value, an FX cue. Tunable in one file.

The scoring engine watches every result-*.json the runner emits, computes the score, applies any combo multiplier, and writes a score_event line into telemetry.jsonl. The viewer was already polling that file every two seconds.

🔌 Two Layers, One File Each

The whole thing is two new modules and one modified one. Below is the actual scoring entry point that fires on every worker end.

// runner/src/lib/scoreboard.js
export function recordScoreFor(rd, name, resultPath, started, timeoutSec) {
  const result = readJSON(resultPath);
  const score  = computeWorkerScore(name, result, started, Date.now(), timeoutSec);
  const combo  = applyCombos(loadScoreboard(rd), score);

  appendTelemetry(rd, {
    event: 'score_event',
    worker: name,
    points: score.total,
    breakdown: score.bonuses,
    achievements: score.achievements,
    combo: combo.current,
    multiplier: combo.multiplier,
    comboBanner: combo.banner,
    ts: Date.now() / 1000,
  });

  writeScoreboardAtomic(rd, applyToBoard(loadScoreboard(rd), score, combo));
}

The viewer side is just a switch on event === 'score_event'. Spawn a popup at the worker's SVG coords, increment the scoreboard, fire the banner if a combo flagged it.

📝 Telling The Agents The Rules

The most important detail isn't the visuals at all. Every worker brief gets a === SCORING === block stitched in at composition time, listing the points table in plain language.

=== SCORING ===
Base: judge dim score × 100.
Bonuses you can earn:
  • SPEED_DEMON  +20  (finish in <50% of phase timeout, with correct output)
  • FLAWLESS     +30  (judge dim score 1.0)
  • CLEAN        +15  (zero warnings in your digest)
  • ONE_SHOT     +25  (success on first attempt — no retries)

Quality drives the score. There are no points for cutting corners.
Speed only counts when the output is correct.

Agents don't see the killstreak banners. The user does. But agents see the points, which means they have a goal beyond “complete the task.” The viewer is the broadcast. The brief is the rules.

🎛️ Try It Yourself

This is the actual scoreboard widget rendered live in your browser, hooked up to the same FX. Press the buttons. Watch the numbers go up.

🔁 The Pipeline End-To-End

Below is the path a single point takes from a worker's result.json all the way to the +N popup on screen. Click any node for what it does.

It's a 200-line scoring engine and ~600 lines of CSS / JS in the viewer. The runner already had everything else.

📉 Before vs After

🐛 Two Bugs We Caught Out Loud

The first build of the FLAWLESS VICTORY overlay used a fixed 140px font size. At 900px wide, the giant neon letters punched straight off the right edge of the screen. So we caught the bug, complained about it on Twitter in our heads, and shipped the fix.

Then we re-tested and found the killstreak banners did the same thing — 92px fixed for TRIPLE KILL!, 110px for GODLIKE!, both clipping at narrow viewports. clamp() with a viewport-relative middle value fixed both.

🏆 The Victory Easter Egg

When a run delivers with no failures and an average judge score above 0.95, the viewer takes the whole stage. Click the trophy below to see it.

💡 What This Buys You