mshumer/Claude-of-Duty

ARCHITECTURE.md is the contract the agents worked against: subsystem interface, directory ownership, the cross-subsystem event vocabulary, and shared surface types.

Tooling

The interesting part of this repo is arguably the harness, not the game.

Two findings worth recording, because both invalidated earlier measurements:

Median frame time hides the actual problem. A static-camera benchmark reported 94 fps while the game was unplayable. Real gameplay at Retina DPR (internal 3.34 MP, not 2.07) ran 12–17 fps with 728–1236 ms stalls caused by 34+ WebGL programs compiling lazily mid-frame. profile.mjs reports p50/p95/p99 and attributes each hitch, which is what surfaced it.

Captures were not reproducible. shotset.mjs reuses one page across all 11 shots, so particle age, decal buffers and exposure state leak forward — two identical runs differed on 10 of 11 shots. baseline.mjs isolates each shot in a fresh page, which is bit-identical and is what makes imagediff.mjs a usable gate.

Performance

Measured on an Apple silicon laptop at 1512×982, DPR 2 (3.34 MP internal), ultra preset, 3 runs, gameplay in motion with AI and firing active:

The optimization pass was constrained to produce zero visual change, enforced by imagediff.mjs rather than by assertion — the shipped build is bit-identical to its pre-optimization reference across all 11 shots.

Shader pre-warm (src/core/prewarm.js) is what removed the stalls. Making it provably pixel-neutral required first fixing subsystems that animated off performance.now() instead of the engine clock, since any change to boot duration otherwise shifted output.

Honest assessment

The goal was to match a modern Call of Duty. It does not.

Eleven independent adversarial critics scored the frames against that bar. Scores went 3.59 → 4.14 → 4.05 → 5.05 out of 10. Two shots reached "CLOSE"; the rest remain "AMATEUR". In a blind A/B, every critic in every round picked the real Call of Duty frame.

Where it falls short, specifically:

• Hands. Blocky finger slabs that don't convincingly grip the weapon.
• Material richness. Surfaces read as procedural noise rather than photographed reality at close range — the ceiling of generating texture from code.
• Characters. Enemies read as mannequins at distance.
• Indirect light. An approximation, not real GI.
• Frame rate. 28–30 fps at Retina. The art passes tripled geometry cost (5.9M → 11.3M triangles) and optimization recovered about half.

A known root cause remains unfixed: the viewmodel light rig in render/index.js delivers roughly 20× the irradiance per unit albedo that the world does — a plain black material in the view scene renders at L=110 against a background of 91, purely from F0=0.04. Every weapon albedo is cheated to a third of physical to compensate, which caps material separation on the most-looked-at object in the game.

Process note

Sequential single-owner passes beat parallel fan-out decisively. Three rounds of six agents each owning one directory moved the score +0.46 and left frame-ruining defects higher than they started (60 → 47 → 66), because tonemapping, sky and indirect light are one coupled system and isolated agents kept breaking each other's assumptions. One sequential pass with a single owner per coupled concern moved it +1.00 and cut defects 66 → 26.

The most valuable single result came from an agent contradicting its own brief. Every critic for three rounds reported the weapon as "untextured". It wasn't — it was specular-dominated, with the diffuse term measured at L=26 against a shipped L=67. Prior rounds had been crushing albedos to fight bright-part complaints, which killed diffuse and made it worse. The fix was the opposite of what was asked for.