Dev Infrastructure Meta Audit — Multiplayer + Classes Readiness

Dev Infrastructure Meta Audit (2026-10-07)

How to make building Dark Cosmos easier, faster and more effective as internet multiplayer and a class system start. Method: a 7-lane audit workflow (Claude harness, tests/verification, autopilot, netcode readiness, class-system readiness, build/feedback loop, external research), a merged plan, two adversarial reviewers checking each recommendation against the repo, then a manual edit pass (git stats, transcripts, scripts verified directly). Read-only; nothing was changed by the audit. Shareable page: https://claude.ai/artifact/XLWbhdrQiSXLFYj2eq26JE

Tags: [Certain] verified in repo · [Likely] strong evidence · [Guessing] judgement.

Uncomfortable truths

1. The team is one person right now. [Certain] Since 2026-08-01: Chris 1,131 commits, Toby 0. Coordination tooling (collision hooks, ownership protocols, CI for a second pusher) solves a problem that doesn’t currently exist. The real bottleneck is Chris’s attention: ~30 open 📍 memory threads, mostly awaiting an eyeball or on-device verdict.
2. The netcode go/no-go hasn’t run. [Certain] Phase 0 of docs/superpowers/specs/2026-10-07-networked-coop-design.md is measuring cross-platform float determinism. Still unmeasured; 26 sim files use sin/cos/atan2/pow. If iPhone/ATV/Android/web disagree, the answer is Apple-only cross-play or a fixed-point rewrite.
3. “Class system” isn’t designed. [Certain] Archetype tags are build labels (the spec excludes class locks/role bonuses/co-op changes); ShipBonuses.TABLE has 7 hulls. Co-op P2 gets no hull stats at all (sim/meta_run_setup.gd applies to P1 only).
4. High throughput, low closure. [Certain] ~1,130 commits in the month to 2026-10-07, ~213 of them fix/revert (~19%). 56 git worktrees, 69 merged branches, NEXT_TASKS.md 47 KB, 84 memory files, 15 of 27 bh-* skills with no recorded Skill-tool invocation in 75 transcripts.

Plan

Tier 0 — this week, all small

1. Cross-platform determinism probe on real devices (M, high). Dev command: seed-1234, 3,600 ticks, net_checksum() every 600 ticks → Remote Console. Run on Mac/iPhone/iPad/ATV/Android/web. Decides internet-play scope. No CI needed (Linux x86 runners say nothing about ARM devices).
2. Fix the two run-costing tooling bugs (S). scripts/economy-careers.sh ends with rm -f "$snap" under set -u → unbound-variable exit on every run + leaked temp dir (should be rm -rf "$snapdir"). test_remote_config_state reads the real user:// dir (hit the gate 3× on 2026-10-07) → inject a test dir.
3. iOS multicast entitlement request now (S). Apple lead time days–weeks. Add NSLocalNetworkUsageDescription to iOS presets (missing). Manual-IP fallback ships meanwhile.
4. One “waiting on Chris” decision page instead of 30 📍 threads (S, high). Memory holds facts, not a to-do list; sessions read verdicts from the page.
5. Prune worktrees/branches and keep them pruned (S). scripts/dev/prune-worktrees.sh (merged into origin/main + clean → remove, dry-run first); SessionStart hook reports count > 15.

Tier 1 — netcode tooling before more netcode code

6. Two-peer bot harness scripts/net-pair.sh (M, high). Host + client headless with SkilledAutoPlayer over LoopbackTransport/ENet 127.0.0.1, flags for latency/jitter/loss/reorder, checksum compare every 30 ticks, first-divergence report. Replaces four separate suggestions (lan-pair-test, chaos DevFlags, networked soak, LAN pair test). Injection knobs already exist in LoopbackTransport but only inside GUT.
7. Per-tick state diff (M, high). Sim.snapshot_dict() + differ → “tick 156, enemies[12].x”. Serves #6, desync telemetry, and determinism re-pins (today pass/fail + manual bisect). On NEXT_TASKS.md since 2026-07-16.
8. Lint new direct sim. writes outside SimCommands (S, high). Gate test with an allowlist ratchet over the ~90 existing out-of-tick mutation sites in main.gd/ui/; the list only shrinks.
9. Lockstep fault-injection tests (S). Corrupt checksum, reordered input, pick not in offer, wrong pilot, duplicate pick, stale packet — additions to test_lockstep_session.gd.
10. Pre-push gate stamp instead of CI, for now (S). gate.sh writes .git/gate-ok-<sha> on green; a pre-push hook refuses main without it. Build GitHub Actions when a second committer is active or the web build needs coverage.

Tier 2 — before the first class

11. Decide what a class is (brainstorm). Hull class, pilot role, or both; per-pilot or per-team; lobby selection. Blocks #13–#16. Peers (Brotato, DRG Survivor, Death Must Die) use characters with starting kit + stat skew rather than hard role locks. [Likely]
12. bible.json validation in the gate (S, high). Schema + node test: required fields per category (archetypes, gold, biomass…) and cross-references. Catches the “silently zero” class (XP-was-zero shipped for months).
13. Per-pilot hull application + pre-tick agreement (M, high). MetaRunSetup loops pilots; networked runs carry every pilot’s hull/class in the welcome frame and validate before tick 0.
14. One bot batch runner scripts/bots.sh (M, high). One JSON row per run (build, seed, hull/class, profile, pilots, run_time, death cause, gold, DPS); summary by class × skill profile; diff vs the previous build. Then recalibrate skill profiles (n=18 human runs, predating two AI reworks). Today soak/economy_sim/autopilot_bench/sim_bench/difficulty_probe/story_trace each have their own output.
15. Run-twice determinism per hull/class × pilot count (S). Self-consistency (+ one replay pass), not more pinned hashes — avoids re-pin churn on every balance change. Keep the three existing pins.
16. bh-add-class skill + completeness test (S). The test iterates the class table and asserts stats, art, bible entry, economy-sim profile and co-op application all exist.

Tier 3 — Claude setup hygiene

17. Trim the skill roster (S). Recorded invocations (75 transcripts): bh-telemetry 24, bh-sdd-concurrent 22, bh-dev-chunk 14, bh-deploy 12, bh-worktree-device-install 10, bh-merge-reviewed-branch 8, bh-vfx-iteration 6, bh-appstore-release 6, bh-overnight-audit-loop 4, bh-diagnose-render-timing / bh-deploy-mac-dev / bh-android-tester-build 2. The other 15 show none (counts may miss cron/slash firings — check each before archiving).
18. Netcode checklist in bh-reviewer, not a new agent (S). Triggered on sim//net/ diffs: RNG stream, new out-of-tick mutation, quantization, protocol version, replay round-trip test. Dispatch counts: general-purpose 320, bh-implementer 300, bh-reviewer 228, bh-visual-qa 14, bh-feedback-triage 2 — move general-purpose code work onto bh-implementer.
19. Shrink per-session context (M). CLAUDE.md 26 KB (warning at 30). Each rule made structural (#8, #10, #12) earns a CLAUDE.md deletion. dark-cosmos/<role>/ROLE.md has zero transcript references — delete or drop its table row.

Overruled from the merged plan

Merged-plan item	Call	Why
GitHub Actions CI, “Now”, XL	Deferred	One active committer; Linux runners can’t test ARM determinism; #10 gives the guarantee at S.
Pre-commit branch-collision hook	Dropped	Collisions came from Chris’s own parallel Claude sessions; worktree isolation covers it.
Toby coordination protocol	Open question	Only matters if Toby returns.
Per-class pinned determinism hashes	Replaced	#15 run-twice self-consistency.
Class balance invariants (“Interceptor finishes by wave 5”)	Dropped	Invented classes.
ui-sweep baseline determinism fix	Refuted	Already fixed 2026-09-23.
Parallel device installs, co-op telemetry session id, chaos DevFlags	Later	Once LAN device testing is routine; #6 covers chaos headlessly.
“26K correction messages”, “hooks dir empty”	Discarded	Grep noise; hooks dir has check-claude-md-size.sh.
LLM balance-testing agents (TITAN-style)	Not yet	Deterministic bots + #14 give the signal cheaper; revisit once classes exist.

Keep

• Pure-RefCounted /sim + fixed-tick determinism — the base for lockstep, bots and replay.
• Host-scheduled input lockstep with quantized inputs; rollback correctly rejected at 2–4 ms/tick on ATV. External research: Vampire Survivors’ online co-op used a host-canonical input table too (with 1 Hz position sync + local prediction for its own enemy model).
• bh-implementer → bh-reviewer in worktrees.
• Telemetry v2, Remote Console, remote config.
• gate.sh + test-count guard; 463 GUT scripts.

Stop

• One memory file / CLAUDE.md line per incident when a test or hook could prevent it.
• New overnight loops while ~30 verdicts are pending.
• Unpruned worktrees, branches and BUILD numbers.
• More netcode before Phase 0 is measured.

Open questions for Chris

1. Is Toby active again for multiplayer? (Moves CI + ownership split to Tier 0 if yes.)
2. Apple-only cross-play acceptable if Android/web diverge? (Alternative: fixed-point rewrite.)
3. What is a class — hull, pilot role, or both? Can co-op pilots differ?
4. Archive the 15 skills with no recorded use (after checking cron/slash usage)?

Chris’s answers (2026-10-07)

1. Toby may come back, but as a playtester who sends change requests, not a coder. Dev stays effectively solo: CI and an ownership split stay deferred; the tester-feedback pipeline (bh-feedback-triage, in-game feedback) matters more than multi-dev tooling.
2. Target is full cross-play eventually: Apple, Android, web AND PlayStation, Xbox, Switch. That spans x86-64 (PS/Xbox), ARM (Switch/Apple/Android) and WASM, so “Apple-only” is off the table. The determinism probe now decides between float lockstep (if every platform matches) and a fixed-point sim / different netcode model (if not). Consoles also need a third-party Godot porting route (Godot ships no console export) — verify options before committing.
3. Classes follow EVE Online / World of Warcraft / League of Legends archetypes: some clearly defined roles first (tank / healer / damage / control style), more exotic hard-to-define ones later. Needs a class design spec; the career archetype tags are the starting vocabulary. (EVE-style economy/fitting/loss still belongs in ~/Claude/sovereign; class roles are fine here.)
4. Don’t archive yet: audit every skill — why it was made, why it went unused, and how to fix it so it gets used.

External research notes (from the research lane)

• Klotho (Godot Asset Library, 2026-07): deterministic lockstep + rollback with FP64 fixed-point — reference only if Phase 4 latency forces rollback.
• Backends for accounts/matchmaking later: Nakama (self-host), Talo, Crux, PlayFab. Not needed for LAN/dev relay phases.
• Cloudflare Durable Objects relay: use the WebSocket Hibernation API so idle rooms bill ~zero; spec cost model ~USD 0.003 per 20-min 2-player match.
• Float determinism: x87/FMA contraction/ARM NEON/WASM differences mean IEEE 754 per-op guarantees don’t make sequences identical across platforms — measure, don’t assume.