Project Status

This page reflects the engineering reality surfaced by the architecture trace
campaign as of `ARCHITECTURE_MAP.md` v0.7.0 (2026-05-13). It supersedes earlier
roadmap documents that framed designed-but-not-wired surface as shipping features.
For navigation, the [interactive workspace map](https://lazyxeon.github.io/AstraWeave-AI-Native-Gaming-Engine/architecture/)
visualises the same information.

What this project is

AstraWeave is an active, solo-built, research-grade AI-native game engine in development as the runtime for the Veilweaver game project. It is not a shipping engine product. Its distinguishing engineering practice is the architecture trace campaign — a 13-subsystem forensic documentation effort that produces evidence-grounded, version-controlled traces of how the engine actually works, separate from how older documentation aspirationally describes it.

The engine is built through AI-augmented development under the Genesis Code Protocol (GCP). The trace campaign is GCP applied at the documentation/audit meta-level — a counterweight to the AI-generated drift that otherwise accumulates in a 850K+ LoC workspace.

What’s working today

• Deterministic ECS substrate — astraweave-ecs archetype storage, generational Entity { id, generation }, deterministic single-threaded scheduler (8 stages, executed in fixed order per tick). ParallelSchedule removed 2026-04-18; parallelism lives at the subsystem level (rayon, tokio, GPU compute) per docs/audits/parallel_schedule_removal_2026-04-18.md.
• AI-first runtime loop — WorldSnapshot → AIArbiter → Orchestrator.plan() → PlanIntent → tool_sandbox → engine-side validate_and_execute. 12,700+ agents at 60 FPS validated. Canonical GOAP + Behavior Trees + LLM orchestrator hybrid (astraweave-ai, astraweave-behavior, astraweave-llm).
• Rendering — astraweave-render 78K LoC + 71 WGSL shaders. Disney BRDF + multi-scatter PBR, 4-cascade CSM shadows, IBL cubemaps, clustered forward+, Lumen GI and VXGI (both implemented), TAA, SSAO/GTAO, SSGI, SSR, volumetric fog, god rays, atmosphere, particle system, impostor LOD3. Editor is fully unified onto this pipeline post-Fix-27 (April 2026).
• Terrain — astraweave-terrain with climate field, Whittaker biome lookup, per-biome parameter blending, regional archetype variation, 32-layer material pipeline driven by MaterialLibrary in astraweave-render.
• Visual editor — tools/aw_editor 224K LoC, 41 panel types, 49 panel files, ~9,397 test annotations. Editor depends on astraweave-render non-optionally post-Fix-27.
• Physics — astraweave-physics wrapping Rapier3D 0.22. PhysicsWorld + CharacterController. Send + Sync.
• Foundation verification — Miri (977 tests across ecs, math, core, sdk with zero undefined behavior) + Kani proofs (71+ harnesses across safety-critical crates).

What’s in active development

• Veilweaver — the game project this engine is built for. Vertical slice shipped February 2026; full game development is the next 12-18 months of primary effort.
• Editor Multi-Tool Architecture Campaign (Sub-phase 3) — Mediator Brush diagnostic, Round 8 closure. §7.7 wrapped-component resource identity trap surfaced at four layers; Real-Fix.A/B/C landed, Real-Fix.D pending. See aw_editor.md §1.
• Editor Behavioral Correctness Audit remediation — 37 fixes across 47 commits shipped; per-audit open items still pending Andrew-gate. See docs/current/EDITOR_BEHAVIORAL_CORRECTNESS_AUDIT.md.
• Architecture trace campaign — 13 of N subsystems traced. Treated as part of the production contract by CLAUDE.md. Adding more traces as subsystems cross the non-trivial-surface threshold.

What’s in-design (research surface, not currently runtime-wired)

Per ARCHITECTURE_MAP.md §5.1, the workspace carries ~200K LoC of in-design-but-tested surface — code that passes its own tests but has zero production callers. This is the signature of breadth-first AI-augmented development, not a flaw. Each item is documented honestly as awaiting a wiring decision:

Plus dormant scaffolding (TODO-only bodies), orphan source files, declared-but-unused Cargo deps, dormant feature flags, and the aspirational documentation tree from commit 28bc94f21. Full taxonomy in ARCHITECTURE_MAP.md §5.

What’s not shipping yet

The engine itself. There is no v1.0 timeline. The engine ships when Veilweaver ships, which is a 12–18 month horizon as of this update (2026-05-15). API stability guarantees do not exist; consumers should pin to a commit.

Open questions (cross-cutting)

ARCHITECTURE_MAP.md §14 documents 23 cross-cutting open questions surfaced by the trace campaign — decisional items that affect more than one subsystem and require explicit Andrew-gate decisions. Highlights:

• Q1 — Long-term plan for legacy astraweave-core::World (dual-World coexistence).
• Q2 — Two GOAP implementations: consolidation roadmap?
• Q3 — Runtime LLM model default (currently phi3:medium despite Qwen3 doc-comments).
• Q4 — Production-hardening surface bypassed by runtime AIArbiter path.
• Q5 — §7.7 wrapped-component resource identity trap, preventive instrumentation.
• Q11, Q12 — Memory pipeline and Fluids dormancy: production-wire, prune, or rebrand?
• Q17 — Standalone server HMAC vs. XOR sign16 mismatch.
• Q20 — Editor god-struct refactor (EditorApp 123 fields).

The list is exhaustive; the architecture map is the canonical reference.

Further reading

• Interactive workspace map — explore the engine.
• ARCHITECTURE_MAP.md — consolidated synthesis (2,500+ lines, version-controlled).
• 13 per-subsystem traces under docs/architecture/.
• Benchmark dashboard — performance measurements.