Point-in-time state of the harness. Update this when a subsystem ships or a verification window opens or closes. Last updated 2026-06-10.
Three paths are live: template, scaffold, and adopt (repository-setup.py --adopt, shipped 2026-06-10 with collision policy and gitignore merge). Eject is implemented through reference scrub and the harness-eject skill, verified end to end on a real Hub-created Unity project the same day (docs/process/2026-06-10-finalize-unity-adopt-plan.md, Phase 5 results). Unity adoption requires Git LFS; the reference set ships at .github/docs/unity.gitattributes.
The harness is configured. copilot-instructions.md and system-index.md no longer carry <!-- CUSTOMIZE --> markers. The remaining markers live in .gitignore, the Repo Signals block of research.instructions.md (fill it during project setup), and the deprecated patterns.instructions.md placeholder.
Two language-specific code-standards files are populated: csharp-code-standards.instructions.md and lua-code-standards.instructions.md. ADRs are authored under docs/adr/, including the project-memory decision (adr-learn-establish-shared-project-memory) and the scaffolding CLI decision (adr-scaffold-introduce-ah-ide-cli). docs/business-rules/ has no rules yet.
The ah-ide scaffolder ships with four templates (three C# layouts on .NET 10 with NUnit, one Lua), validated by the scaffold-matrix.yml workflow and check 23 of validate-system.py.
The learning pipeline is wired and live, recording observations and producing proposals. One earlier proposal (skill-post-save-revision) was reviewed and applied, adding the Skill Authoring rule to agent-guardrails.instructions.md. The committed project-memory digest is populated and curated through the continuous-learning skill; review any pending proposals with that skill before acting on them.
As of 2026-06-05 the pipeline's signal layer is implemented in full: transcript-parse correction capture and evidence-based session-clock staleness (both ADRs Active), provenance-tagged instincts, the proposal quality gate, the #correction marker, and the contradiction reducer. The verification windows are open: classifier precision is unproven until the first real correction batches are reviewed, per docs/process/2026-06-04-learning-signal-fidelity-plan.md (baseline metrics recorded there).
A volatility-based design pipeline (behavioral-requirements, volatility-decomposition, architecture-layering, orchestrated by architecture-design-pipeline) runs requirements through to a layered architecture under docs/design/{slug}/, following Löwy's Righting Software method.