-
Notifications
You must be signed in to change notification settings - Fork 1
hot
Aryan Iyappan edited this page Apr 28, 2026
·
3 revisions
A ~500-word semantic snapshot of recent activity. Updated after every major write operation.
- [2026-04-28] UPDATE removed decisions/ pages (ADR-008, ADR-009) — folded into concept pages. Updated Home.md with wiki index.
- [2026-04-28] INGEST sources="wiki/harness-implementation-plan.md + 8 layer docs" pages_created=11 — full agentic harness knowledge base: 8 layer concepts, 2 ADRs, 1 project page
- Agentic Harness Implementation: 8-layer mandatory pipeline for ultimate-pi. Key ADRs: black-box QA (ADR-008), claude-obsidian Mode B for memory (ADR-009). Build order: schemas → memory → spec → planner → execution → QA → critics → observability → Archon.
- The harness pipeline is always-on — layers cannot be disabled, only tuned
- QA tests are generated from spec ONLY (ADR-008) — prevents gaming
- Memory layer uses claude-obsidian skills (ADR-009) instead of Vectra — eliminates ~87MB deps
- Search is LLM-native: hot.md → index.md → pages (3 depth modes)
- Archon handles orchestration (DAG, loops, approval gates); pi.dev extensions handle intelligence
- Token budget: ~17,500 per subtask, ~83,500 for typical 5-subtask plan
None yet.