ORNITH PIPELINE — session package
=================================
Local agentic coding pipeline built for an 8GB Apple Silicon Mac, plus this session's
full transcripts. Models are NOT included (re-download per instructions below).

CONTENTS
--------
ornith-pipeline/        The pipeline (code, docs, configs). Highlights:
  README.md, docs/      Full docs (ARCHITECTURE, WORKFLOW, BACKEND_SETUP, MODEL_GUIDE,
                        USER_GUIDE, EVAL_GUIDE) — start here.
  ornith                CLI launcher (init / doctor / up / tui / coord / mcp / run / verify)
  pipeline_core.py      The batch pipeline as one callable (interpret->discover->plan->execute->verify)
  coordinator.py        Local-model coordinator (high-level tools: apply_change/discover/done)
  tui.py                curses TUI over the coordinator
  hermes_agent.py       Low-level interactive tool-calling agent
  mcp_server.py         Stdlib MCP server exposing apply_change/discover/verify (for Hermes/Claude Code)
  ornith-apply          Bridge so Aider's /run can call the pipeline
  think_off_proxy.py    :11999 reverse proxy (think:false + num_ctx) in front of Ollama
  planner_v2.py         Grammar-constrained per-file planner
  execute_forge.py (v2) str_replace executor via forge, windowed context + completeness loop
  verify.py             Generic, OS-resolving compile gate
  build_conventions.json  Ecosystem build data (dotnet/rust/go/node/maven/gradle/python)
  machine.json          Per-machine env shim
  Modelfile.iq3m/.plan/.32k  Ollama alias definitions (num_ctx capped)
  v2/discover_cm.py     codebase-memory discovery (+ word-boundary grep fallback for properties/types)
  NOTE: v2/codebase-memory-mcp (macOS-arm64 binary) and models were EXCLUDED — re-fetch per your OS.

transcripts/            Full session JSONL (75652420 = pre-compaction, 8120feb3 = continuation)
configs/                hermes-config.yaml (Hermes wired to local Ollama + ornith MCP),
                        aider.model.settings.yml, claude-mcp.example.json

RE-DOWNLOAD / SETUP ON A NEW BOX
--------------------------------
1. Ollama:  ollama pull maxwell1500/ornith-9b:IQ3_M
            ollama create ornith-iq3m -f ornith-pipeline/Modelfile.iq3m
            ollama create ornith-plan -f ornith-pipeline/Modelfile.plan
            (serve with OLLAMA_FLASH_ATTENTION=1 OLLAMA_KV_CACHE_TYPE=q8_0)
2. codebase-memory-mcp: download the build for THIS OS (Linux/Windows) -> ornith-pipeline/v2/codebase-memory-mcp
3. forge:   pip install "forge-guardrails[anthropic]"   (its own venv/uv on Windows)
4. Proxies: python3 ornith-pipeline/think_off_proxy.py   +   forge-proxy (see docs/BACKEND_SETUP.md)
5. Run:     python3 ornith-pipeline/ornith doctor   then   ornith tui  (or wire Hermes via configs/)

CONTEXT NOTE
------------
On >=16GB (e.g. duskbox, RTX 3050) you can run Ornith at 64K context, which unlocks the real
Hermes TUI (Hermes hard-requires >=64K; the 8GB Mac cannot fit it). On the Mac keep num_ctx=8192.
See docs/MODEL_GUIDE.md and the transcripts for the full reasoning.
