Skip to content

roadmap.md

Latest commit

 

History

History
118 lines (92 loc) · 5.9 KB

roadmap.md

File metadata and controls

118 lines (92 loc) · 5.9 KB
Raw

Roadmap

Planning index only. Architecture and acceptance tests stay in spec.md. C4 views: c4.md. Do not put hostnames, LAN addresses, or SKUs here.

Current line

Phase Status What it is
1 — inference host Downstairs NVIDIA path is the live target; IDE box is MCP-only Ollama on downstairs WSL GPUs; starter tags; no paper inference on the IDE machine
2 — MCP bridge Done local-coding-slm stdio tools on the IDE workstation; Cursor / Copilot / Claude adapters
3 — measure Protocol + corpus + apply gate + CI; live rates need downstairs powered Layered scoring; next live rows go through downstairs
T12 Part A downstairs Next hardware work (early Nov 2026) Second PC / WSL NVIDIA via SSH — was blocked on host power; examples/downstairs-wsl-gpu.md
T12 Part B Copilot A8 / Claude A9 Config ready; operator clicks pending Same-machine only; do not block on downstairs; docs/a8-a9-operator-checklist.md

Order (do not invert): finish Phase 3 live measurement on downstairs (T12 Part A) before treating Halo as the next lab.

  1. First work — downstairs: power on, WSL Ollama on the usable NVIDIA GPUs, ssh -L from the IDE workstation, A4-class check, scrubbed live harness / run_eval.py --live rows.
  2. Then — keep measuring Part A on downstairs through November.
  3. Later — Halo / 395 (~Black Friday): bring-up and A13 only after downstairs is measuring.

Public-safe downstairs notes: examples/downstairs-wsl-gpu.md.

Halo / Ryzen AI Max+ 395: planned acquire around Black Friday week Nov 2026. Claim Phase 4 only after A13. Until then Phase 3 live rates stay on downstairs. The IDE workstation runs MCP only — do not treat local Ollama there as a paper or Phase 3 path. Paper track: embabel-slm paper calendar.

Paper track: Nov 2026–Mar 2027 Zenodo preprint (DOI by 31 Mar 2027). Live pass@1 / pass@end freeze by early January (downstairs default; Halo only if A13 green); all paper numbers freeze by late February.

Phase 4 — Halo host profile (planned ~Black Friday Nov 2026)

Use an AMD Ryzen AI Max+ 395 / Halo-class box as another private Ollama host. Target arrival: around Black Friday week November 2026. The MCP server stays on the workstation. Premium agents still plan and review.

Until the box exists and A13 passes, do not claim this phase — keep Phase 3 live rates on the downstairs host (T12 Part A). Do not skip downstairs to chase Halo.

This is not a second product. It is the same bridge with a different OLLAMA_BASE_URL (or the same loopback URL behind ssh -L).

Public-safe notes: examples/halo-ryzen-ai.md.

Already satisfied (do not redo)

  • Stdio MCP on the workstation; no listening port
  • Tools: local_status, local_code, local_refactor, local_generate_tests, local_explain, local_review
  • Env contract: OLLAMA_BASE_URL, OLLAMA_FAST_MODEL, OLLAMA_STRONG_MODEL, OLLAMA_NUM_CTX
  • Starter tags: qwen3.5:9b / devstral-small-2
  • No public tunnels; cloud agents out of scope
  • Official Ollama library tags only; output treated as untrusted
  • Deployment checker (A12)

Halo-specific work (when the box exists)

  1. Validate current Ollama on the Halo with its supported AMD backend; record whether ROCm or Vulkan is used. Confirm accelerated placement after a short chat before selecting larger models.
  2. Pull the starter pair. Leave larger coding tags for a later benchmark.
  3. Reach the API from the workstation. Prefer ssh -N -T -o ExitOnForwardFailure=yes -L 127.0.0.1:11436:127.0.0.1:11434 user@<halo-host> and keep Ollama on Halo localhost. Use any unused workstation port in place of 11436. A private-interface bind plus a workstation-only firewall is optional.
  4. Point gitignored .env at http://127.0.0.1:11436. Reload desktop MCP. Re-run scripts/check_deployment_safety.py.
  5. Repeat acceptance A1–A7 and A11–A12. Run A4 (the workstation succeeds through SSH; an unauthorized LAN client fails). A8/A9 only if those clients are in use.
  6. After the starter pair feels usable, record a Halo benchmark (model + quant, effective context, tok/s, time to first token, peak unified memory, whether output is reviewable). Do not treat a large advertised context window as a reason to send the whole repo.

Explicit non-changes

  • Do not rename tools to write_code / ollama_status. local_* is the contract already in Cursor, Copilot, and Claude configs.
  • Do not add SLM_* environment aliases unless a later consumer cannot use OLLAMA_*.
  • Do not expose Halo in Cursor's model picker.
  • Do not implement an automatic task classifier (still Phase 3).
  • Do not start Halo install or wrapper changes until this phase is claimed.

Optional hardening (only if Halo work shows the gap)

These are not required to start Phase 4. Schedule them if the Halo path needs them:

  • Wrapper preflight: reject a non-private OLLAMA_BASE_URL and fail clearly when /api/tags is down (no public fallback).
  • Reject oversized tool payloads (files + task over a configured character cap).
  • Keep dual fast/strong timeouts; a single SLM_TIMEOUT_SECONDS is unnecessary unless operators ask for one knob.

Later than Phase 4

  • Automatic routing / classifiers (only after Phase 3 numbers exist)
  • Larger-than-starter models on Halo unified memory
  • Application-level shared secret, only if Ollama can do it without breaking local IDE use

Out of scope (unchanged)

  • OpenRouter or Cursor OpenAI-base-URL override
  • Public Ollama, ngrok, Cloudflare Tunnel
  • Cloud-agent access to the private GPU in this home-lab profile
  • A second MCP server just for Halo