Skip to content

Latest commit

 

History

History
46 lines (37 loc) · 2.65 KB

TASKS.md

File metadata and controls

46 lines (37 loc) · 2.65 KB

Task breakdown

Mirrors the GitHub Issues on this repo (create those from this list — one issue per row, assigned to the named person). Check items off here or on GitHub, whichever the team actually uses day to day.

Delfi — Android app + on-device inference (GPU laptop)

  • Open android-app/ in Android Studio, confirm it builds and runs with StubReviewEngine
  • Pull in llama.cpp's examples/llama.android as the JNI/CMake starting point
  • Bundle Qwen2.5-Coder-1.5B-Instruct GGUF (Q4_K_M), get it loading on-device
  • Implement LlamaCppReviewEngine.review() — prompt, call model, parse output into Findings
  • Swap MainActivity's engine to LlamaCppReviewEngine, confirm real findings on sample-diffs/01-null-pointer.diff
  • See android-app/NOTES_LLAMACPP.md for the detailed plan.

Nambert — laptop-bridge + escalation tier (GPU laptop)

  • Install Ollama, ollama pull qwen2.5-coder:7b (4.7 GB), server running on port 8000
  • Fixed requirements.txt for Python 3.14 (pydantic wheel issue)
  • Confirmed /review via all three sample diffs — 2/3 clean catches, 1 known 7B-model miss (off-by-one), documented and acceptable for demo (see laptop-bridge/README.md)
  • Improved server.py: richer prompt, proper error handling (503/504), /status endpoint, CORS, structured logging
  • Built the laptop-IP settings UI in the app (settings/LaptopSettings.kt, ⚙ button → bottom sheet → persists via SharedPreferences, no recompile needed)
  • Share your laptop's local IP with Delfi/Mukesh for Wi-Fi testing (ipconfig)
  • See laptop-bridge/README.md.

Mukesh — voice trigger, demo, submission (this repo's owner)

  • Implement mic button + VoiceTrigger wiring in MainActivity (runtime mic permission request, "review this" → calls engine.review())
  • Curate/expand sample-diffs/ if more demo cases are needed
  • Record 60–90s screen capture: paste diff → voice trigger → findings, fully offline
  • Fill in Team section of ../phase1-submission-draft.md
  • Turn the draft into the deck (PDF/PPT)
  • Before submitting: flip this repo to public (or add the organizer as a collaborator) — it's private right now, deliberately, until the demo video is ready. The submission form's "Prototype URL" field should point at the demo video once recorded, with this repo as the backup/source-code link — not the other way around.

Shared / whoever gets to it first

  • Decide phone↔laptop pairing UX — manual IP entry via ⚙ settings sheet in the app (done)
  • End-to-end dry run: all three engines working, escalation flow demoed live