PROGRESS.md

PROGRESS — Hackathon Task State

Live task tracking. Updated at the end of every work block. Source of truth = this file.

Started: 2026-04-23 Target submission: 2026-04-25 EOD (2-day sprint)

---

Pre-flight

• Repo created (GeoForce-CCHackathon) + pushed
• initial/ seeded from ForceX-AI v2-real-transform planning docs
• Hackathon research complete (rules, prizes, Discord, prior winners)
• HACKATHON-PLAN.md v2 (dual-tool pivot) committed
• CLAUDE.md, AGENTS.md, PROGRESS.md scaffolded
• .claude/ subagents defined (8)
• .claude/ skills defined (6)
• .claude/ commands defined (4)
• .claude/settings.json — model pinned to claude-opus-4-7
• Fresh .venv + pyproject.toml
• ANTHROPIC_API_KEY in .env (gitignored) + round-trip tested
• .gitignore protects secrets
• [~] Registration + Discord acceptance email — deferred per user (2026-04-23)

Day 1 — Build both engines

Morning (3h) — COMPLETE 2026-04-23

• Copy geoforce_cnn_v1.1.pt to surrogate/weights/ (248,595 bytes, from ForceX-AI/products/model_registry)
• Port ReservoirCNN class + encoding to surrogate/ (model.py, encoding.py, predict.py)
• tests/test_surrogate_smoke.py green — 5/5 pass, 1.7s total
• Claude scaffolding discoverable (subagents + skills load; surrogate-operator.md corrected to match real v1.1 encoding)

Afternoon (4h) — Agent team builds GeoForce-Solver — COMPLETE 2026-04-23

• solver/properties.py — IAPWS-IF97 wrappers
• solver/grid.py — 2D vertical section structured grid
• solver/darcy.py — pressure solver
• solver/energy.py — heat solver
• solver/wells.py — source terms
• solver/coupled.py — implicit backward-Euler
• solver/benchmarks/theis.py — analytical pressure
• solver/benchmarks/conduction_1d.py — analytical temperature
• tests/test_solver_theis.py passes (<5% error)
• tests/test_solver_conduction.py passes (<5% error)

Evening (2h) — CHECKPOINT + wiring — COMPLETE 2026-04-23

• GO/NO-GO checkpoint: both analytical tests green → GO
• tools/predict_solver.py + tools/predict_surrogate.py
• agent/runtime.py — claude-agent-sdk boot
• CLI answers Q1 end-to-end
• Commit + push — milestone: both engines live

Day 2 — Polish + demo + ship

Morning (3h) — Backend + Streamlit fallback — COMPLETE 2026-04-23

• agent/api.py — FastAPI + SSE (+ /predict, + SPA static mount)
• app/app.py — Streamlit fallback
• demo/scenarios.yaml — Q1, Q2, Q3
• tools/monte_carlo.py + tools/sensitivity.py

Afternoon (4h) — React dashboard — COMPLETE 2026-04-23

• dashboard/ scaffold (Vite+TS+zustand)
• tokens.css from claude-design-system skill
• Components: Header, QueryInput, ScenarioPicker, AgentTrace, AnswerPanel, FieldPlot, FieldPanel
• api/client.ts SSE consumer + predictFields helper
• Canvas-based magma heatmap (no plotly — bundle stays <160 kB)
• Dry-run 3 scenarios through /predict + SSE smoke of /query
• Dockerfile multi-stage (node build → python runtime, CPU-only torch)

Evening (3h) — Ship

• Deploy to HF Spaces (Dockerfile space)
• Root README.md + ASCII architecture diagram
• 90s demo video (recorded against React dashboard)
• demo/brady_validation.ipynb (compressed)
• Tag v0.1-hackathon + submit via Cerebral Valley portal

Stretch (only if all above shipped)

• .mcp.json exposing GeoForce-Solver via MCP
• Two-phase stub (saturation variable plumbing, no flash)

---

Fallback Plan (if Day 1 Evening checkpoint fails)

If either analytical-benchmark test fails:

• Drop solver/ from critical path
• Reframe as "Claude Opus 4.7 agent orchestration over deployed surrogate"
• Keep all Day 2 deliverables intact
• Still viable for "Best use of Claude Managed Agents" $5K

---

Decisions Log

Date	Decision	Source
2026-04-23	1st-place target, Managed Agents as safety net	user
2026-04-23	Solver + Surrogate dual-tool	user
2026-04-23	Solver name = GeoForce-Solver (provisional)	claude
2026-04-23	Single-phase water (not two-phase)	user approval of pitch
2026-04-23	NREL Brady for validation cameo	claude proposal
2026-04-23	Day 1 evening GO/NO-GO trigger	claude proposal
2026-04-23	Solver renamed TinyTOUGH → GeoForce-Solver	user
2026-04-23	Two-phase: single-phase for Day 1; optional stub Day 2 evening only if solver green + submitted	user
2026-04-23	Demo framing: Ulubelu-inspired synthetic (liquid-dominated, on-brand)	user approved claude recommendation
2026-04-23	Demo UI: React dashboard (primary) + Streamlit (fallback), Claude design language, HF Spaces Dockerfile deploy	user
2026-04-23	Added 8th subagent: ui-engineer	claude (per user request)
2026-04-23	Added 5th skill: claude-design-system	claude (per user request)
2026-04-23	Added 6th skill: tough-reference (deep TOUGH family knowledge from LBNL manuals)	claude (per user request)

---

Blocker Log

• 2026-04-23 — Solver pressure output for q1_drill_temperature and q3_well_placement reports values ~10² MPa after the unit conversion, which is physically implausible (reservoir should stay around 15 MPa with the small doublet stress). The /predict endpoint itself works; the issue is either a unit in the solver output dict or boundary treatment in solver.coupled. Flagged for the reviewer agent before the demo video is recorded.