Publish v3 as the airops-v3 bundle on the b6-v3 pack

.github/workflows/ci.yml

@@ -26,7 +26,7 @@ jobs:
 
   test:
     runs-on: ubuntu-latest
-    timeout-minutes: 20
+    timeout-minutes: 40
     strategy:
       matrix:
         python: ["3.11", "3.12"]

Dockerfile.v3

@@ -0,0 +1,17 @@
+# The AirOps v3 env: the airops-v3-day MCP server with the b6-v3 pack, on port 18765.
+#   Build: agent-env airops-v3 setup   (docker build -f Dockerfile.v3 -t agentenv-airops-v3-env .)
+FROM python:3.12-slim
+LABEL org.opencontainers.image.source="https://github.com/earakely-scale/agentenv-airops-plugin" \
+      org.opencontainers.image.licenses="Apache-2.0" \
+      org.opencontainers.image.description="AirOpsEnv v3 as an AgentEnv environment. Built from public U.S. records: BTS on-time data of every carrier, the FAA registry, FAA Command Center advisories and NWS METARs, SPECIs and TAFs via the Iowa Environmental Mesonet."
+ENV PYTHONUNBUFFERED=1 PIP_NO_CACHE_DIR=1 PIP_DISABLE_PIP_VERSION_CHECK=1 AIROPS_V3_TASKS=/app/data/b6-v3
+RUN pip install "agentenv-framework-protocol==0.1.338" "fastmcp==3.4.8" "mcp==1.30.0" "pydantic==2.13.5" "tzdata" \
+    "httpx==0.28.1" "numpy==2.5.4" "pandas==3.0.6" "highspy==1.15.1" "airportsdata==20260905"
+WORKDIR /app
+COPY LICENSE NOTICE /app/
+COPY data/b6-v3 /app/data/b6-v3
+COPY src/agentenv_airops /app/agentenv_airops
+RUN useradd --create-home --uid 1000 airops
+USER airops
+EXPOSE 18765
+CMD ["python", "-m", "agentenv_airops.v3day.server"]

Dockerfile.v3.dockerignore

@@ -0,0 +1,10 @@
+*
+!LICENSE
+!NOTICE
+!data/b6-v3
+!src/agentenv_airops
+src/agentenv_airops/data
+src/agentenv_airops/web
+src/agentenv_airops/bundles
+**/__pycache__
+**/*.pyc

README.md

@@ -391,6 +391,30 @@ The [replay Space](https://huggingface.co/spaces/earakely-scale/AirOpsEnv-AgentE
 viewer (`web/`) under `viewer/`, reading the episodes straight from the dataset's resolve URLs at the tag, which
 huggingface.co serves with CORS headers. `--space-out` assembles it and `--space-repo` pushes it.
 
+## AirOpsEnv v3: the realism layers
+
+v3 plays the same whole-network storm days from 04:00 inside the FAA's programs, with the constraints v2's world
+leaves out, each enforced alike in the sim and the hindsight MIP: hub departure caps from every carrier's BTS
+pushbacks, ramp closures from lightning at the field (METAR and SPECI thunderstorms, as departure and gate-in rate cuts
+at a rate fitted on train dates) and outstation arrival caps. The desk sees them only as it could: caps 4 hours ahead,
+a closure once it begins, and the FAA advisories, METARs, SPECIs and TAFs at their delivery lags, with a watch each hour
+and one more on each event. One `airops-llm` desk plays a day on `airops-v3-day` and is graded as v1 is, against the
+hindsight optimum. The pack is b6-v3 (`data/b6-v3`: 94 storm dates, 24 eval and 70 train, and 2 controls, with each
+date's anchor and follower costs stored, so a load never waits on a solve); `docs/v3/DESIGN.md` and
+`docs/v3/REPORT.md` give the design and its check against v2.
+
+```bash
+agent-env airops-v3 setup --agent            # builds airops-v3-day (Dockerfile.v3) and airops-llm
+agent-env run airops-v3 --task wiring-noplay # the first eval date flown with no change, then the grade
+agent-env run airops-v3 --task b6-20240630-0400 --model anthropic/claude-sonnet-5-5
+agent-env hf run earakely-scale/AirOpsEnv-AgentEnv --bundle airops-v3 --task wiring-noplay
+```
+
+The dataset holds the bundle `airops-v3` (the eval dates and `wiring-noplay`) from the first release
+`scripts/hub_dataset.py` builds with it. `agent-env airops-v3 tasks --split train` writes the training dates as a
+bundle; `scripts/v3_pack.py --bases none --refs` builds the pack and `scripts/ship_v3_pack.py` ships it as
+`data/b6-v3`.
+
 ## AirOpsEnv v1: one desk at one hub
 
 v1, released in v0.2.0, is one operations desk at one hub in the 3 to 6 hours of a storm day where planning matters