docs/publishing.md

# Publishing the TradeOpsSim release

The release contains a Hugging Face dataset with executable bundles, recorded trials, and the final video; a static Space for watching the video and inspecting recorded runs; and optional Docker Space inputs for a live guided workflow. AgentEnv Hub registration is separate.

## Assemble

The prepared `release/` directory is sufficient for upload. Rebuilding the dataset requires the original local AgentEnv trial store; the public source archive intentionally excludes runtime stores and credentials. From the development source with that store present:

```bash
python -m pip install -e '.[dev,integration]'
python -m pip install 'agentenv-hf @ git+https://github.com/earakely-scale/agentenv-hf-plugin@c116bf1b321fa33d671632b0ee1b77406dbe3b58'
python scripts/build_bundles.py --host
python scripts/export_hf.py
python scripts/assemble_release.py
```

The assembler uses an explicit public-file allowlist. The source tarball contains application code, tests, documentation, dependency locks, frozen bundles, and the static viewer source. The Docker Space contains only the application and its build inputs. Dataset media comes from the verified video package; raw render logs, downloaded reference media, local stores and credentials are excluded.

## Hugging Face dataset and Space

Authenticate locally with `hf auth login`, then confirm the identity with `hf auth whoami`. Keep access tokens out of command arguments, documentation and datasets. Select repository IDs belonging to the intended account or organization:

```bash
export TRADEOPS_DATASET='OWNER/TradeOpsSim-AgentEnv'
export TRADEOPS_SPACE='OWNER/TradeOpsSim'

python scripts/export_hf.py --repo-id "$TRADEOPS_DATASET" --space-id "$TRADEOPS_SPACE"
python scripts/build_replay_space.py --dataset-id "$TRADEOPS_DATASET"
python scripts/assemble_release.py --dataset-id "$TRADEOPS_DATASET"

hf repos create "$TRADEOPS_DATASET" --type dataset --public --exist-ok
hf upload "$TRADEOPS_DATASET" release/huggingface . --repo-type dataset \
  --commit-message 'Publish TradeOpsSim environment and recorded trials'

hf repos create "$TRADEOPS_SPACE" --type space --sdk static --public --exist-ok
hf upload "$TRADEOPS_SPACE" release/space-static . --repo-type space \
  --commit-message 'Publish TradeOpsSim video and recorded walkthrough'
```

Replace `OWNER` before running. Use repository names dedicated to this release. These commands do not delete destination files. The static Space serves recorded results and the video without running an agent or a backend.

On October 15, 2026, Hugging Face rejected Docker Space creation on this account with HTTP 402: hosting Docker or Gradio on `cpu-basic` requires a PRO subscription. Static Spaces remain free. The release therefore publishes a static viewer and keeps the complete runnable environment in the dataset.

After upload, record the returned commit IDs. Download the dataset at the published revision into a fresh directory, verify `SHA256SUMS`, and run its baseline bundle:

```bash
hf download "$TRADEOPS_DATASET" --repo-type dataset --revision COMMIT_SHA \
  --local-dir /tmp/tradeops-published
cd /tmp/tradeops-published
shasum -a 256 -c SHA256SUMS
agent-env run ./bundles/tradeops-baseline --task late_fill_duplicate
```

Verify the static Space loads, the video plays, both recorded cases render, and download links resolve. No model execution or fresh evaluation occurs on the static page.

## Optional live Docker Space

With an eligible Hugging Face plan, create a separate Docker Space and upload `release/space`. That application has one shared synthetic workspace and no model credentials. **Guided workflow** runs the transparent rules baseline. Run the application locally for isolated sessions or host CLI model execution.

Confirm the Docker Space finishes building, `/healthz` responds, and the main guided case reaches 600 filled, 400 cancelled, zero remaining, and +600 booked. A successful upload is distinct from a healthy running Space.

## AgentEnv Hub

Follow [hub-publishing.md](hub-publishing.md). The helper requires an explicit destination configuration and project ID. It verifies the frozen bundle, pins all references, retains ownership, and writes a registration receipt. The Hub runtime and configured sandbox must support the registered environment and task-step versions.

## Video

`release/huggingface/media/tradeops-demo.mp4` explains how a team at GTS could use AgentEnv to develop and evaluate a trade-reconciliation agent. It shows recorded synthetic cases and outlines a proposed internal review trial. The same directory contains `poster.png`, `captions.srt`, `captions.vtt`, and `TRANSCRIPT.md`. `release/huggingface/downloads/tradeops-video-package.zip` contains the editable HyperFrames source, frozen render assets, model-run evidence and provenance. No reference film footage or branding is included.

The video is prepared for sharing; this release process does not post to social media.