Fieldwork is a set of eight plugins for ChatGPT and Codex. Each one wraps a single-file Python CLI behind a SKILL.md, so the agent runs a bounded command that prints JSON instead of improvising. The catalog site downloads the same scripts, runs them in Pyodide on committed public inputs, and prints what each tool prints. The tools cover Minecraft frame captures, video size caps, exact CSV aggregates, release checks, file tidying, CAD drawings, GitHub Actions spend and glTF budgets.
StatusShipped, catalog live on Vercel, not in the OpenAI plugin directory
RuntimePython 3.11+ CLIs, Pyodide 314.0.7 in a module worker, static site
Proof hooks80 unit tests, eight catalog demos run against the real tools, ZIP smoke from clean extraction

How it works

Each plugin lives in plugins/<name>/ with a portable plugin.json, a Codex manifest in .codex-plugin/, and a skill directory holding SKILL.md and one script. The SKILL.md tells the agent which subcommand to run and how to translate a request into its arguments. tablebeam.py, for example, takes a JSON spec with where, group_by and aggregate, and its skill tells the agent to ask rather than guess when a column is ambiguous. Seven of the scripts use only the standard library. CAD Rescue needs ezdxf and Fit to Upload calls ffmpeg and ffprobe. Every command prints one JSON envelope with status, summary, data and warnings. Inputs are explicit paths, and output files must not already exist. Safe Tidy, for example, writes plans and receipts with O_CREAT | O_EXCL and moves a file by linking it into place before it unlinks the original (_move_link_then_unlink). build_site.py copies site/, each plugin’s icon and CLI, and every demo input into .artifacts/site, and inlines catalog.json into index.html. In the browser, worker.js loads Pyodide 314.0.7 from jsDelivr in a module worker, writes the sample inputs into /work, fetches the tool’s script and runs each step of the demo. harness.py compiles the script once and executes it as __main__ with the demo’s argv, capturing stdout, stderr, exit code and run time. The browser lacks three things the tools use, and harness.py substitutes each one per run.
  • install_link_fallback replaces os.link with an exclusive create and a byte copy. The copy doesn’t share an inode, so Safe Tidy’s apply refuses to run and its demo stops at the plan.
  • XhrOpener routes ShipProof’s HTTP request through a blocking XMLHttpRequest.
  • prepare replays recorded ffprobe and ffmpeg output for Fit to Upload from recorded-binaries.json. The byte budget is still computed live from that metadata.
run_demos.py runs the same demo steps with a local interpreter. It fails when a tool exits non-zero, prints non-JSON, reports error, omits a value the site displays, or skips a file the demo promises. It also checks that the recorded ffprobe size matches the committed clip.

Hosting

The site is on Vercel at fieldwork-plugins.vercel.app, built by vercel.json with python3 scripts/build_site.py --output .artifacts/site. No workflow in the repository deploys to Vercel. The Pages workflow publishes the same build to fortunexbt.github.io/fieldwork-plugins on each push to main. On 18 November 2026 both answered HTTP/2 200, and both had been modified at 18:41 UTC, the minute PR #2 merged. The README links the Vercel URL. The repository homepage, catalog.json and every plugin manifest’s website, support, privacy and terms URLs still point at the GitHub Pages copy.

Verify it

These ran on 18 November 2026 against main at 16470b5, with Python 3.11 on an Apple Silicon Mac:
The unit tests printed Ran 80 tests in 15.714s and OK (skipped=2). ffmpeg and rsvg-convert were installed, so the Fit to Upload and CAD Rescue preview tests ran. run_demos.py printed ok for all eight plugins in 1.5 s. The release ZIP build and archive smoke weren’t run for this page. CI runs them on every push. The 3.13 job of the CI run for 16470b5 passed, and the 3.11 job was still running when this page was written. To run one tool by hand:
It returned status needs_attention with valid false and the reason Missing route completion sidecar. It still measured the capture: 10,569 frame intervals, 502.6 average FPS, a 15.71 ms p99 frame time and a 53.5 fps 1% low. The Tablebeam example in the README (query samples/tablebeam/penguins.csv --spec samples/tablebeam/penguins-query.json) returned the mean body mass per species over 342 rows as exact decimals, such as 3700.662251655629139072847682119205 for Adelie.

Limits

  • The plugins aren’t in the OpenAI plugin directory. The README says submission waits on developer identity verification. There are no GitHub releases, so you install from the repository or build the ZIPs with scripts/release.py build.
  • The catalog site doesn’t run everything for real. Fit to Upload replays recorded FFmpeg output, and Safe Tidy stops at its plan because the browser filesystem has no hard links.
  • ShipProof’s demo and run_demos.py fetch from jsDelivr, so they need network access.
  • Two unit tests skip unless you supply a real DWG file and LibreDWG runtime, or raw Minecraft captures. DWG support in CAD Rescue depends on that optional runtime.
  • Frame Lab measures CPU frame-production intervals, not GPU or display latency. Asset Check checks budgets and missing resources and doesn’t measure frame rate.
  • The Frame Lab demo capture can’t be ranked because its harness writes route.csv, not the route.json sidecar Frame Lab expects. The site shows that rejection on purpose.
  • The Vercel deployment has no workflow in the repository, so it can drift from main unless it’s redeployed.
  • CAD Rescue is GPL-3.0-only. The rest of the repository is MIT.

Source

Reviewed 18 November 2026 against 16470b5.