rca-umb/five-desks-code

Five Desks Code: agentic coding on real libraries from every domain

Five Desks Code is an OpenEnv environment for the OpenEnv Arena. Every task is real agentic software work: explore an unfamiliar repository, reproduce a problem, read code, run things, change code, check that nothing else broke, and submit. The repositories are pinned releases of open-source Python libraries from the arena's eight domains, so the coding happens inside finance, science, scheduling, subtitles, signing, pipe-flow and media code, and the media tasks use real production tools (ffmpeg, ImageMagick, Pillow).

Why coding: RL on software tasks has been reported to carry over to other reasoning tasks (for example SWE-RL), and on this arena's own leaderboard the software-repair runs that had a learning signal averaged 4.0/40 against about 3.0 for all runs, including two of the three best single runs (a gap inside the 1 to 4 that runs with no learning at all have scored, so it is a hunch, not evidence). Three runs of our own narrower diligence tasks (Five Desks v1-v3) learned their tasks and scored 2, 2 and 1 out of 40.

• Image: ghcr.io/rca-umb/five-desks (code tag, linux/amd64, public); the arena pins the digest.
• Source: source/: environment, tests, the catalog builder and the calibration code.

Tasks

Bug fixing (bugfix-<domain>-l1|l2). Each instance takes one library and one bug from a build-time catalog of 525: a single token changed inside a function (a comparison, an operator, a constant, and/or, a not) that 1 to 8 of the library's own tests catch. The catalog builder plants every candidate change in a container, runs the library's test suite, keeps the changes the tests catch, and then checks each one inside the final image (the untouched repository must fail, the original code must pass, at both levels). Level 1 is a failing CI run: the issue shows the failing test's output and the suite is in the repository. Level 2 is a user report: the issue shows only the failing check, and the tests that cover it are removed, so the policy has to reproduce, locate and verify on its own.

Media production (media-production-l1|l2). Each instance generates its own media and a client spec plus production notes that add one requirement the spec does not mention: a trimmed, scaled clip (with or without a mono audio track), square thumbnails from photos rotated by EXIF (skipping placeholders), a contact sheet ordered by capture time, or a loudness-normalised FLAC delivery (with or without the leading silence).

Tools (names and argument names are drawn per episode, e.g. read_file/cat/open_file, so the policy has to read the tool list): list files, read a file with line numbers, search with a regular expression, replace an exact snippet, write a file, run a shell command (60 s), run pytest (100 s), submit. One generic action for everything: {"tool": "<name>", "args": {...}}.

Reward and grading

All or nothing. A bug fix scores 1 only if every catching test passes and no other test in those files fails that passes on the original library. Grading copies only the library's own source files from the workspace into a clean copy of the original project and runs the original tests there, so edits to tests, configuration or conftest.py change nothing, and a package containing symlinks scores 0. Media deliverables are measured, not trusted: stream properties with ffprobe, the first frame against the right moment of the source, thumbnail and tile content against references computed from a private copy of the inputs, and the peak level and start of the audio.

The policy's commands run as the unprivileged nobody user with CPU, memory, file-size and process limits; the server, the bug catalog and the original sources are root-only. source/codedesk/tests checks: oracle 1 and untouched 0 for every task id, test-file edits and conftest tricks score 0, symlinks out score 0, the catching tests are absent at level 2, the policy cannot read the server, and a set of hostile tool calls (path escapes, a fork bomb, endless output, killing the server) cannot escape or crash an episode.

Calibration (Qwen3.8-27B, thinking off, arena reply parsing)

HF Inference Providers (DeepInfra), temperature 1.0, arena-style system prompt with the action schema, budgets as submitted, replies parsed by their first JSON object (which matched the arena's training rewards on our earlier tasks). The trained tasks, one per domain plus media production:

task id	domain	code	what the policy gets	solved by the untrained model
bugfix-office-l1	office	croniter, isodate, humanize	failing CI test given	0.56 (16)
bugfix-finance-l1	finance	numpy-financial	failing CI test given	0.50 (16)
bugfix-science-l1	natural science	uncertainties	failing CI test given	0.31 (16)
bugfix-industrial-l2	industrial	fluids (pumps, valves, pipe flow, fittings)	user report only; catching tests removed	0.38 (16)
bugfix-media-l2	media	srt, webvtt-py, python-slugify	user report only; catching tests removed	0.56 (16)
bugfix-security-l2	cybersecurity	PyJWT, itsdangerous	user report only; catching tests removed	0.50 (16)
bugfix-software-l2	software engineering	toolz, tabulate, inflection	user report only; catching tests removed	0.69 (16)
bugfix-math-l1	math	more-itertools	failing CI test given	0.50 (16)
media-production-l2	media	ffmpeg, ImageMagick, Pillow	clip, thumbnails, contact sheet or audio delivery from spec.md + notes.md	0.44 (16)

All picked tasks have mixed outcomes, so GRPO groups carry signal. About a third of the losses are replies that are not one valid action (arguments outside args, invalid JSON escapes in long code strings, prose with no action): keeping tool calls well-formed while carrying code is part of what these tasks train. Episodes are long (about 13 tool calls), so a 4-hour run takes fewer optimizer steps than with short tasks.

Libraries

Pinned sources as published on PyPI, shipped with their licenses: croniter 6.2.4 (MIT), isodate 0.7.2 (BSD), humanize 4.16.0 (MIT), numpy-financial 1.0.0 (BSD-3-Clause), uncertainties 3.2.3 (BSD-3-Clause), fluids 1.3.1 (MIT), srt 3.5.3 (MIT), webvtt-py 0.5.1 (MIT), python-slugify 9.1.3 (MIT), PyJWT 2.15.1 (MIT), itsdangerous 2.2.0 (BSD-3-Clause), toolz 1.2.0 (BSD-3-Clause), tabulate 0.10.0 (MIT), inflection 0.5.1 (MIT), more-itertools 11.1.0 (MIT). The bug catalog (which line was changed) is the answer key and is not published here; source/codedesk/build rebuilds it from these versions.