ShiftDesk v3

An OpenEnv environment for OpenEnv Arena. 47 tasks from 27 procedural families covering all eight arena domains (software 5, industrial 6, natural science 5, office 5, finance 7, math and formal reasoning 9, cybersecurity 5, media 5). Every reset generates a fresh instance (files, question, answer) from a seeded generator, so there are no fixed answer keys and this dataset holds none: tasks.jsonl lists each task id with a seed-0 example question, the answer format and the file names.

What the agent does

Each episode is a small workspace of text files and four plain tools, listed with one example call each in the first observation:

• list_files {}
• read_file {"path": ..., "offset": optional first line} (2,500 characters per call)
• run_python {"code": ...} (Python 3 standard library, task files in the working directory, runs as nobody)
• submit {"answer": ...} ends the episode

Answers are small JSON objects with named fields (for example {"best_project": "P41", "npv": 18234.55}).

Domain	Families (levels)
Software engineering	change-impact (1,2): rebuild set and deterministic build order with dev-only edges; version-resolve (1,2): semver ranges, plugin constraints, yanked releases; module-repair (3)
Industrial & physical	bom-explosion (1,2): BOM explosion with scrap and stock netting; tank-balance (1,2): level and alarm time with overflow and a tripped pump; line-oee (2,3)
Natural science	beer-lambert (1,2): least-squares calibration with blank subtraction, an excluded standard, dilution; stoichiometry (1,2): limiting reagent, purity, process yield; kinetics-lab (3)
Office & white-collar	meeting-slot (1,2): earliest common slot across time zones with a late calendar change; invoice-match (1,2): three-way match with tolerance and credit notes; expense-policy (3)
Finance & economics	npv-choice (1,2): NPV by risk class with errata; bond-desk (1,2): price and modified duration; fifo-basis (1,2): FIFO realized gain with fees and a split; loan-desk (3)
Math & formal reasoning	congruences (1,2): CRT with non-coprime moduli; logic-grid (1,2): unique-solution seating puzzles; shortest-route (1,2): closures and a waypoint; exact-probability (1,2): reduced fractions; exact-count (3)
Cybersecurity	firewall-audit (1,2): first-match rules with an approved change; vuln-triage (1,2): numeric semver ranges and vendor backports; auth-hunt (3)
Media & content	edit-timeline (1,2): EDL frames and timecode after review notes; headline-style (1,2): a two-page style guide; subtitle-qc (3)

What it is meant to teach

General desk habits rather than one tool's quirks: read every file (level-2 tasks put a correction, errata, change log or second page in a separate file), compute with Python instead of by eye, and submit exactly the requested JSON fields. Tool names are conventional and fixed; there is no ritual end action.

Reward

All or nothing: 1.0 only when every requested field is right (numeric tolerances are stated in the answer format; ids, lists, versions and timecodes match exactly, lists in any order unless an order is asked for), otherwise 0.0. No answer, an empty answer, running out of calls, or an invalid reply scores 0.0. The finish action {"tool": "submit", "args": {"answer": ""}} scores 0.0.

Same instance for a group

Rollouts that reset the same task within the same 5-minute window get the same instance (seeded from a secret in the image), so the parallel rollouts of an optimizer step are compared on one problem.

Checks

For every task over 100+ seeds: the reference answer scores 1 as an object and as a JSON string, plausible wrong answers (the answer you get by skipping the caveat file) and empty answers score 0, and generation is deterministic per seed. For all 19 new families an independent solver that reads only the question and files reproduces every label. openenv validate passes and every task replays to a terminal reward over WebSocket.

Source: https://github.com/nice-bills/shiftdesk