skillsdesk

skillsdesk is an OpenEnv environment for OpenEnv Arena. Every task is a piece of desk work for a terminal agent: a workspace with INSTRUCTIONS.md, a skill document (skills/<skill>/SKILL.md) that carries the knowledge the task needs (a formula, a format specification, a procedure), and input files (CSV, JSON, logs, code). The agent inspects the files with a shell, writes one output file at an exact path, and submits.

A bundle task puts three such tasks, of three families and three domains, in one episode: one folder each (part1/ to part3/) with its own INSTRUCTIONS.md, skill and inputs, and its own output (a new file, or an input edited in place). A top-level INSTRUCTIONS.md lists the parts; one submit scores all of them. A competent agent needs about 20 tool calls.

This dataset lists the 13 tasks of the arena submission skillsdesk-run3 (image ghcr.io/shaal/skillsdesk:run3). The arena runs only the image, pinned by digest; this dataset documents it.

What it teaches

• Reading a task and its skill document, then applying the rules exactly: units, rounding, time zones, edge cases the skill spells out. Each generator plants the cases its skill exists for, so a realistic misreading of a rule costs reward.
• Working with tools one call at a time. Tool output is cut to 2,000 characters, so the agent has to look at large inputs with commands or small scripts instead of reading them whole.
• Producing machine-checked output: JSON numbers, exact timestamps, files at exact paths.

Domains: each bundle task combines three of the eight arena domains (software engineering, industrial and physical systems, natural science, office and white-collar work, finance and economics, math and formal reasoning, cybersecurity, media and content production).

Episode

• reset(task_id="<family>-<seed>") builds a fresh /workspace. The task id alone fixes its content, so every episode of a task sees the same files. The first observation holds the instruction and a short file tree.

• Each turn the model sends exactly one JSON object, one tool call:

  Action	What it does
  {"tool": "bash", "command": "..."}	Runs it with bash in /workspace, at most 60 s.
  {"tool": "read_file", "path": "..."}	Shows a file inside the workspace.
  {"tool": "write_file", "path": "...", "content": "..."}	Replaces a file.
  {"tool": "submit"}	Ends the episode and scores it. It is also the finish action.

• Observations are JSON and short: tool output, an exit code, an error. The instruction is never repeated.

Task families

Family	Arena domain	What the agent does	Output	Tasks
bundle	mixed (three per task)	Three tasks of three other skillsdesk families (16 in all, two per domain), from three domains, in one episode: one folder each (part1/ to part3/) with its own instructions, skill and inputs, and one submit for all three.	partN/<its output>	13

Reward

• On submit (or the finish action), the task's verifier runs 4 to 6 independent checks, each about a different part of the answer. The reward is the fraction of checks passed, a float in [0, 1]. Every other step returns 0.0.
• A bundle task's reward is the mean of its three parts' rewards, each part scored by its own family's 4 to 6 checks. A part with no output scores 0.
• Partial work earns partial credit. A well-formed but wrong answer, a copy of the input, or an untouched workspace scores 0.0. Tolerances are at most about one unit in the last reported decimal, so writing several guesses cannot pass.
• The verifier recomputes the expected answer from the seed at scoring time; it never exists in the workspace. Run the documented way (the image's own user, root), the server starts the model's tools as an unprivileged user that cannot read the verifiers, the generators or the server.

Tasks file

tasks.jsonl has one arena task per line: task_id (<family>-<seed>), split and the task's budgets, exactly as submitted. 13 tasks (13 train). Held-out seeds (split eval) exist for memorization checks; they are not part of this submission.

Limit per task	Value
reset_wall_s	180	the arena's default
rollout_wall_s	1800	the arena's default
verifier_wall_s	120	the arena's default
tool_wall_s	120	the arena's default
tool_calls_total	48	set by the submission
tool_calls_per_minute	60	the arena's default
completion_tokens	28672	set by the submission
context_tokens	32768	set by the submission
memory_gib	2	the arena's default
cpu_floor_vcpus	1	the arena's default
workspace_gib	10	the arena's default

Origin

All tasks are original. Each family is a procedural generator written for this environment: the seed draws the data, the planted edge cases and the expected answer. No task is copied from SkillsBench or any other benchmark, and none is derived from the arena's private evaluation tasks.