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 78 tasks of 14 arena submissions of skillsdesk. The arena runs a submission's image, pinned by digest, with the tasks and budgets of its request; this dataset documents them. The Runs section says which submission uses which tasks: skillsdesk-run1, skillsdesk-run2, skillsdesk-run3, skillsdesk-run4, skillsdesk-run5, skillsdesk-run5-r1, skillsdesk-run6, skillsdesk-run5-r2, skillsdesk-run5-r3, skillsdesk-run7, skillsdesk-run7-r1, skillsdesk-run7-r2, skillsdesk-run5-r4, skillsdesk-run8.

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: 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; 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. A generator family's task id, <family>-d<level>, draws a fresh instance of that level at every reset instead. 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", "code": "..."}	Ends the episode and scores it. Only a submit with the episode's code keeps the whole score. {"tool": "submit"} is the finish action.

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

The tools and limits above are those of skillsdesk-run8; the Runs table lists each run's env knobs.

Task families

Family	Arena domain	What the agent does	Output	Tasks
semver-fix	software engineering	Fix the 4 broken rules (of 8) in a semantic-versioning module without breaking the other 4. The visible unit tests catch only 2 of the bugs.	lib/semver.py	1
alarm-log	industrial and physical systems	Find the high alarms in 40 minutes of historian samples from 4 sensors: on-delay, deadband, ERR faults, alarms still open at the end.	alarms.json	4
lab-solutions	natural science	Compute molar masses from formulas (repeated elements, parentheses, hydrates), then the grams to weigh or the stock volume to dilute, across M/mM and mL/L.	prep.json	3
meeting-slots	office and white-collar work	Book 5 meetings in order: quarter-hour UTC starts, fixed UTC offsets, working hours, weekdays, busy blocks and the earlier bookings.	bookings.json	2
loan-schedule	finance and economics	Amortize 3-4 loans with cent rounding and lump-sum prepayments: payment, payments made, total interest, final payment, payoff month.	loan_summary.json	5
prop-logic	math and formal reasoning	Parse 8 propositional formulas (precedence, right-grouping ->), then build truth tables: variables, class, model count, first model.	truth.json	4
ssh-bruteforce	cybersecurity	Triage an sshd auth log: failures (with rsyslog repeat lines), sliding-window detection, a CIDR allowlist, logins after detection.	incident.json	2
subtitle-retime	media and content production	Clean sound annotations from an SRT file and retime it for a frame-rate change and a negative offset: drop or clamp at zero, cap, minimum gap, renumber.	subs/<name>.retimed.srt	3
config-migrate	software engineering	Migrate a service config to schema v2 in place: host and port, durations to integer milliseconds, binary sizes to bytes, log levels, feature flags, deprecated keys.	config/service.json	5
cable-sizing	industrial and physical systems	Size the cables of 6 circuits: design current, derating for ambient temperature and grouping, ampacity, then voltage drop against each circuit's limit.	cable_schedule.csv	4
primer-screen	natural science	Screen 12 PCR primers (GC content, Tm, base runs, 3' clamp) and pick the best forward/reverse pair of each target gene by Tm difference.	primer_report.json	4
expense-audit	office and white-collar work	Audit expense claims: daily FX rates with a weekend fallback, rejection reasons in order, meal and hotel caps, reimbursement per employee.	audit.json	3
bank-recon	finance and economics	Reconcile a bank statement with the cash book: cheques by number, other entries by amount and a date window, items in transit, book errors, the adjusted balance.	reconciliation.json	3
congruences	math and formal reasoning	Solve 7-8 systems of linear congruences with non-coprime moduli (generalised CRT): smallest solution, modulus, solutions in a range, or none.	solutions.json	3
firewall-audit	cybersecurity	Evaluate flows against a first-match firewall (CIDR masks, port ranges, any), then find the shadowed and the redundant rules.	firewall_audit.json	5
loudness-gain	media and content production	Compute loudness-normalization gains for a streaming delivery: silence, target, true-peak ceiling, gain cap, and one duration-weighted gain per album.	gains.json	4
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
rundown-timing	media and content production	Time a broadcast rundown with durations in mixed notations: each item's in-time, the total, the out-time and the signed over- or under-run, as timecodes.	timing.json	1
caption-qc	media and content production	Check every cue of an SRT file against a house style (lines, characters per line, reading speed, duration, gap in frames) and list the rules each breaks.	qc.json	1
delivery-qc	media and content production	Check probed media files against a delivery spec (frame rate as a fraction, size, audio, duration, aspect, naming): PASS or FAIL with reason codes.	delivery.json	1
chapter-markers	media and content production	Turn a chapter cue sheet in mixed time notations into ordered chapters with starts and ends, dropping duplicates and chapters that are too short.	chapters.json	2
molar-prep	natural science	Molar masses from formulas (groups, hydrates) and the grams to weigh or the stock millilitres to dilute, across M/mM and mL/L.	prep.json	2
crt-solve	math and formal reasoning	Solve systems of linear congruences with the generalised Chinese remainder procedure: smallest solution, modulus, solutions in a range, or null.	crt.json	3

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.
• Self-submit: each reset shows a new code, and only a submit that carries it keeps the whole reward. Any other ending (the finish action after a reply that is not one JSON object, or at the end of a budget) gets 25% of it, so breaking the reply format costs reward.
• 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 task per line: task_id (<family>-<seed> or <family>-d<level>), split and its wall-time limits, exactly as in the tasks file of every run that submits it, so a task in several runs has one line. A run sets its own token and tool-call budgets on its tasks, and can set shorter wall-time limits over these lines: the Runs section lists them.

78 tasks (78 train). Held-out seeds (split eval) exist for memorization checks; no run here submits them.

Runs

The arena submissions this dataset documents, oldest first. A submission's tasks are their lines above with its budgets set on each; the arena's defaults fill every other limit.

Submission	Image	Tasks	Budgets	Env knobs
skillsdesk-run1	ghcr.io/shaal/skillsdesk:run1	24	completion_tokens 12288, context_tokens 16384, tool_calls_total 30	defaults
skillsdesk-run2	ghcr.io/shaal/skillsdesk:run2	50	completion_tokens 12288, context_tokens 16384, tool_calls_total 30	defaults
skillsdesk-run3	ghcr.io/shaal/skillsdesk:run3	13	completion_tokens 28672, context_tokens 32768, tool_calls_total 48	families = ["bundle"]; reset_family = "bundle"
skillsdesk-run4	ghcr.io/shaal/skillsdesk:run4	10	completion_tokens 11264, context_tokens 12800, tool_calls_total 18	defaults
skillsdesk-run5	ghcr.io/shaal/skillsdesk:run5	9	completion_tokens 11264, context_tokens 12800, tool_calls_total 21	self_submit = true
skillsdesk-run5-r1	ghcr.io/shaal/skillsdesk@sha256:30ff331ba2e1f0b4b5e8e800d6afcae726fded65b506ffa17b63ac9ffb9be203	9	completion_tokens 11264, context_tokens 12800, tool_calls_total 21	self_submit = true
skillsdesk-run6	ghcr.io/shaal/skillsdesk:run6	7	completion_tokens 8704, context_tokens 9728, tool_calls_total 14	defaults
skillsdesk-run5-r2	ghcr.io/shaal/skillsdesk@sha256:30ff331ba2e1f0b4b5e8e800d6afcae726fded65b506ffa17b63ac9ffb9be203	9	completion_tokens 11264, context_tokens 12800, tool_calls_total 21	self_submit = true
skillsdesk-run5-r3	ghcr.io/shaal/skillsdesk@sha256:30ff331ba2e1f0b4b5e8e800d6afcae726fded65b506ffa17b63ac9ffb9be203	9	completion_tokens 11264, context_tokens 12800, tool_calls_total 21	self_submit = true
skillsdesk-run7	ghcr.io/shaal/skillsdesk:run7	7	completion_tokens 5632, context_tokens 7168, tool_calls_total 12	families = ["rundown-timing", "caption-qc", "delivery-qc", "chapter-markers", "molar-prep", "crt-solve"]
skillsdesk-run7-r1	ghcr.io/shaal/skillsdesk@sha256:152030ffc3be45ede71600f18ea0319b13c301cbe273f8faf77ee20b8d378fec	7	completion_tokens 5632, context_tokens 7168, tool_calls_total 12	families = ["rundown-timing", "caption-qc", "delivery-qc", "chapter-markers", "molar-prep", "crt-solve"]
skillsdesk-run7-r2	ghcr.io/shaal/skillsdesk@sha256:152030ffc3be45ede71600f18ea0319b13c301cbe273f8faf77ee20b8d378fec	7	completion_tokens 5632, context_tokens 7168, tool_calls_total 12	families = ["rundown-timing", "caption-qc", "delivery-qc", "chapter-markers", "molar-prep", "crt-solve"]
skillsdesk-run5-r4	ghcr.io/shaal/skillsdesk@sha256:30ff331ba2e1f0b4b5e8e800d6afcae726fded65b506ffa17b63ac9ffb9be203	9	completion_tokens 11264, context_tokens 12800, tool_calls_total 21	self_submit = true
skillsdesk-run8	ghcr.io/shaal/skillsdesk:run8	26	completion_tokens 12288, context_tokens 13824, tool_calls_total 21/27, reset_wall_s 120, rollout_wall_s 600, verifier_wall_s 90	self_submit = true

skillsdesk-run1

Submitted as "skillsdesk: SkillsBench-style file tasks in eight domains": 24 tasks, in this order: alarm-log-1, lab-solutions-1, loan-schedule-1, prop-logic-1, ssh-bruteforce-1, subtitle-retime-1, alarm-log-2, meeting-slots-2, loan-schedule-2, subtitle-retime-2, alarm-log-3, lab-solutions-3, meeting-slots-3, loan-schedule-3, prop-logic-3, alarm-log-4, lab-solutions-4, loan-schedule-4, prop-logic-4, semver-fix-5, loan-schedule-5, prop-logic-5, ssh-bruteforce-5, subtitle-retime-5.

skillsdesk-run2

Submitted as "skillsdesk: SkillsBench-style file tasks in eight domains": 50 tasks, in this order: prop-logic-5, cable-sizing-1, config-migrate-5, expense-audit-2, loudness-gain-2, loudness-gain-1, loudness-gain-3, bank-recon-1, prop-logic-3, config-migrate-1, primer-screen-4, prop-logic-4, subtitle-retime-2, config-migrate-3, firewall-audit-4, primer-screen-3, bank-recon-5, congruences-5, firewall-audit-2, primer-screen-2, prop-logic-1, bank-recon-4, cable-sizing-3, cable-sizing-2, cable-sizing-4, meeting-slots-2, primer-screen-5, subtitle-retime-5, loan-schedule-2, loudness-gain-4, subtitle-retime-1, expense-audit-5, loan-schedule-3, semver-fix-5, meeting-slots-3, config-migrate-2, config-migrate-4, firewall-audit-3, expense-audit-1, congruences-4, firewall-audit-1, lab-solutions-1, loan-schedule-4, alarm-log-2, loan-schedule-1, congruences-2, firewall-audit-5, alarm-log-1, lab-solutions-3, ssh-bruteforce-1.

skillsdesk-run3

Submitted as "skillsdesk: three SkillsBench-style file tasks per episode, in mixed domains": 13 tasks, in this order: bundle-9, bundle-15, bundle-13, bundle-3, bundle-16, bundle-12, bundle-11, bundle-8, bundle-5, bundle-1, bundle-14, bundle-4, bundle-2.

skillsdesk-run4

Submitted as "skillsdesk: ten SkillsBench-style file tasks the model mostly solves": 10 tasks, in this order: primer-screen-3, congruences-5, cable-sizing-3, loan-schedule-2, loudness-gain-4, expense-audit-5, meeting-slots-3, firewall-audit-1, lab-solutions-1, alarm-log-2.

skillsdesk-run5

Submitted as "skillsdesk: nine SkillsBench-style file tasks; only the model's own submit scores in full": 9 tasks, in this order: loan-schedule-3, primer-screen-3, loudness-gain-2, bank-recon-1, expense-audit-5, prop-logic-5, cable-sizing-4, congruences-2, alarm-log-2.

skillsdesk-run5-r1

Submitted as "skillsdesk: nine SkillsBench-style file tasks; only the model's own submit scores in full (redraw run5-r1)": 9 tasks, in this order: loan-schedule-3, primer-screen-3, loudness-gain-2, bank-recon-1, expense-audit-5, prop-logic-5, cable-sizing-4, congruences-2, alarm-log-2.

skillsdesk-run6

Submitted as "skillsdesk: seven SkillsBench-style science, math and media file tasks": 7 tasks, in this order: prop-logic-5, loudness-gain-2, loudness-gain-1, primer-screen-4, subtitle-retime-2, congruences-5, lab-solutions-1.

skillsdesk-run5-r2

Submitted as "skillsdesk: nine SkillsBench-style file tasks; only the model's own submit scores in full (redraw run5-r2)": 9 tasks, in this order: loan-schedule-3, primer-screen-3, loudness-gain-2, bank-recon-1, expense-audit-5, prop-logic-5, cable-sizing-4, congruences-2, alarm-log-2.

skillsdesk-run5-r3

Submitted as "skillsdesk: nine SkillsBench-style file tasks; only the model's own submit scores in full (redraw run5-r3)": 9 tasks, in this order: loan-schedule-3, primer-screen-3, loudness-gain-2, bank-recon-1, expense-audit-5, prop-logic-5, cable-sizing-4, congruences-2, alarm-log-2.

skillsdesk-run7

Submitted as "skillsdesk: short media, science and math file tasks, a fresh instance per episode": 7 tasks, in this order: crt-solve-d3, delivery-qc-d2, caption-qc-d3, rundown-timing-d1, chapter-markers-d3, chapter-markers-d2, molar-prep-d2.

skillsdesk-run7-r1

Submitted as "skillsdesk: short media, science and math file tasks, a fresh instance per episode (redraw run7-r1)": 7 tasks, in this order: crt-solve-d3, delivery-qc-d2, caption-qc-d3, rundown-timing-d1, chapter-markers-d3, chapter-markers-d2, molar-prep-d2.

skillsdesk-run7-r2

Submitted as "skillsdesk: short media, science and math file tasks, a fresh instance per episode (redraw run7-r2)": 7 tasks, in this order: crt-solve-d3, delivery-qc-d2, caption-qc-d3, rundown-timing-d1, chapter-markers-d3, chapter-markers-d2, molar-prep-d2.

skillsdesk-run5-r4

Submitted as "skillsdesk: nine SkillsBench-style file tasks; only the model's own submit scores in full (redraw run5-r4)": 9 tasks, in this order: loan-schedule-3, primer-screen-3, loudness-gain-2, bank-recon-1, expense-audit-5, prop-logic-5, cable-sizing-4, congruences-2, alarm-log-2.

skillsdesk-run8

Submitted as "skillsdesk: 26 file tasks in all eight domains; only the model's own submit scores in full": 26 tasks, in this order: bank-recon-5, expense-audit-1, bank-recon-4, crt-solve-d3, loan-schedule-2, loan-schedule-3, crt-solve-d1, crt-solve-d2, molar-prep-d1, molar-prep-d2, subtitle-retime-2, expense-audit-2, primer-screen-3, loudness-gain-2, loudness-gain-3, bank-recon-1, firewall-audit-2, lab-solutions-4, prop-logic-5, congruences-4, cable-sizing-4, config-migrate-2, alarm-log-2, config-migrate-1, meeting-slots-2, ssh-bruteforce-5.

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.