Syndicate604/skills-desk

Skills Desk

An OpenEnv environment for the OpenEnv Arena that trains reliable problem solving with tools: understand the request, find the relevant information, compute or repair something, produce the required output, and check it. Each episode is a small workspace under /app with a short instruction. The agent runs commands (exec), writes files (write_file) and sends evaluate; a hidden grader checks the result and returns 1 or 0.

Image: ghcr.io/syndicate604/skills-desk (linux/amd64). Source: source/. Forked from Science Desk and dnakhla/terminal-desk v2 (Apache-2.0); see NOTICE.

Interface

Terminal Desk v2's copy of OpenEnv's upstream tbench2_env contract: exec, write_file, evaluate, close; the instruction repeated in every observation; output not cut; per-task budgets of 4096 completion tokens (model tokens plus every observation after the first), 8192 context tokens and 32 tool calls. The image has Python 3.12 with numpy, scipy, pandas and astropy, plus jq, bc and sqlite3.

Training tasks (v2: 36)

manifest/training_tasks_v2.txt. v2 keeps only tasks that give the trainer mixed outcomes on first visit, the selection rule behind vibrantlabsai's usaspending-recon and pool-desk runs. The rule (manifest/v2-PREREGISTRATION.md) was frozen before screening: 52 candidates (v1's 40 tasks and 12 contract-reconciliation tasks), 6 episodes each of untrained Qwen3.8-27B under the arena's budgets; a task is kept when it was solved 1 to 5 times of 6. 36 were kept; per-candidate counts are in manifest/v2-selection.json. Each id is one fixed problem, so a GRPO group compares answers to the same question; ids shared with v1 serve exactly their v1 instances.

Skill	Tasks
Find and combine evidence	backorders (2 variants), calibration windows, duplicate customers, bank reconciliation, first release that fixed an issue
Contract reconciliation (real data)	contract obligations against a finance ledger, levels 1, 2 and 4 (5 tasks)
Compute from data	channel statistics, rainfall with gauge scales and QC, loan schedule
Transform structured data	contacts CSV to typed JSON, log lines to JSON, XML to CSV, long to wide, Markdown table to JSON, JSONL sessions, regional dates to ISO
Repair and verify	toolkit module, broken report script, pagination, log-line parser, unit conversion, business days, deduplication script, sorted log merge
Follow constraints	roster (2 sizes), correct a one-error draft roster, complete a half-agreed roster, cargo choice, fewest disks, delivery days

Left out because the screen solved them every time: deliveries, badge access, rainfall L1, meter registers, final grades, FIFO stock cost, rolling-median exceedances, fixed-width report. Left out because it never solved them: course selection, contract levels 3, 5 and 6 and one level-2 contract (mostly the 4096-token budget, which counts query output). The v1 bank (manifest/training_tasks.txt, 40 tasks) is still served.

The contract tasks serve the CC0 task databases of vibrantlabsai/usaspending-recon (FY2024 U.S. EPA contract actions from USAspending.gov) unchanged, with Skills Desk's own instruction, output file and grader; see NOTICE.

Reserved evaluation tasks (20)

manifest/reserved_tasks.txt: four per skill, two fresh applications of a familiar skill and two combining skills, with different file layouts and solution structures. They are served by the same image so a trained checkpoint can be evaluated on them, but they are never trained on and never used to choose or tune the training tasks. The set is frozen (v1); a change would be published as a new version.

tasks.jsonl lists the 36 training and 20 reserved ids with their split, skill, output path and fixed instruction.

Reward

Binary on evaluate. Content and structure are graded together: correct JSON with wrong values fails, correct values in a malformed file fail, and a CSV row with extra or missing fields fails. Harmless differences pass (key order, CSV quoting, line endings, and either rounding of an exact half where the task states no rounding rule; a task that states half-up rounding requires it). Tasks with many correct answers (rosters, schedules, plans) are graded against their rules. Repair tasks run the repaired code as nobody on hidden input. Nothing rewards a particular tool sequence, reply length or claims of checking.

Verification

Every task: the reference output scores 1, no output and corrupted output score 0, an independent agent-style reference script scores 1 on four instances inside the image, and a deliberate common mistake scores 0. Instances are byte-identical in any process and frozen by fingerprints in source/tests/fixtures/. See source/README.md to build and test.

Limitations

The selection screen is small (6 episodes per task) and uses the untrained model through a provider, not the arena's sampler; earlier screens underestimated the arena-trained model, and no correction was applied. Whether training on these tasks transfers can only be measured by evaluating a trained checkpoint on the reserved tasks; the arena's private score is a separate measure.