Syndicate604/skills-desk-v3

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 (v3: 48)

manifest/training_tasks_v3.txt: the 40 v1 tasks, unchanged and in v1 order, plus 8 output-format tasks, one inserted after every fourth v1 task. Each id is one fixed problem, so a GRPO group compares answers to the same question; v1 ids serve exactly their v1 instances.

Output format (new in v3). The facts are already correct and supplied in the workspace; the agent only has to write them in exactly the requested form: key names and JSON types, list order, half-up rounding and digit grouping, CSV quoting, nulls for missing values, and late change requests in a spec file. The reward stays binary on the whole output.

Family	Task	Levels trained
fmt_keys	checked order facts to an exactly typed JSON object (L2: spec with a change request; L3: nested)	1, 2, 3
fmt_csv	JSON records to a CSV report (L3: spec with a filter and a change request)	1, 3
fmt_numbers	exact decimal strings to display text with half-up rounding (L3: German number style)	1, 3
fmt_schema	a free-form support ticket to schema-exact JSON (mixed date and duration spellings)	2

Selection followed a rule frozen before screening (manifest/v3-PREREGISTRATION.md): 12 candidates (four families, levels 1-3), 6 episodes each of untrained Qwen3.8-27B under the arena's budgets; a candidate is kept unless it was solved all 6 times. Counts are in manifest/v3-selection.json. Most screen failures were the reply format itself (prose or a fenced block instead of one JSON action, an unknown action type), not wrong content.

v1 skill	Tasks
Find and combine evidence	backorders (2 variants), calibration windows, deliveries, badge access, duplicate customers, bank reconciliation, first release that fixed an issue
Compute from data	channel statistics, rainfall (2 variants), meter registers, final grades, FIFO cost, rolling-median exceedances, loan schedule
Transform structured data	contacts CSV to JSON, log lines to JSON, XML to CSV, long to wide, Markdown table to JSON, JSONL sessions, fixed-width report, 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), cargo choice, draft roster fix, roster completion, course selection, fewest disks, delivery days

The image also serves the v2 contract-reconciliation tasks (CC0 USAspending data, see NOTICE), which v3 does not train on.

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 48 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.