naklitechie/openenv-terminal-habits

OpenEnv Arena terminal habits (v2)

Shell tasks in the same action and observation format as OpenEnv's tbench2_env, which the arena's own terminal smoke tests use. Each task gives input files in /app and asks for one exact answer in /app/answer.txt. The files are too large to print within the arena's budget, so a solve needs the habits the private eval rewards on terminal tasks:

• reply with one well-formed JSON action per turn;
• read /app/NOTES.md, then inspect the data with commands instead of printing it;
• notice what the notes do not list: summary rows, rows exported twice, stray whitespace, a byte-order mark;
• compute with a script (python3, awk, jq are installed) that applies every rule;
• write one exact output file, then send evaluate.

Why v2

v1 (image :v1, dataset revision 9d75217) asked one-rule questions. The arena trained on it in run d05536ba and the model solved 98% of episodes from the first step, so 250 of 270 optimizer steps had no reward variance and no gradient. The run scored 2/40, inside the range of runs that never update. v2 keeps the shape and makes every task stack rules.

Tasks

8 families, one per arena domain, 2 levels each (16 task ids, tasks.jsonl):

family	the question	rules and artefacts
office	category spend in USD across three monthly CSVs	column order differs by file, quoted thousands, category case, resubmitted ids, approved only, EUR conversion; level 2: monthly rates, date window, refunds. Artefacts: a SUBTOTAL row per file, spaces after status, a BOM
finance	net balance of one account for Q3	two date formats, reversals, a merged account, fees; level 2: currencies, quoted thousands, pending rows. Artefacts: rows exported twice, a TOTAL row
industrial	readings above a temperature limit	rotated logs incl. gzip, Fahrenheit readings, calibration runs, maintenance windows; level 2: lines written to two files, a relabelled machine. Artefacts: repeated lines, rotation markers, CRLF
science	blank-corrected concentration of one sample	three units, QC-flagged replicates, per-plate blank subtraction; level 2: detection limit, mean of plate means. Artefacts: repeated rows, per-sample summary rows, a BOM
security	IPs with repeated failed logins	two log forms, other services' lines, a time window, CIDR allowlist; level 2: IPv4-mapped IPv6, a 10-minute burst rule. Artefacts: repeated lines, a BOM
math	sum of distinct primes / twin-prime pairs	five integer notations, several values per line, comments, [a..b] runs. Artefacts: header and footer lines
media	on-screen seconds of cues with a word	two-line and italic cues, whole-word match, overlap counted once; level 2: sound cues, a time window. Artefacts: blocks written twice, a translator's NOTE, a BOM
software	p95 latency of one service	rotated logs incl. gzip, truncated lines, three latency formats, /healthz, nearest rank; level 2: retries by request_id, a time window. Artefacts: lines shipped twice, a BOM

Every reset without a seed builds a fresh random instance, so the rollouts of a GRPO group see different data. A given (task_id, seed) is deterministic.

Reward

1.0 when /app/answer.txt holds the expected answer (numbers within half a unit of the stated rounding, lists exactly in the stated order), otherwise 0.0. The answer is computed when the instance is built and kept in the server's memory; it is never written to disk.

Interface

• Action: {"action_type": "exec", "command": "..."}, {"action_type": "write_file", "file_path": "answer.txt", "content": "..."}, {"action_type": "evaluate"}. Other tbench2_env fields are accepted; other action types return an error observation.
• Observation: instruction (repeated every turn, as tbench2_env does), output (full command output, not cut), success, error, task_id.
• Commands run as an unprivileged user in /app with a 20-second limit; 32 actions at most, after which the episode is evaluated.

Checks before release

• Unit tests over every task id and 6 seeds: the reference solution scores 1; missing, empty and wrong answers score 0; the same seed gives the same instance; the main input exceeds 1,500 tokens; NOTES.md exists and seeds differ. A wider pass: the reference matched the generator on 25 more seeds per task.
• Stand-in calibration (source/calibrate.py, DeepSeek-chat at temperature 1 under the arena's budgets, replies read leniently so format slips do not count): 100% on v1, against the arena model's 98%; 51% on v2, with per-task rates from 25% to 100%.
• Container replay over WebSocket and openenv validate --url: source/evidence/.

Image and source

• Image: ghcr.io/naklitechie/openenv-terminal-habits:v2 (linux/amd64).
• Source: source/ in this dataset, pinned to OpenEnv 86a180ede21e044f7929b9a7783ad83aa67d83a3.