dnakhla/terminal-desk

Terminal Desk

An OpenEnv environment for the OpenEnv Arena. Each episode is a fresh workspace (under /root for most families, /app for a few) with a task instruction in the prose shape of a SkillsBench task. The agent runs commands, writes one output file, and sends evaluate; a hidden set of named checks reads the file and returns 1 if every check passes, else 0.

Image: ghcr.io/dnakhla/terminal-desk (linux/amd64, public). Source: source/ in this repo. Earlier versions: v1 (JSON-answer interface, 3/40), v2 (the interface below with 16 short families, 3/40 including the first natural-science solve on the arena board), v3c (the families below without the gate, 2/40).

The interface: the same as the private evaluation

The arena operator's own test Spaces (benchflow/arena-tb2-*, dataset benchflow/openenv-smoke-fix-git) serve Terminal-Bench 2 tasks through OpenEnv's upstream envs/tbench2_env wrapper at the arena's pinned revision. Terminal Desk reproduces that contract so that what the policy learns here is what it meets on the private tasks:

• Action and observation schemas equal the operator's published schema (tests/fixtures/tb2_schema.json; a test diffs the live /schema against it). exec runs a command, write_file writes a file, evaluate ends the episode with a binary reward, close ends it without one; session actions fail softly with success: false.
• The instruction is repeated in every observation and command output is not cut (a 60,000-character safety cap only), so the arena's token budget, not the environment, is what punishes dumping files.
• Per-task limits are copied verbatim from the operator's smoke submission: completion_tokens 4096 (which the arena counts across the model's tokens and every observation after the first), context_tokens 8192, tool_calls_total 32, tool_wall_s 110.

v3: SkillsBench-shaped tasks screened for mixed difficulty

v2 trained (reward 0.75 to 0.84 over the run) but 70% of its sampling groups had zero reward variance by the end: the tasks were too short and too easy, and three of its science tasks sat at reward 1.00 from the first step. v3 keeps the interface and rebuilds the tasks:

• 24 procedural families, three per arena domain, three levels each (72 task ids, <family>-L<n>). Every reset draws a fresh instance; client seeds are ignored.
• Instructions are 600 to 1,000 characters in SkillsBench prose: a scenario, the inputs with absolute paths, the rule with its edge cases, the exact output path and format with one example row, and rounding. No hints about tools, budgets or brevity.
• Every family has one input whose first ten lines cost under 300 tokens and whose full dump costs over 1,500 (a few hundred rows under 100 characters), so a schema peek is cheap and cat costs the episode.
• The reference solution of every family is one script under 600 tokens; levels add rows, one more rule, or a second input file rather than more code.
• Checks are named (test_file_exists, test_row_count, test_balances, ...) and reported pytest-style in the evaluate output; info carries assertions_passed and assertions_total. Reward is 1 only if every check passes. Floats are compared with the tolerance the instruction states.
• 22 of the 24 families need only the standard library. off3_pdf_xlsx_diff reads pdf and xlsx with pdfplumber and openpyxl, and med3_sprite_count reads png with Pillow and numpy; the image installs that library set either way, so the policy meets it.
• The submissions train only on ids the untrained model solved between 1 and 3 times of 4 in the screen below, and only in office, software and finance (v3c: all 20 such ids; v4: 17 of them, see below). The leaderboard keeps each account's best per domain, and across 74 arena runs the private tasks that were ever solved cluster in a few domains: office 2 happens in 46% of runs and office 3 in 3%, a software solve in 22%, a finance solve in 11%, a media solve in 4%, and math, security and science solves once each. This account's industrial 2 and science 1 are already the most any run has scored, so training there cannot add a point, and math, security and media have no private task within reach of a modest lift.

Domain	Families and output file
software engineering	sw3_fix_module repair a helper module against hidden tests, /app/src/toolkit.py; sw3_rate_limit replay token-bucket rate limiting, /root/rejected.json; sw3_dep_lock resolve a lock file against a package index, /root/lock.txt
industrial and physical systems	ind3_service_due machines due for service, /root/service_report.csv; ind3_tank_balance a flow log into /root/tank_report.json; ind3_stl_mass solid volumes and masses from an STL, /root/mass_report.json
natural science	sci3_kinetics_fit rate constant, order and activation energy, /root/kinetics_report.json; sci3_stoich_batch limiting reagent and yield from a stockroom table, /root/batch_report.json; sci3_lightcurve_period transit period from a light curve, /root/period.txt
office and white-collar work	off3_ar_recon receivables with memos and credit notes, /root/balances.csv; off3_meeting_slot earliest slot across time zones, /root/meeting.json; off3_pdf_xlsx_diff a PDF roster against a workbook, /root/diff_report.json
finance and economics	fin3_loan_schedule amortize a loan from the loan book, /root/schedule.csv; fin3_portfolio_value value a brokerage account at the last close, /root/positions.csv; fin3_invoice_fraud screen invoices for fraud, /root/fraud_report.json
math and formal reasoning	math3_route_plan fastest route with closures, /app/route.json; math3_count_items raffle counts under divisor rules, /root/results.json; math3_exam_slots an exam schedule that minimizes conflicts, /root/output/schedule.csv
cybersecurity	sec3_auth_triage brute-forced sources from auth logs, /root/blocked.txt; sec3_fw_verdicts first-match firewall verdicts, /root/verdicts.csv; sec3_exfil_flows exfiltration alerts from flow logs, /root/alerts.json
media and content production	med3_silence_report silences in a WAV, /app/compression_report.json; med3_subtitle_retime subtitles after cuts, /root/output/episode.srt; med3_sprite_count sprites per frame from PNGs, /root/counting_results.csv

tasks.jsonl has one line per task id: family, domain, level, output file, an example instruction, whether the task is in the arena submission, and the screen statistics below.

Isolation and anti-gaming

• The server runs as root outside the workspace; exec and write_file run as nobody with CPU, memory, file size and process limits, and every nobody process is killed after each action and before verification.
• Expected outputs exist only in server memory. The checks only parse the output file (symlinks, FIFOs and oversized files are refused; xlsx and docx are parsed in a nobody subprocess); sw3_fix_module runs its hidden tests on a copy of the module as nobody.
• Every reset wipes /root and /app and draws a fresh instance, so a second session opened from the agent's shell learns nothing.

Screen against the untrained model

Qwen/Qwen3.8-27B, thinking off, temperature 1.0 (OpenRouter), 4 episodes per task id against the image under the arena's budgets emulated from token usage (the episode ends when the budget runs out, then evaluate runs). The action parse is lenient, as the arena's is: the first JSON object with an action_type anywhere in the reply, with a 2,400-token reply cap. 288 episodes, 72 ids, mean reward 0.41. Endings: budget or context exhausted 124, a reply with no action object 61, evaluate or finish 59, a reply cut at the cap 26, an action outside the schema 17. Mean 2.9 actions per episode; write_file was used once. The budget deaths are what the task shape is for: the model dumps an input and the episode ends.

Solves per task id (v3c trained on the 20 office, software and finance ids solved 1 to 3 times; v4 trains on 17 of them, flagged in_submission in tasks.jsonl; "out" = never or always solved):

Family	L1	L2	L3
sw3_fix_module	2/4	0/4 (out)	0/4 (out)
sw3_rate_limit	3/4	2/4	1/4
sw3_dep_lock	1/4	0/4 (out)	0/4 (out)
ind3_service_due	3/4	2/4	0/4 (out)
ind3_tank_balance	4/4 (out)	2/4	1/4
ind3_stl_mass	3/4	1/4	1/4
sci3_kinetics_fit	2/4	3/4	1/4
sci3_stoich_batch	4/4 (out)	1/4	0/4 (out)
sci3_lightcurve_period	0/4 (out)	0/4 (out)	0/4 (out)
off3_ar_recon	3/4	3/4	2/4
off3_meeting_slot	1/4	1/4	1/4
off3_pdf_xlsx_diff	2/4	1/4	0/4 (out)
fin3_loan_schedule	2/4	0/4 (out)	1/4
fin3_portfolio_value	3/4	2/4	1/4
fin3_invoice_fraud	4/4 (out)	3/4	3/4
math3_route_plan	2/4	3/4	1/4
math3_count_items	3/4	2/4	3/4
math3_exam_slots	1/4	1/4	1/4
sec3_auth_triage	2/4	3/4	1/4
sec3_fw_verdicts	1/4	1/3	1/4
sec3_exfil_flows	3/4	3/4	1/4
med3_silence_report	2/4	2/4	2/4
med3_subtitle_retime	3/4	0/4 (out)	0/4 (out)
med3_sprite_count	3/4	3/4	0/4 (out)

The 35 ids at 1-3 of 4 in the other five domains are held back from this submission (see above). The screen is about 0.2 harsher than the arena's own untrained reward on v2's tasks, so 3/4 here is close to saturated there. Six ids were re-screened after their generators were eased (sci3_stoich_batch L1 and L2, sci3_kinetics_fit-L1, med3_subtitle_retime all levels); sec3_fw_verdicts-L2 has one episode fewer because the grader raised on a ragged CSV row, fixed before the image was built. sci3_lightcurve_period lost every episode and is left out whole.

Raw episodes: screen_v3.jsonl; selected ids: selected_tasks.txt.

terminal-desk-v3c: pre-registration and result

Pre-registered before the POST: same interface and budgets as v2, with the 20 office, software and finance tasks above; judged on new points for the account (industrial 2, office 1, science 1 banked), not on the total; expected total 2 to 5 of 40; new points = office 2 or more, any software solve, any finance solve (null rates 46%, 22%, 11%); evidence over luck = two of the three in one run (about 15% under the null) or office 3 (3%).

Result: 2/40 (industrial 1, office 1), run aee36bddad8c082ce3ae3546, 125 updates. No new point, so the run is null by its own rule. Inside the run the reward rose from 0.46 to 0.65 by 20-update block, 30 to 50% of groups kept gradient and reply length stayed near 530 tokens. The arena's own passes label the tasks better than any local screen: 7 of the 20 were saturated from the first pass (sw3_rate_limit L1-L3, fin3_loan_schedule-L1, fin3_portfolio_value-L1, fin3_invoice_fraud L2-L3), 3 scored 0 on every pass (off3_ar_recon-L3, off3_pdf_xlsx_diff-L2, sw3_fix_module-L1: the model dumps a 6,000-character input and the budget ends the episode), and off3_ar_recon-L2 went from 0 to 1.0 by the sixth pass. Curve: https://openenvarena-training.hf.space/?project=dnakhla&runs=aee36bddad8c082ce3ae3546

v4: the decisiveness gate (terminal-desk-v4)

Three runs on this account (142, 162 and 125 updates) scored 3, 3 and 2, inside the band of every trained run on the board, while runs that trained with zero gradient score 0 to 2 (naklitechie's null controls). Training does reach the private score, but nothing domain-matched has shown for any entrant, so what transfers looks generic: fitting the action format and the token budget. v4 trains that directly.

The :v4 image is the v3 image with one environment variable set at build time, TERMINAL_DESK_MAX_ACTIONS=3: evaluate pays 1 only if every check passes and the episode used at most 3 exec or write_file actions since reset (failed actions count; evaluate, close and session actions do not). An over-limit solve scores 0, the evaluate output's first line reads decisiveness gate: N actions used, limit 3, and info carries actions_used, gate_limit and gate_passed. Every reference solution pays under the gate (72 ids at two seeds). In a corrected screen of the untrained model (screen_v3_fixed.jsonl, 220 episodes, 4 per id on 55 ids, with the budget accounting fixed to end the episode where the arena does), 13 of the 18 solves on the v3c tasks used 3 actions or fewer, and over all ids the solves by action count were 2: 15, 3: 26, 4: 26, 5: 13, 6: 5. So the gate removes about a quarter of the untrained reward and leaves the rest for the policy to recover by committing earlier. Tasks: the 20 of v3c minus the three the arena scored 0 on every pass, 17 ids (office 6, software 4, finance 7; selected_tasks.txt holds the current submission's list).

Pre-registration, written before the POST. Judged on new points for the account (industrial 2, office 1, science 1 banked): office 2 or more, any software solve, any finance solve; a media, math or security solve would count too. Null rates over 82 runs: office 2 or more 41%, software 21%, finance 11%, media 4%. Since 10-14 01:00 UTC only 2 of 22 runs scored office 2 or more (32 of 60 before, Fisher p 0.0003, within-user p 0.0004; industrial did not move), so office 2 in this run would be stronger evidence than the 82-run rate suggests. Evidence over luck is two of the three in one run (about 15% under the null) or office 3 (2%). Expected total 2 to 5 of 40. Training-side predictions: reward starts near 0.35 (the gate's bite) and recovers above 0.55 within 60 updates; mean reply length falls below v3c's 530 tokens. If the gated reward sits below 0.2 for the first 40 updates, the run is a no-update draw and 0 to 2 is expected.

Result (run 4f3e6ad598da8089b5abf4ad, evaluated 2026-10-16 17:19 UTC): 3 of 40, industrial 1, office 2, delta -1 against the 4/40 reference. Office 2 is a new point for the account (industrial 2, office 2, science 1 banked, 5 of 40). Against the pre-registration: one of the three named points, not two, so not evidence for the gate by the rule above; since 10-14 01:00 UTC office 2 has come in 6 of 34 runs, so one hit is about a one-in-six draw. Training side: 153 updates; reward by 20-step block 0.47, 0.49, 0.49, 0.49, 0.59, 0.57, 0.57, 0.58 (the predicted recovery above 0.55 came near update 90, not 60); gradient on 80 of 153 steps; mean reply length flat at 590 tokens, not below v3c's 530; entropy 0.099 to 0.092. Per task over 9 visits: off3_ar_recon-L1 0.25 to 1.0; off3_ar_recon-L2 0 for five visits (41-token episodes: one reply, then the budget), then 0.25, 0.25, 1.0, 0.75; off3_pdf_xlsx_diff-L1 0 on every visit; fin3_portfolio_value-L3 and off3_meeting_slot-L3 above 0 on one and two visits. Curve: analysis/trackio_v4.json.

v5: the same regime again (terminal-desk-v5)

Why a replicate: office 2 after the gate is the first hit any of this account's pre-registrations named, and one hit at a one-in-six rate says nothing. A second office 2 under the same regime would have P about 0.03 under the post-change rate and would be the first repeatable signal on this board; a miss costs nothing the account has banked. Runs 2 to 4 also trained under a budget that ends an episode when the model reads a whole input file (off3_ar_recon-L2 averaged 41 completion tokens for five visits; the same task code behind another entrant's wrapper that sends the instruction once scored 0.81 over 13 visits, bochencs-multidomain-v5r1, card revision 5f854031), and whether the private evaluation starves its episodes the same way is unknown (asked in HF discussion #18 and board #240; no answer). The non-starving variant (8,192 / 16,384 token budgets; the guide allows up to 32,768) is the planned run 6, not this one, so that the replicate stays clean.

What changes: nothing in the image (:v4, digest 765d9c5c, gate K=3) or the budgets (4,096 / 8,192 tokens, 32 calls). The task list drops the one id that scored 0 on every visit (off3_pdf_xlsx_diff-L1) and adds five media ids (med3_silence_report L1 to L3, med3_sprite_count-L2, med3_subtitle_retime-L1; 0.5 to 0.75 on the corrected screen), because a media solve is a point the account lacks. 21 ids: office 5, software 4, finance 7, media 5 (selected_tasks.txt).

Pre-registration, written before the POST. Board points: office 3, any software, any finance, any media. Replication: office 2 or more again is the signal this run exists for (about 0.16 under the post-change rate, 0.03 for two in a row); office 1 or less says the run-4 hit was a draw, which is the expected outcome. Expected total 2 to 5. Training side: first-pass mean reward over the 16 shared ids within 0.1 of run 4's first pass on them (0.52), gradient on 40 to 60% of steps, media ids between 0.3 and 0.8 on the first pass.

Analysis: the public training curves of all 82 evaluated arena runs

analysis/ holds 34 Trackio features per run (entropy level and change, gradient-step count and share, completion length and change, reward level and gain, clipped ratio, tokens, steps, task count) against the private scores, with the chance checks (analysis/README.md). None predicts the score: the strongest Spearman is 0.23 and random shuffles beat it 53% of the time; leave-one-user-out models do no better than the mean. Only having any gradient lines up with score (the five zero-gradient runs average 1.6 solved, the other 77 average 2.95; permutation p 0.017), and the amount does not matter. The board total tracks runs submitted (Spearman 0.71). The one pattern that is not noise is on the evaluation side: office scores fell on 10-14 (table in analysis/README.md).

Verification

• 1,453 unit tests (52 of them on the gate: off by default, refused at start on a bad value, counted and paid as described above): for every task id and three seeds, the oracle script written through the environment scores 1; no file, an empty file, a file with one perturbed value, a symlink or a FIFO at the path scores 0; every level's extra rule changes the answer and an oracle that skips it scores 0; the live schema equals the operator's; session actions fail softly; the call cap auto-evaluates; the workspace is wiped per reset; a malformed CSV row is a failed check, not a server error. All pass locally and as root inside the image (where commands drop to nobody).
• Anonymous pull of the image, /health, openenv validate --url, and a WebSocket replay of all 72 task ids to a terminal reward.