| --- |
| license: apache-2.0 |
| pretty_name: ShiftDesk |
| tags: [openenv, reinforcement-learning, agents, tool-use] |
| configs: |
| - config_name: default |
| data_files: |
| - split: train |
| path: tasks.jsonl |
| --- |
| # ShiftDesk v4 |
|
|
| An OpenEnv environment for OpenEnv Arena. v4 adds a terminal: a `bash` tool and a `write_file` tool next to |
| `list_files`, `read_file` and `run_python`, a working directory that persists for the whole episode, and 32 new |
| **file-deliverable** tasks (16 families x 2 levels) in the four domains where earlier versions never scored: |
| natural science, math and formal reasoning, cybersecurity, and media. Each asks for a real output file |
| (`output/results.csv`, `output/incident.json`, `output/out.srt`, `output/article.html`, ...) that hidden |
| checks grade on submit. All 47 v3 answer-style tasks are still served. |
|
|
| The training list (`in_training_list` in `tasks.jsonl`) has 50 tasks, the arena's maximum: the 32 new ones (64%) |
| interleaved with 18 v3 tasks from software, industrial, office and finance (36%), so the domains that already |
| score keep training. |
| Every reset generates a fresh instance from a seeded generator, so there are no fixed answer keys and this |
| dataset holds none: `tasks.jsonl` lists each task id with a seed-0 example question, its deliverable path or |
| answer format, and its file names. |
|
|
| ## What the agent does |
|
|
| Each episode is a working directory with the task files and an empty `output/` folder. Tools, each shown with a |
| one-line JSON example in the first observation: |
|
|
| - `list_files` `{}`: everything in the working directory, including files the agent wrote |
| - `read_file` `{"path": ..., "offset": optional first line}` (2,500 characters per call) |
| - `write_file` `{"path": ..., "content": ...}`: create or overwrite a file inside the working directory |
| - `bash` `{"command": ...}`: a shell with python3, grep, awk, sed, sort, jq, sqlite3, bc, openssl, file |
| - `run_python` `{"code": ...}`: Python 3 with the standard library, numpy, sympy and PIL |
| - `submit` `{"answer": ...}` ends the episode (for file tasks the answer can be empty) |
|
|
| `bash` and `run_python` run as `nobody` with CPU, memory, file-size and process limits and a 20 s wall timeout; |
| `write_file` writes through the same sandboxed child. Outbound IPv4/IPv6 sockets are refused by a seccomp |
| filter. The instance salt and the server code are root-only. |
|
|
| | Domain | New v4 file families (levels 1, 2) | |
| |---|---| |
| | Natural science | assay-reduction: plate-reader calibration with mixed units, QC-flagged wells, replicates and out-of-range samples -> results.csv; peak-picking: spectrum peak table with baseline windows and cosmic-ray removal -> peaks.json; dilution-plan: direct and serial dilution plans from mg/mL stocks -> plan.csv; pk-fit: non-compartmental PK (AUC, k_el, half-life) with BLQ handling and clock-time sampling -> pk.json | |
| | Math & formal reasoning | linear-system: exact rational solutions with a redundant and a voided equation -> solution.json; recurrence-id: smallest linear recurrence (with constant) and a far term mod p -> answer.json; claim-check: exhaustive checks with smallest counterexamples -> verdicts.json; board-expectation: exact expected turns of a dice game with exact-landing and bonus rolls -> answer.json | |
| | Cybersecurity | web-forensics: nginx log attacker, UTC timeline and successful payloads in an unordered merged log -> incident.json; decode-chain: layered base64/hex/rot13/URL/XOR blobs and salted hash cracking -> decoded.json; sshd-audit: effective sshd_config (first value wins, defaults, Match blocks) against a baseline -> findings.json; jwt-inspect: HS256 verification with kid, exp/nbf, issuer, audience and a check order -> report.json | |
| | Media & content | srt-retime: offset and two-point linear retiming with credit removal and renumbering -> out.srt; image-audit: real format, EXIF-rotated display size and extension checks -> report.csv; audio-audit: WAV duration, peak dBFS and clipping -> audio_report.csv; markdown-html: Markdown to house-style HTML (slug ids, demoted h1s, external-link attributes) -> article.html | |
| |
| v3 answer-style families (all still served): change-impact, version-resolve, module-repair (software); |
| bom-explosion, tank-balance, line-oee (industrial); beer-lambert, stoichiometry, kinetics-lab (science); |
| meeting-slot, invoice-match, expense-policy (office); npv-choice, bond-desk, fifo-basis, loan-desk (finance); |
| congruences, logic-grid, shortest-route, exact-probability, exact-count (math); firewall-audit, vuln-triage, |
| auth-hunt (security); edit-timeline, headline-style, subtitle-qc (media). |
| |
| ## What it is meant to teach |
| |
| Terminal work the way held-out skills tasks are shaped: read every file (level 2 hides a caveat in a second file: |
| QC flags, an SOP step, errata, a sync sheet, a style guide), compute with code instead of by eye, and leave a |
| correctly formatted output file where the task asks for it. |
| |
| ## Reward |
| |
| Binary. File tasks: 1.0 only when the deliverable exists at the requested path, is a regular file (symlinks are |
| refused), parses, and passes every hidden check (stated numeric tolerances; ids, lists, timestamps and HTML |
| structure exact); otherwise 0.0. The deliverable is graded on `submit`, on the finish action, and when calls run |
| out. Answer tasks (v3): 1.0 only when every requested JSON field is right. The finish action |
| `{"tool": "submit", "args": {"answer": ""}}` scores 0.0 unless the deliverable file is already correct. |
| |
| ## Same instance for a group |
| |
| Rollouts that reset the same task within the same 5-minute window get the same instance (seeded from a |
| secret in the image), so the parallel rollouts of an optimizer step are compared on one problem. |
| |
| ## Checks |
| |
| For every task over 60+ seeds: the reference scores 1, plausible wrong outputs (what you get by skipping the |
| caveat) and empty outputs score 0, and generation is deterministic per seed. For all 16 new families an |
| independent solver that reads only the question and files plays full episodes through `write_file`, `bash` and |
| `run_python` and earns 1 on every seed tried (150 seeds per task offline); an empty submit, the right content |
| passed only as the answer, an empty file, a symlinked deliverable and a wrong file all earn 0. `openenv validate` |
| passes and every task replays to a terminal reward over WebSocket with the new tools. |
|
|
| Source: https://github.com/nice-bills/shiftdesk |
|
|
