Datasets:
Skills Desk v2: 36 screened training tasks (pre-registered rule), ev_contracts on CC0 USAspending data; reserved set unchanged
8685f4e verified | """Reference solutions played as an agent would: write one script, run it, evaluate (3 actions). | |
| Each must score 1 on several fixed instances. Common mistakes must score 0.""" | |
| import os | |
| import tempfile | |
| import pytest | |
| os.environ.setdefault("TERMINAL_DESK_APP", tempfile.mkdtemp(prefix="td-solve-")) | |
| from terminal_desk.families import all_task_ids # noqa: E402 | |
| from terminal_desk.models import Tbench2Action # noqa: E402 | |
| from terminal_desk.server.environment import TerminalDeskEnvironment # noqa: E402 | |
| HERE = os.path.join(os.path.dirname(__file__), "solvers") | |
| def solver(family: str) -> str: | |
| return open(os.path.join(HERE, f"{family}.py")).read() | |
| def play(task_id: str, script: str, extra: list | None = None) -> float: | |
| env = TerminalDeskEnvironment() | |
| env.reset(task_id=task_id) | |
| if "{UEI}" in script: # values an agent reads from the instruction | |
| script = script.replace("{UEI}", env._ep["built"].notes["uei"]) | |
| acts = [Tbench2Action(action_type="write_file", file_path="solve.py", content=script), | |
| Tbench2Action(action_type="exec", command="python3 solve.py")] + (extra or []) | |
| for act in acts: | |
| obs = env.step(act) | |
| assert obs.success, obs.error | |
| return env.step(Tbench2Action(action_type="evaluate")).reward | |
| def test_reference_solution_scores_one(task_id): | |
| family = task_id.split("-L")[0] | |
| assert play(task_id, solver(family)) == 1.0 | |
| MISTAKES = { | |
| # ignores cancellations | |
| "ev_backorder": ("cancel = {l.strip() for l in open(\"cancellations.txt\") if l.strip()}", "cancel = set()"), | |
| # forgets that revoked windows do not count | |
| "ev_calibration": ('if r["status"] == "active"]', "]"), | |
| # leaves videos without views out of total_hours as well | |
| "cmp_channels": (' s = secs(r["duration"])', ' if not r["views"].strip():\n continue\n s = secs(r["duration"])'), | |
| # sums duplicates instead of keeping the correction | |
| "cmp_rainfall": ('latest[(r["station"], r["date"])] = float(r["amount"]) * f[r["unit"]]', | |
| 'latest[(r["station"], r["date"], len(latest))] = float(r["amount"]) * f[r["unit"]]'), | |
| # keeps the N/A email as a string | |
| "tx_contacts": ('None if e in ("", "n/a") else e', 'None if e == "" else e'), | |
| # keeps escaped quotes in msg | |
| "tx_events": ('re.sub(r"\\\\(.)", r"\\1", m.group(3))', "m.group(3)"), | |
| # fixes only the crash | |
| "fix_report": ("src = src.replace('t[\"cost\"] / t[\"shipments\"]', 't[\"cost\"] / t[\"weight\"]')\n", ""), | |
| # greedy by value per kg | |
| "plan_load": ("for r in range(len(pool) + 1):", "for r in [len(pool)] if False else range(0):"), | |
| } | |
| # Second batch: one plausible mistake each. | |
| MISTAKES.update({ | |
| "ev_deliveries": ('if r["status"] == "received":', "if True:"), | |
| "ev_badges": (' role = roles["inherits"][role]; out |= set(roles["zones"][role])', " break"), | |
| "ev_dupes": ('("p", "".join(c for c in r["phone"] if c.isdigit()))', '("p", r["phone"])'), | |
| "ev_bankrec": ("0 <= (date.fromisoformat(b[\"booked\"]) - ld).days <= 3", "True"), | |
| "ev_fixedin": ("key < first[iss][0]", "v < first[iss][1]"), | |
| "cmp_meter": ("(b - a) % 100000", "abs(b - a)"), | |
| "cmp_grades": ("q = sorted(d[\"q\"])[1:]", "q = d[\"q\"]"), | |
| "cmp_fifo": ("t * lots[0][1]", "t * lots[-1][1]"), | |
| "cmp_smooth": ("sorted(v[i - 1:i + 2])[1] > limit", "v[i] > limit"), | |
| "cmp_loan": ("p = pay if m < n else bal + i", "p = pay"), | |
| "tx_xml": ('";".join(t.text for t in p.findall("tag"))', '",".join(t.text for t in p.findall("tag"))'), | |
| "tx_pivot": ('[cells[k].get(m, "") for m in metrics]', '[cells[k].get(m, "0") for m in metrics]'), | |
| "tx_mdtable": ('.replace("\\\\|", "|")', ""), | |
| "tx_sessions": (' if not line.strip(): continue\n', ""), | |
| "tx_fixedwidth": ('f"{-b if neg else b:.2f}"', 'f"{b:.2f}"'), | |
| "tx_dates": ('"us_branch": "%m/%d/%Y"', '"us_branch": "%d/%m/%Y"'), | |
| "fix_paginate": ('s = s.replace("start = number * per_page - per_page + 1", "start = (number - 1) * per_page")\n', ""), | |
| "fix_logline": ('s = s.replace("(\\\\d+)$", "(\\\\d+|-)$")\n', ""), | |
| "fix_units": ("s = s.replace(\"LENGTH_M[to_unit]\\n return value\\n\", 'LENGTH_M[to_unit]\\n raise ValueError(f\"cannot convert {from_unit} to {to_unit}\")\\n')\n", ""), | |
| "fix_bizdays": ('.replace("step = 1\\n", "step = 1 if n >= 0 else -1\\n")', ""), | |
| "fix_dedupe": ('row["email"] = row["email"].strip().lower()', 'row["email"] = row["email"].strip()'), | |
| "fix_mergelogs": ('.replace(" out += a[i:]\\n", " out += a[i:]\\n out += b[j:]\\n")', ""), | |
| "plan_disks": ("for k in range(1, len(names) + 1):", "for k in range(len(names), len(names) + 1):"), | |
| "plan_delivery": ("load[d] + int(o[\"weight_kg\"]) > cap[d]", "False"), | |
| # reports obligated minus posted | |
| "ev_contracts": ("posted[(a.upper(), mod(m))] - ob", "ob - posted[(a.upper(), mod(m))]"), | |
| }) | |
| # Mistakes that only the L2 variants catch (the extra rule is ignored). | |
| MISTAKES_L2 = { | |
| "ev_backorder/receipts": ('if os.path.exists("receipts.csv"):', "if False:"), | |
| "ev_backorder/aliases": ('if os.path.exists("sku_aliases.csv") else {}', "if False else {}"), | |
| "cmp_rainfall/scale": ('if os.path.exists("gauges.csv") else []', "if False else []"), | |
| "cmp_rainfall/qc": ('if os.path.exists("qc.csv") else set()', "if False else set()"), | |
| } | |
| def test_common_mistake_scores_zero(family): | |
| old, new = MISTAKES[family] | |
| src = solver(family) | |
| assert old in src, family | |
| bad = src.replace(old, new) | |
| if family == "plan_load": # no search: load only the priority items | |
| bad = bad.replace("best = None", "best = (sum(int(i['value']) for i in must), base, sorted(i['item_id'] for i in must))") | |
| rewards = [play(f"{family}-L1" + (f"-i{k}" if k else ""), bad) for k in range(4)] | |
| assert max(rewards) == 0.0, rewards | |
| def test_l2_mistake_scores_zero(name): | |
| family = name.split("/")[0] | |
| old, new = MISTAKES_L2[name] | |
| src = solver(family) | |
| assert old in src, name | |
| rewards = [play(f"{family}-L2" + (f"-i{k}" if k else ""), src.replace(old, new)) for k in range(4)] | |
| assert max(rewards) == 0.0, rewards | |
| def test_unrepaired_report_with_correct_summary_scores_zero(): | |
| """Hand-writing out/summary.json without repairing report.py must fail the hidden-input run.""" | |
| from terminal_desk.core.grading import render_artifact | |
| env = TerminalDeskEnvironment() | |
| env.reset(task_id="fix_report-L1") | |
| art = env._ep["built"].artifact | |
| env.step(Tbench2Action(action_type="write_file", file_path=art.path, content=render_artifact(art))) | |
| assert env.step(Tbench2Action(action_type="evaluate")).reward == 0.0 | |
| def test_roster_breaking_one_rule_scores_zero(): | |
| env = TerminalDeskEnvironment() | |
| env.reset(task_id="plan_shifts-L1") | |
| oracle = env._ep["built"].notes["oracle"] | |
| rows = [[s, p] for s, ps in oracle.items() for p in ps] | |
| rows = rows[:-1] # one shift is now short of its headcount | |
| content = "shift_id,staff\n" + "".join(f"{s},{p}\n" for s, p in rows) | |
| env.step(Tbench2Action(action_type="write_file", file_path="roster.csv", content=content)) | |
| assert env.step(Tbench2Action(action_type="evaluate")).reward == 0.0 | |
| def test_course_plan_breaking_only_prerequisites_scores_zero(): | |
| """Selecting by every rule except prerequisites must fail. A solver that ignores one rule can | |
| still land on a valid plan, so this searches for a plan that breaks only that rule.""" | |
| import csv | |
| import io | |
| import itertools | |
| import json | |
| import re | |
| checked = 0 | |
| for k in range(4): | |
| task = "plan_courses-L1" + (f"-i{k}" if k else "") | |
| env = TerminalDeskEnvironment() | |
| env.reset(task_id=task) | |
| f = env._ep["built"].files | |
| lo, hi = map(int, re.search(r"between (\d+) and (\d+)", f["PLAN_RULES.md"]).groups()) | |
| done = {x.strip() for x in f["transcript.txt"].splitlines()[1:] if x.strip()} | |
| req = {x.strip() for x in f["required.txt"].splitlines() if x.strip()} | |
| cs = {r["code"]: r for r in csv.DictReader(io.StringIO(f["courses.csv"]))} | |
| opts = [c for c, r in cs.items() if c not in done and r["offered"] == "yes"] | |
| blocked = lambda c: not set(filter(None, cs[c]["prerequisites"].split(";"))) <= done # noqa: E731 | |
| bad = next((sel for n in range(1, 7) for sel in itertools.combinations(opts, n) | |
| if any(blocked(c) for c in sel) and req <= set(sel) and len({cs[c]["slot"] for c in sel}) == n | |
| and lo <= sum(int(cs[c]["credits"]) for c in sel) <= hi), None) | |
| if bad is None: | |
| continue | |
| checked += 1 | |
| env.step(Tbench2Action(action_type="write_file", file_path="plan.json", content=json.dumps({"courses": list(bad)}))) | |
| assert env.step(Tbench2Action(action_type="evaluate")).reward == 0.0 | |
| assert checked >= 2 | |