agent-harness / scripts /build_study4_tasks.py
cuber12's picture
Publish agent harness research code and paper artifacts
d61821a verified
Raw
History Blame Contribute Delete
6.75 kB
"""Deterministically select fresh, test-backed tasks for Study 4.
This outcome-blind selector reads repository history and test outcomes only. It
excludes every commit already present in the task catalog, validates that public
tests pass at the parent, hidden tests expose the bug, and the gold production
patch repairs it, then retains every candidate decision.
"""
from __future__ import annotations
import argparse
from dataclasses import asdict
import json
from pathlib import Path
from build_study2_tasks import (
candidate_commits,
changed_symbols,
manifest_text,
precheck,
run_git,
validate_candidate,
load_repository,
)
from agent_harness.specs import load_tasks
REPOSITORIES = ("R001", "R002", "R003")
def select_fresh(root: Path, repository_id: str, count: int, write: bool) -> dict:
rule = load_repository(root, repository_id)
catalog = load_tasks(root)
excluded_commits = {task.gold_commit for task in catalog.values()}
selected: list[str] = []
audits = []
manifest_dir = root / "tasks" / "manifests"
patch_dir = root / "tasks" / "patches"
validation_dir = root / "tasks" / "validation" / "study4"
selection_dir = root / "tasks" / "selection" / "study4"
if write:
validation_dir.mkdir(parents=True, exist_ok=True)
selection_dir.mkdir(parents=True, exist_ok=True)
for commit in candidate_commits(rule):
if len(selected) >= count:
break
audit = precheck(rule, commit, excluded_commits)
audits.append(audit)
if audit.status != "eligible":
continue
assert audit.parent is not None
source_patch = str(
run_git(rule.path, "diff", "--binary", audit.parent, commit, "--", *audit.source_paths)
)
test_patch = str(
run_git(rule.path, "diff", "--binary", audit.parent, commit, "--", *audit.test_paths)
)
validation = validate_candidate(rule, audit, source_patch, test_patch)
audit.validation = validation
audit.elapsed_seconds += float(validation["elapsed_seconds"])
if not validation["valid"]:
audit.status = "rejected"
audit.reason = str(validation["reason"])
print(f"REJECT {repository_id} {commit[:12]} {audit.reason}", flush=True)
continue
task_id = f"TASK_S4_{repository_id}_{len(selected) + 1:03d}"
source_name = f"{task_id}_source.patch"
test_name = f"{task_id}_tests.patch"
symbols = changed_symbols(rule, audit)
audit.task_id = task_id
audit.status = "selected"
audit.reason = None
selected.append(task_id)
excluded_commits.add(commit)
if write:
text = manifest_text(
rule, audit, task_id, source_name, test_name, symbols
).replace(
"deterministic Study 2 held-out validation",
"deterministic Study 4 fresh-task validation",
)
(manifest_dir / f"{task_id}.toml").write_text(text, encoding="utf-8")
(patch_dir / source_name).write_text(source_patch, encoding="utf-8")
(patch_dir / test_name).write_text(test_patch, encoding="utf-8")
(validation_dir / f"{task_id}.json").write_text(
json.dumps(
{
"schema_version": 1,
"study": "Study 4 fresh-task retrieval replication",
"task_id": task_id,
"repository_id": repository_id,
"commit": commit,
"parent": audit.parent,
"valid_end_to_end": True,
"validation": validation,
},
indent=2,
sort_keys=True,
)
+ "\n",
encoding="utf-8",
)
print(
f"SELECT {task_id} {commit[:12]} source={len(audit.source_paths)} "
f"tests={len(audit.test_paths)} lines={audit.source_changed_lines}",
flush=True,
)
report = {
"schema_version": 1,
"study": "Study 4 fresh-task retrieval replication",
"repository": asdict(rule) | {"path": str(rule.path)},
"selection_policy": {
"fresh_count": count,
"excluded_prior_task_commits": len({task.gold_commit for task in catalog.values()}),
"order": "reverse chronological first-parent history since 2023-01-01",
"outcome_blind": True,
},
"selected_task_ids": selected,
"selected_count": len(selected),
"complete": len(selected) == count,
"audits": [asdict(item) for item in audits],
}
if write:
(selection_dir / f"{repository_id}_selection.json").write_text(
json.dumps(report, indent=2, sort_keys=True) + "\n", encoding="utf-8"
)
if len(selected) != count:
raise RuntimeError(f"{repository_id} yielded {len(selected)}/{count} fresh valid tasks")
return report
def freeze_split(root: Path, count: int) -> tuple[str, ...]:
catalog = load_tasks(root)
expected = [
f"TASK_S4_{repository_id}_{number:03d}"
for repository_id in REPOSITORIES
for number in range(1, count + 1)
]
missing = [task_id for task_id in expected if task_id not in catalog]
if missing:
raise RuntimeError(f"cannot freeze Study 4 split; missing {missing}")
path = root / "tasks" / "splits" / "study4_fresh.txt"
path.write_text(
"# Fresh Study 4 split; frozen before any E10 LLM inference.\n"
+ "\n".join(expected)
+ "\n",
encoding="utf-8",
)
return tuple(expected)
def main() -> None:
parser = argparse.ArgumentParser()
parser.add_argument("--root", type=Path, default=Path(__file__).resolve().parents[1])
parser.add_argument("--repository-id", choices=REPOSITORIES)
parser.add_argument("--count", type=int, default=10)
parser.add_argument("--write", action="store_true")
parser.add_argument("--freeze-split", action="store_true")
args = parser.parse_args()
root = args.root.resolve()
if args.freeze_split:
split = freeze_split(root, args.count)
print(json.dumps({"split_count": len(split), "task_ids": split}, indent=2))
return
if not args.repository_id:
parser.error("--repository-id is required unless --freeze-split is used")
report = select_fresh(root, args.repository_id, args.count, args.write)
print(json.dumps({"repository_id": args.repository_id, "selected": report["selected_count"]}))
if __name__ == "__main__":
main()