File size: 5,617 Bytes
9368cc4 | 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 | #!/usr/bin/env python3
"""End-to-end episode tests with scripted agents -- no model involved.
Three controls:
perfect fixes the real cause -> must resolve
noop changes nothing -> must not resolve
symptom-fix silences the error at the symptom site -> must NOT resolve
The third is the one that matters. It is the fix a model reaches for when it
reads the traceback and stops there, and a benchmark that scored it as correct
would be rewarding exactly the behaviour it is meant to detect.
"""
import json, sys, tempfile, shutil
from pathlib import Path
EVAL = Path(__file__).resolve().parents[2] / "examples" / "llama-eval"
sys.path.insert(0, str(EVAL))
sys.path.insert(0, str(Path(__file__).parent))
from agentic_eval import (AgenticTask, Workspace, ToolBox, Runner, LANGS, # noqa: E402
materialise, run_episode, run_tests_against, score)
from build_corpus import sandbox_for # noqa: E402
fails = []
def check(name, cond, detail=""):
print(f" {'PASS' if cond else 'FAIL'} {name}" + (f" [{detail}]" if not cond else ""))
if not cond:
fails.append(name)
def scripted(*batches):
"""Turn lists of (tool, args) into an OpenAI-shaped chat callable."""
turns = list(batches)
state = {"i": 0}
def chat(messages, tools):
i = state["i"]
state["i"] += 1
if i >= len(turns):
return {"choices": [{"message": {"content": "done", "role": "assistant"}}],
"usage": {"prompt_tokens": 100, "completion_tokens": 10}}
calls = [{"id": f"c{i}_{j}", "type": "function",
"function": {"name": n, "arguments": json.dumps(a)}}
for j, (n, a) in enumerate(turns[i])]
return {"choices": [{"message": {"content": "", "role": "assistant",
"tool_calls": calls}}],
"usage": {"prompt_tokens": 500 * (i + 1), "completion_tokens": 60}}
return chat
corpus = [json.loads(l) for l in
(Path(__file__).parent / "agentic-corpus.jsonl").read_text().splitlines()]
by_id = {r["task_id"]: r for r in corpus}
task = AgenticTask.from_record(by_id["ledger-since-inclusive"])
print(f"task: {task.task_id} f2p={len(task.fail_to_pass)} p2p={len(task.pass_to_pass)}")
runner = Runner(sandbox_for("python"), LANGS["python"])
root = Path(tempfile.mkdtemp(prefix="episode-"))
AGENTS = {
"perfect": scripted(
[("list_files", {})],
[("read_file", {"path": "ledger/projections.py", "start_line": 1, "end_line": 60})],
[("read_file", {"path": "ledger/store.py"})],
[("edit_replace", {"path": "ledger/store.py",
"old_text": "if e.seq >= seq", "new_text": "if e.seq > seq"})],
[("lint", {"path": "ledger"})],
[("finish", {"summary": "made since() exclusive again"})],
),
"noop": scripted(
[("list_files", {})],
[("finish", {"summary": "nothing to do"})],
),
"symptom-fix": scripted(
[("read_file", {"path": "ledger/projections.py"})],
# delete the ordering guard so the crash goes away
[("edit_replace", {
"path": "ledger/projections.py",
"old_text": (" if event.seq <= self.last_seq:\n"
" raise SequenceError(\n"
" f\"event {event.seq} already applied "
"(at {self.last_seq})\")\n"),
"new_text": ""})],
[("finish", {"summary": "removed the exception"})],
),
}
try:
results = {}
for name, chat in AGENTS.items():
ws_root = root / name
materialise(task.files, ws_root)
ws = Workspace(ws_root)
tb = ToolBox(ws, linter=lambda t, w=ws_root: runner.lint(w, t))
ep, _ = run_episode(task, tb, chat, max_turns=20)
final = ws.snapshot()
outcomes = run_tests_against(final, task, runner, root, name)
sc = score(task, outcomes)
results[name] = (ep, sc)
print(f" [{name}] stop={ep.stop_reason} calls={ep.tool_calls} "
f"edits={ep.edits} errs={ep.tool_errors} -> "
f"resolved={sc['resolved']} f2p={sc['f2p_passed']}/{sc['f2p_total']} "
f"regressions={sc['n_regressions']}")
ep, sc = results["perfect"]
check("perfect agent resolves the task", sc["resolved"], json.dumps(sc))
check("perfect agent hit no tool errors", ep.tool_errors == 0, str(ep.tool_errors))
check("perfect agent terminated via finish", ep.stop_reason == "finished", ep.stop_reason)
check("perfect agent caused no regressions", sc["n_regressions"] == 0)
ep, sc = results["noop"]
check("noop agent does not resolve", not sc["resolved"], json.dumps(sc))
check("noop agent fixes nothing", sc["f2p_passed"] == 0, str(sc["f2p_passed"]))
ep, sc = results["symptom-fix"]
check("symptom-only fix made an edit", ep.edits == 1, str(ep.edits))
check("symptom-only fix is NOT scored as resolved", not sc["resolved"], json.dumps(sc))
check("symptom-only fix is caught as a regression",
sc["n_regressions"] > 0, json.dumps(sc))
# token accounting must be populated
ep, _ = results["perfect"]
check("token accounting recorded", ep.peak_context > 0 and ep.prompt_tokens > 0,
f"peak={ep.peak_context} prompt={ep.prompt_tokens}")
finally:
shutil.rmtree(root, ignore_errors=True)
print(f"\n{'ALL PASS' if not fails else 'FAILURES: ' + ', '.join(fails)}")
sys.exit(1 if fails else 0)
|