Files
claude-config/workflow-bench/handoff_bench.py
T

176 lines
8.5 KiB
Python

#!/usr/bin/env python3
"""Bench how a worker's step to a stronger model hands the chunk over: handoff_bench.py <work dir> start <case>-<n> ..., then run|grade <case>-<arm>-<n> ..., then summary
The question: when loop-claude steps a worker to a rung on another model, which handoff serves the stronger model
best? Each case is an archived chunk of this repo, as in advisor_bench.py, whose judge and files this bench shares.
- `start` runs WEAK on the item in the clone <work dir>/<case>-<n> until it has spent START_BUDGET, which
stands in for a run that ends asking for a higher rung. A worker that finishes within the budget leaves no
handoff, so `start` fails on it. It saves the change so far (<case>-start-<n>.diff), and a copy of the clone
(.tree) and of the session's transcript (.session.jsonl), which every arm starts from.
- `run` puts the clone and the session back as `start` left them, then continues on STRONG, one arm of a case at a
time. The resume arm resumes WEAK's session with handoff-resume-prompt.md. The compact arm does the same once WEAK
has compacted the session with `/compact`, whose stream it saves as <name>.compact.jsonl. The brief arm starts a
session that has the item and handoff-brief-prompt.md, and no transcript.
- `grade` is advisor_bench.py's, over these cases. `summary` counts an arm's cost and seconds from its own runs,
the compaction included, and leaves out the start, which the arms share.
Run each arm of each case once; the table shows a direction, not a significance.
"""
import concurrent.futures
import collections
import json
import pathlib
import shutil
import subprocess
import sys
import threading
import uuid
import advisor_bench
from advisor_bench import Arm
BENCH = pathlib.Path(__file__).resolve().parent
# The ladder's first step that changes the model: bin/model_routing/routing-ladder.json.
WEAK = Arm("claude-sonnet-5-5", "high")
STRONG = Arm("claude-opus-5-5", "medium")
ARMS = ("resume", "compact", "brief")
# Claude Code keeps a session's transcript at <CLAUDE_PROJECTS>/<the working directory's slug>/<session id>.jsonl.
CLAUDE_PROJECTS = pathlib.Path.home() / ".claude/projects"
# One per case and run number: its arms share one clone, so they run one at a time.
CLONE_LOCKS: collections.defaultdict = collections.defaultdict(threading.Lock)
# L 3 chunks, by todo item. advisor_bench's L 1 and L 2 chunks are too small: Sonnet 5.5 finished five of the six for
# under $0.20, before any handoff.
CASES = {
"143": ("45cb331a4", "545f3c347"),
"163": ("d4cf6e7fe", "d1b0fe3a4"),
"198": ("4a0ba5921", "fd60b6da"),
"210": ("4764d3d52", "dcbdc970"),
"227": ("eb7b50698", "3bde9cd6d"),
"239": ("d583f8d48", "47a40757"),
}
# Dollars WEAK may spend before the handoff. In a trial, $0.36 stopped Sonnet 5.5 on items 163 and 227 after 11 to 14
# turns, with about 250 lines of diff.
START_BUDGET = 0.36
CLAUDE_FLAGS = ["--setting-sources", "project", "--permission-mode", "bypassPermissions", "--output-format", "stream-json", "--verbose"]
def stopped_at_budget(lines):
"""Whether a claude stream's first result line, the worker's own, says the run stopped at --max-budget-usd."""
for line in lines:
try:
entry = json.loads(line)
except ValueError:
continue
if entry.get("type") == "result":
return entry.get("subtype") == "error_max_budget_usd"
return False
def run_prompt(case):
"""advisor_bench's prompt for the case's item, as the item's text stood at the start commit."""
base = CASES[case][0]
text = advisor_bench.item_text(advisor_bench.git(advisor_bench.REPO, "show", f"{base}:todo.md"), case)
return (BENCH / "advisor-run-prompt.md").read_text().format(base=base, item=case, text=text)
def claude(prompt, flags, clone, stream):
with stream.open("w") as out:
subprocess.run(["claude", "-p", prompt, *flags, *CLAUDE_FLAGS], cwd=clone, stdin=subprocess.DEVNULL, stdout=out, stderr=subprocess.STDOUT)
def start(work, name):
"""WEAK's capped run of <case>-<n>, and the tree and session it leaves the arms; the files are in the module docstring."""
case, number = name.split("-")
base = CASES[case][0]
clone, saved = work / name, work / f"{case}-start-{number}"
if clone.exists():
raise ValueError(f"{name} already started: remove {clone}")
advisor_bench.fresh_clone(base, clone)
session = str(uuid.uuid4())
stream = saved.with_suffix(".jsonl")
claude(run_prompt(case), [*WEAK.flags(), "--session-id", session, "--max-budget-usd", str(START_BUDGET)], clone, stream)
if not stopped_at_budget(stream.read_text().splitlines()):
raise ValueError("the weak worker finished within its budget, or its run failed, so there is no handoff to bench")
transcripts = list(CLAUDE_PROJECTS.glob(f"*/{session}.jsonl"))
if len(transcripts) != 1:
raise ValueError(f"expected one transcript of session {session} under {CLAUDE_PROJECTS}, found {len(transcripts)}")
shutil.copy(transcripts[0], saved.with_suffix(".session.jsonl"))
saved.with_suffix(".session").write_text(str(transcripts[0]))
# The worker may have committed against the bench rules, so the arms get its whole change since base, uncommitted.
advisor_bench.git(clone, "add", "--all")
saved.with_suffix(".diff").write_text(advisor_bench.git(clone, "diff", "--cached", base))
advisor_bench.git(clone, "reset", "-q", "--mixed", base)
shutil.copytree(clone, saved.with_suffix(".tree"), symlinks=True)
return f"{name}: started"
def run(work, name):
"""STRONG's run of one arm, <case>-<arm>-<n>, from the tree and session `start` left."""
case, arm, number = advisor_bench.split_name(name)
if arm not in ARMS:
raise ValueError(f"the arm {arm!r} is none of {ARMS}")
clone, saved = work / f"{case}-{number}", work / f"{case}-start-{number}"
if not saved.with_suffix(".tree").exists():
raise ValueError(f"start {case}-{number} first")
stream = work / f"{name}.jsonl"
if stream.exists():
raise ValueError(f"{name} already ran: remove {stream}")
with CLONE_LOCKS[clone]:
shutil.rmtree(clone)
shutil.copytree(saved.with_suffix(".tree"), clone, symlinks=True)
transcript = pathlib.Path(saved.with_suffix(".session").read_text())
shutil.copy(saved.with_suffix(".session.jsonl"), transcript)
resumed = ["--resume", transcript.stem]
if arm == "compact":
claude("/compact", [*WEAK.flags(), *resumed], clone, work / f"{name}.compact.jsonl")
if arm == "brief":
claude(run_prompt(case) + "\n" + (BENCH / "handoff-brief-prompt.md").read_text(), [*STRONG.flags(), "--no-session-persistence"], clone, stream)
else:
prompt = (BENCH / "handoff-resume-prompt.md").read_text().format(rung=f"{STRONG.model} at {STRONG.effort} effort")
claude(prompt, [*STRONG.flags(), *resumed], clone, stream)
outcome(work, name)
advisor_bench.save_change(work, name, clone, CASES[case])
return f"{name}: ran"
def outcome(work, name):
"""The cost and seconds of an arm's run: its STRONG stream's, plus its compaction's where it made one."""
streams = [stream for stream in (work / f"{name}.compact.jsonl", work / f"{name}.jsonl") if stream.exists()]
results = [advisor_bench.read_result(stream.read_text().splitlines()) for stream in streams]
return {"cost": sum(result.cost for result in results), "seconds": sum(result.seconds for result in results)}
def rows_of(work):
"""One row per graded run: its arm, verdict, cost and seconds."""
rows = []
for verdict_file in sorted(work.glob("*.verdict")):
name = verdict_file.stem
rows.append({"arm": advisor_bench.split_name(name)[1], "verdict": json.loads(verdict_file.read_text())["verdict"], **outcome(work, name)})
return rows
def summary(rows):
"""A markdown table with one row per arm, in ARMS' order: its runs, passes, partials, mean cost and mean seconds."""
lines = ["| arm | runs | pass | partial | mean cost $ | mean seconds |", "|---|---|---|---|---|---|"]
for name in ARMS:
own = [row for row in rows if row["arm"] == name]
if own:
passes, partials = (sum(row["verdict"] == kind for row in own) for kind in ("pass", "partial"))
cost, seconds = (sum(row[key] for row in own) / len(own) for key in ("cost", "seconds"))
lines.append(f"| {name} | {len(own)} | {passes} | {partials} | {cost:.2f} | {seconds:.0f} |")
return "\n".join(lines)
def main(argv):
work, step, names = pathlib.Path(argv[0]), argv[1], argv[2:]
if step == "summary":
print(summary(rows_of(work)))
return
steps = {"start": start, "run": run, "grade": lambda work, name: advisor_bench.grade(work, name, CASES)}
with concurrent.futures.ThreadPoolExecutor(max_workers=3) as pool:
for result in pool.map(lambda name: advisor_bench.attempted(steps[step], work, name), names):
print(result, flush=True)
if __name__ == "__main__":
main(sys.argv[1:])