Files
headroom/benchmarks/codemode_ab.py
6b7b6358d5 fix(benchmarks): refuse a dirty or non-git --worktree in codemode_ab (#4736)
## Description

`benchmarks/codemode_ab.py` runs `git reset --hard -q` and `git clean
-qfd` in `--worktree` after any task that leaves `git status
--porcelain` non-empty (`main()`, lines 319-324 on `1da8930a`). The
module docstring assumes a clean detached worktree, but nothing checks
that `--worktree` is a git work tree or was clean before the first run.
If it holds uncommitted work, that work is deleted after the first task.

With this PR `main()` checks once before the loop: if `--worktree` is
not inside a git work tree, or `git status --porcelain` is not empty, it
prints a message on stderr and returns exit code 2. The per-run reset is
unchanged.

Fixes #4731

## Type of Change

- [x] Bug fix (non-breaking change that fixes an issue)
- [ ] New feature (non-breaking change that adds functionality)
- [ ] Breaking change (fix or feature that would cause existing
functionality to change)
- [ ] Documentation update
- [ ] Performance improvement
- [ ] Code refactoring (no functional changes)

## Changes Made

- `benchmarks/codemode_ab.py`: after `wt =
Path(args.worktree).resolve()`, run `git status --porcelain` in `wt`. A
non-zero exit gives `error: --worktree <path> is not inside a git work
tree`; non-empty output gives an error naming the reset/clean commands
and suggesting commit, stash or a clean detached worktree. Both return 2
before any `claude` call.
- `tests/test_codemode_ab_worktree_guard.py`: new regression tests for
the dirty, non-git and clean cases (follow-up commit).

## Testing

- [x] Unit tests pass (`pytest`, new file only)
- [ ] Linting passes (`ruff check .`)
- [ ] Type checking passes (`mypy headroom`)
- [x] New tests added for new functionality
- [x] Manual testing performed

`tests/test_codemode_ab_worktree_guard.py` (new) calls `main()` on a tmp
git repo with `run_one` stubbed, so the `claude` CLI is never reached:
dirty tree (modified tracked file plus untracked file) must return 2
with the files intact; a plain directory must return 2 with the
not-a-git-work-tree message; a clean repo must return 0. Only that file
was run, not the full suite; `ruff format` and `ruff check` on it are
clean; `mypy` was not run. The script was also checked end to end
against a stub `claude` (below).

pytest, Ubuntu on WSL2, Python 3.12, `-p no:cacheprovider`.

Pass after (this branch):

```text
collected 3 items

tests/test_codemode_ab_worktree_guard.py ...                             [100%]

============================== 3 passed in 21.33s ==============================
```

Fail before (same test file, `benchmarks/codemode_ab.py` from
`origin/main` `1da8930a`):

```text
E   assert 0 == 2
E   assert 0 == 2
FAILED tests/test_codemode_ab_worktree_guard.py::test_dirty_worktree_is_refused_and_left_intact
FAILED tests/test_codemode_ab_worktree_guard.py::test_non_git_directory_is_refused
========================= 2 failed, 1 passed in 17.63s =========================
```

On `origin/main` the dirty and non-git cases return 0 instead of 2
because there is no guard (the test stops at that first assertion; the
reset of the dirty tree is shown by the stub run above). The clean case
passes on both, as it should.

### Test Output

Same stub `claude` (prints a fixed JSON result, touches nothing) and
same fixture on both versions of the script: a git repo with one commit,
then ` M tracked.txt` and an untracked `notes.md`. Ubuntu on WSL2,
Python 3, `python3 -I`, `HOME` pointed at an empty dir.

Fail before (`benchmarks/codemode_ab.py` from `origin/main` `1da8930a`):

```text
before: [ M tracked.txt ?? notes.md ]
  ccr_py_count           steered  rep0  $0.0000  turns=1  calls=None  ok=True
    !! worktree dirtied, resetting: M tracked.txt
?? notes.md
after: status=[] tracked=[committed ] notes.md=NO
```

Pass after (this branch):

```text
before: [ M tracked.txt ?? notes.md ]
error: --worktree /var/tmp/cb-u16-ab/wt-branch has uncommitted changes; the script runs `git reset --hard` and `git clean -fd` there after each run. Commit or stash first, or use a clean detached worktree.
exit=2
after: status=[ M tracked.txt ?? notes.md ] tracked=[committed MY EDIT ] notes.md=yes
```

Not a git directory (this branch): `error: --worktree
/var/tmp/cb-u16-ab/nogit is not inside a git work tree`, exit 2.

Clean fixture (this branch), full run with the stub: `exit=0`, last line
`wrote .../out/r.json`, 16 records written.

## Real Behavior Proof

- Environment: Ubuntu on WSL2, Python 3, git, stub `claude` first on
`PATH`, `HOME` redirected to an empty directory. No real `claude` run
and no model call.
- Exact command / steps: `python3 -I -u benchmarks/codemode_ab.py
--worktree <fixture> --plugin-dir <empty dir> --out <tmp>/r.json`, once
per script version and fixture.
- Observed result: on `1da8930a` the dirty fixture ended with an empty
status, `tracked.txt` back to `committed` and `notes.md` deleted; on
this branch the script exited 2 with the message above and both files
were untouched. On a clean fixture this branch still completed all runs
with exit 0.
- Not tested: a real `claude -p` run; a `--worktree` path that does not
exist (the `git` call would raise `FileNotFoundError`); worktrees made
with `git worktree add`.

## Runtime Rollout Safety

- Rollout-managed feature(s): none; benchmark script only.
- Minimum rollout channel: n/a
- Stable/default behavior changed: no, for the library or proxy. The
script now exits 2 instead of resetting a dirty `--worktree`.
- Kill switch / disable path: n/a
- Unsafe override required: no
- Qualification impact: none
- Rollback path: revert this commit.

## Review Readiness

- [x] I have performed a self-review
- [x] This PR is ready for human review

## Checklist

- [x] My code follows the project's style guidelines
- [x] I have performed a self-review of my code
- [x] I have commented my code, particularly in hard-to-understand areas
- [ ] I have made corresponding changes to the documentation
- [x] My changes generate no new warnings
- [x] I have added tests that prove my fix is effective or that my
feature works
- [ ] New and existing unit tests pass locally with my changes (only the
new test file was run; see Testing)
- [x] I did **not** edit `CHANGELOG.md` — it is generated by
release-please from my Conventional Commit PR title (a CI guard enforces
this)

## Additional Notes

- Known limit: a worktree edited while a benchmark is running is still
reset by the existing per-run cleanup. The check runs once, before the
first run.
- A user who wants to reuse a dirty tree now has to commit or stash
first. I kept the check to the smallest form; no `--force` flag was
added.

Produced with AI assistance (Claude Code) operated by me (HetCreep); the
stub-claude runs above were done on my machine against a clone of the
repository.

🤖 Generated with [Claude Code](https://claude.com/claude-code)

---------

Co-authored-by: Claude Sonnet 5.5 <noreply@anthropic.com>
Co-authored-by: JD Davis <mxjerrett@gmail.com>
2026-11-11 09:58:13 -06:00

353 lines
12 KiB
Python

#!/usr/bin/env python3
"""Live A/B for the headroom-codemode probe-batching directive.
Runs a fixed exploration-task suite through `claude -p` twice per task -- once
with the steering plugin active (HEADROOM_CODEMODE=1) and once without -- in a
clean detached worktree, then scores cost, turns, burst structure, bytes
fetched and answer correctness from the resulting session transcripts.
Replay benchmarks cannot measure this: steering changes which calls the model
emits, so the counterfactual is not present in recorded sessions.
"""
from __future__ import annotations
import argparse
import json
import os
import re
import subprocess
import sys
import time
from pathlib import Path
PROJECTS = Path.home() / ".claude" / "projects"
# (id, prompt, ground truth, kind)
TASKS = [
(
"ccr_py_count",
"How many .py files are directly in headroom/ccr/? Answer with just the number.",
"10",
"num",
),
(
"max_rounds",
"What is the default value of max_retrieval_rounds in ResponseHandlerConfig? Answer with just the number.",
"3",
"num",
),
(
"plugin_dirs",
"How many directories are directly under plugins/? Answer with just the number.",
"5",
"num",
),
(
"logcompressor_lines",
"How many lines are in crates/headroom-core/src/transforms/log_compressor.rs? Answer with just the number.",
"1795",
"num",
),
(
"openclaw_version",
"What version is declared in plugins/openclaw/package.json? Answer with just the version string.",
"0.37.0",
"exact",
),
(
"test_ccr_files",
"How many files directly in tests/ (not in subdirectories) have names starting with test_ccr? Answer with just the number.",
"24",
"num",
),
(
"crate_names",
"List the directory names under crates/ as a comma-separated sorted list, nothing else.",
"headroom-core,headroom-parity,headroom-proxy,headroom-py,headroom-simulators",
"set",
),
(
"agenthooks_events",
"How many distinct hook event names are defined in plugins/headroom-agent-hooks/hooks/hooks.json? Answer with just the number.",
"2",
"num",
),
]
# Multi-probe suite: each task needs several small independent facts, i.e. the
# exact shape the directive targets. This measures the CEILING for steering,
# not a field average -- real observed bursts average 2.6 calls.
TASKS_MULTI = [
(
"crate_rs_counts",
"For each directory under crates/, report how many .rs files it contains recursively. "
"Answer as name=count pairs, comma-separated, sorted by name, nothing else.",
"headroom-core=99,headroom-parity=3,headroom-proxy=86,headroom-py=2,headroom-simulators=7",
"seq",
),
(
"file_existence",
"Do these paths exist? pyproject.toml, uv.lock, Cargo.toml, mkdocs.yml, tsconfig.json. "
"Answer as five yes/no values, comma-separated, in that order, nothing else.",
"yes,yes,yes,no,no",
"seq",
),
(
"dir_counts",
"Report three numbers: how many .py files are directly in benchmarks/, how many .md files "
"are directly in wiki/, and how many directories are directly under plugins/. "
"Answer as three comma-separated numbers in that order, nothing else.",
"25,35,5",
"seq",
),
(
"ccr_symbols",
"In headroom/ccr/, report how many .py files contain the string 'CCR_TOOL_NAME', how many "
"contain 'async def', and how many contain 'class '. Answer as three comma-separated numbers "
"in that order, nothing else.",
"6,5,8",
"seq",
),
(
"plugin_versions",
"Report the version field from each of: plugins/openclaw/package.json, "
"plugins/headroom-agent-hooks/.claude-plugin/plugin.json. "
"Answer as two comma-separated version strings in that order, nothing else.",
"0.37.0,0.37.0",
"seq",
),
(
"line_counts",
"Report the line count of each of these files: headroom/ccr/response_handler.py, "
"headroom/ccr/tool_injection.py, headroom/ccr/mcp_server.py. "
"Answer as three comma-separated numbers in that order, nothing else.",
"1095,615,1199",
"seq",
),
]
def score(kind: str, truth: str, answer: str) -> bool:
"""Grade an answer.
Answers often carry a trailing citation ("3\n\n(file.py:72)"), so numeric
grading reads the FIRST number, not the last -- otherwise a line number in
the citation is graded as the answer.
"""
a = (answer or "").strip().replace("`", "")
if kind == "num":
nums = re.findall(r"\d[\d,]*", a)
return bool(nums) and nums[0].replace(",", "") == truth
if kind == "exact":
return truth in a
if kind == "seq":
want = [t.strip().lower() for t in truth.split(",")]
# take the first len(want) comma/whitespace-separated fields of the
# first line that has at least that many -- ignores trailing prose
for line in [a] + a.splitlines():
got = [t.strip().lower() for t in re.split(r"[,\s]+", line) if t.strip()]
if len(got) >= len(want) and got[: len(want)] == want:
return True
return False
if kind == "set":
got = {t.strip().lower() for t in re.split(r"[,\s]+", a) if t.strip()}
want = {t.strip().lower() for t in truth.split(",")}
return want.issubset(got)
return False
def trace(session_id: str) -> dict:
"""Pull burst structure and bytes fetched out of the session transcript."""
hits = list(PROJECTS.glob(f"**/{session_id}.jsonl"))
if not hits:
return {"trace_found": False}
rows = []
for line in hits[0].open(errors="replace"):
try:
rows.append(json.loads(line))
except Exception:
pass
calls = bash_calls = 0
bursts: list[int] = []
run = 0
fetched = 0
pend: set[str] = set()
batched = 0 # compound shell commands (a proxy for directive compliance)
def text_of(x):
if isinstance(x, str):
return x
if isinstance(x, list):
return "\n".join(
b.get("text", "")
if isinstance(b, dict) and b.get("type") == "text"
else json.dumps(b)
for b in x
)
return json.dumps(x) if x is not None else ""
for d in rows:
t = d.get("type")
if t == "assistant":
tus = [
b
for b in (d.get("message") or {}).get("content") or []
if isinstance(b, dict) and b.get("type") == "tool_use"
]
if tus:
run += 1
for b in tus:
calls += 1
pend.add(b.get("id"))
if b.get("name") == "Bash":
bash_calls += 1
cmd = (b.get("input") or {}).get("command") or ""
if re.search(r";|&&", cmd):
batched += 1
else:
if run:
bursts.append(run)
run = 0
elif t in ("user", "attachment"):
for b in (d.get("message") or {}).get("content") or []:
if (
isinstance(b, dict)
and b.get("type") == "tool_result"
and b.get("tool_use_id") in pend
):
pend.discard(b.get("tool_use_id"))
fetched += len(text_of(b.get("content")))
if run:
bursts.append(run)
return {
"trace_found": True,
"tool_calls": calls,
"bash_calls": bash_calls,
"compound_bash": batched,
"bursts": len(bursts),
"max_burst": max(bursts) if bursts else 0,
"mean_burst": round(sum(bursts) / len(bursts), 2) if bursts else 0,
"fetched_chars": fetched,
}
def run_one(task, arm: str, wt: Path, plugin_dir: Path, model: str | None) -> dict:
tid, prompt, truth, kind = task
env = dict(os.environ)
env["HEADROOM_CODEMODE"] = "1" if arm == "steered" else "0"
cmd = [
"claude",
"-p",
prompt,
"--output-format",
"json",
"--permission-mode",
"auto",
"--plugin-dir",
str(plugin_dir),
]
if model:
cmd += ["--model", model]
t0 = time.time()
proc = subprocess.run(cmd, cwd=wt, env=env, capture_output=True, text=True, timeout=600)
wall = time.time() - t0
try:
d = json.loads(proc.stdout)
except Exception:
return {
"task": tid,
"arm": arm,
"error": (proc.stdout or proc.stderr)[:300],
"wall_s": wall,
}
u = d.get("usage") or {}
rec = {
"task": tid,
"arm": arm,
"session_id": d.get("session_id"),
"cost_usd": d.get("total_cost_usd"),
"num_turns": d.get("num_turns"),
"wall_s": round(wall, 1),
"input_tokens": u.get("input_tokens", 0),
"cache_creation": u.get("cache_creation_input_tokens", 0),
"cache_read": u.get("cache_read_input_tokens", 0),
"output_tokens": u.get("output_tokens", 0),
"answer": str(d.get("result", ""))[:200],
}
rec["correct"] = score(kind, truth, rec["answer"])
rec.update(trace(rec["session_id"] or ""))
return rec
def main() -> int:
ap = argparse.ArgumentParser()
ap.add_argument("--worktree", required=True)
ap.add_argument("--plugin-dir", required=True)
ap.add_argument("--reps", type=int, default=1)
ap.add_argument("--model", default=None)
ap.add_argument("--out", default="benchmark_results/codemode_ab.json")
ap.add_argument("--suite", choices=["simple", "multi"], default="simple")
args = ap.parse_args()
wt = Path(args.worktree).resolve()
# The per-run cleanup below is `git reset --hard` + `git clean -fd`: refuse a
# worktree that is not a clean git checkout so it cannot destroy user work.
# --untracked-files=all: `status.showUntrackedFiles=no` would hide the untracked
# files that `git clean -fd` deletes.
git = subprocess.run(
["git", "status", "--porcelain", "--untracked-files=all"],
cwd=wt,
capture_output=True,
text=True,
)
if git.returncode != 0:
print(f"error: --worktree {wt} is not inside a git work tree", file=sys.stderr)
return 2
if git.stdout.strip():
print(
f"error: --worktree {wt} has uncommitted changes; the script runs "
"`git reset --hard` and `git clean -fd` there after each run. "
"Commit or stash first, or use a clean detached worktree.",
file=sys.stderr,
)
return 2
pd = Path(args.plugin_dir).resolve()
out = Path(args.out)
out.parent.mkdir(parents=True, exist_ok=True)
results = json.loads(out.read_text()) if out.exists() else []
done = {(r["task"], r["arm"], r.get("rep")) for r in results}
suite = TASKS_MULTI if args.suite == "multi" else TASKS
for rep in range(args.reps):
for task in suite:
# interleave arms so prompt-cache warmth cannot favour one side
for arm in ("steered", "control") if rep % 2 == 0 else ("control", "steered"):
if (task[0], arm, rep) in done:
continue
rec = run_one(task, arm, wt, pd, args.model)
rec["rep"] = rep
results.append(rec)
out.write_text(json.dumps(results, indent=2))
print(
f" {task[0]:22s} {arm:8s} rep{rep} "
f"${rec.get('cost_usd', 0) or 0:.4f} turns={rec.get('num_turns')} "
f"calls={rec.get('tool_calls')} ok={rec.get('correct')}",
flush=True,
)
dirty = subprocess.run(
["git", "status", "--porcelain"], cwd=wt, capture_output=True, text=True
).stdout.strip()
if dirty:
print(f" !! worktree dirtied, resetting: {dirty[:120]}", flush=True)
subprocess.run(["git", "reset", "--hard", "-q"], cwd=wt)
subprocess.run(["git", "clean", "-qfd"], cwd=wt)
print(f"\nwrote {out}")
return 0
if __name__ == "__main__":
sys.exit(main())