[BUG] Race condition in security-guidance's _agentic_review_with_race causes redundant security analysis

### Preflight Checklist

- [x] I have searched [existing issues](https://github.com/anthropics/claude-code/issues?q=is%3Aissue%20state%3Aopen%20label%3Abug) and this hasn't been reported yet
- [x] This is a single bug report (please file separate reports for different bugs)
- [x] I am using the latest version of Claude Code

### What's Wrong?

In `plugins/security-guidance/hooks/security_reminder_hook.py`, the function `_agentic_review_with_race` runs two threads in a race: an agentic reviewer thread (`_agentic`) and a delayed fallback reviewer thread (`_fallback`). They communicate their completion using a shared queue `q` of size 1.

If the agentic thread finishes first, it puts its result into `q`. The main thread immediately retrieves this result using `q.get()`, which empties the queue, and then returns.

When the fallback thread wakes up after its delay (`delay_s`), it checks if the agentic thread finished by calling `if not q.empty(): return`. Because the main thread already consumed the item from `q`, the queue is empty. The fallback thread incorrectly assumes the agentic thread hasn't finished, so it proceeds to run the expensive `analyze_code_security` function anyway. This results in redundant LLM/security analysis calls and wasted resources — silently, with no error or indication anything went wrong.

### What Should Happen?

The fallback thread should correctly detect that the agentic reviewer already finished (via a dedicated completion signal, not queue emptiness) and skip re-running `analyze_code_security` entirely when the agentic reviewer already succeeded within the delay window.

### Error Messages/Logs

```shell
None — this is a silent correctness/efficiency bug, not a crash. That's part of what makes it easy to miss: analyze_code_security simply runs twice with no warning.
```

### Steps to Reproduce

This was found via static code review, not live reproduction — no special setup is needed to see the bug, just reading the logic:

1. Open `plugins/security-guidance/hooks/security_reminder_hook.py`.
2. Locate `_agentic_review_with_race`. Note `q = queue.Queue(maxsize=1)`.
3. Trace what happens when `_agentic` finishes before the `delay_s` timeout: it does `q.put_nowait(("agentic", r))`, and the main thread's `winner, (g, v, m) = q.get()` immediately drains that item.
4. Trace `_fallback()`: after `time.sleep(delay_s)`, it checks `if not q.empty(): return`. Since the queue was already drained in step 3, this check is always False in this scenario — `_fallback` proceeds to call `analyze_code_security` even though `_agentic` already completed successfully.

To confirm at runtime: add a log line at the start of `_fallback()`'s `analyze_code_security` call, make a commit where the agentic reviewer typically finishes well within `SG_AGENTIC_RACE_DELAY_S` (default 180s), and observe the fallback's expensive path still fires.

### Claude Model

None

### Is this a regression?

I don't know

### Last Working Version

_No response_

### Claude Code Version

N/A — found via static code review of the plugin source (plugins/security-guidance/hooks/security_reminder_hook.py on the main branch), not a live reproduction tied to a specific runtime/OS/terminal combination. The bug is in pure Python thread-synchronization logic, independent of platform.

### Platform

Anthropic API

### Operating System

macOS

### Terminal/Shell

PyCharm terminal

### Additional Information

Suggested fix — use a dedicated threading.Event instead of queue emptiness as the completion signal:

def _agentic_review_with_race(
repo_root: str,
diff_files: List[Tuple[str, str]],
rel_touched: List[str],
previous_findings: List[Dict[str, Any]],
) -> Tuple[Optional[str], List[Dict[str, Any]], Dict[str, Any]]:
"""Race the agentic reviewer against a delayed single-shot fallback."""
import queue as _queue
import threading as _th
import time as _t

if os.environ.get("SG_AGENTIC_NO_RACE") == "1":
return agentic_review(repo_root, diff_files, rel_touched)

delay_s = int(os.environ.get("SG_AGENTIC_RACE_DELAY_S", "180"))
q: "_queue.Queue[Tuple[str, Any]]" = _queue.Queue(maxsize=1)
fallback_started = _th.Event()
agentic_finished = _th.Event()

def _agentic() -> None:
try:
r = agentic_review(repo_root, diff_files, rel_touched)
except Exception as e:
r = (None, [], {"agentic_fallback": f"race_crash:{type(e).__name__}"})
try:
q.put_nowait(("agentic", r))
except _queue.Full:
pass
finally:
agentic_finished.set()

def _fallback() -> None:
_t.sleep(delay_s)
if agentic_finished.is_set():
return # agentic finished within the delay — never start fallback
fallback_started.set()
try:
g, v = analyze_code_security(
diff_files, is_diff=True, previous_findings=previous_findings
)
except Exception as e:
g, v = None, []
try:
q.put_nowait(("fallback", (g, v, {"agentic": False})))
except _queue.Full:
pass

_th.Thread(target=_agentic, daemon=True).start()
_th.Thread(target=_fallback, daemon=True).start()

winner, (g, v, m) = q.get()
m = dict(m)
m["race_winner"] = 1 if winner == "agentic" else 2
m["race_delay_s"] = delay_s
m["race_started"] = 1 if fallback_started.is_set() else 0
return g, v, m

Confidence: high — directly visible in the thread-synchronization logic; confirmed by tracing queue state through the code, no runtime logs needed.

Found via an AI code-investigation tool (Meeba Brain) while testing its bug-discovery accuracy on real open-source repos — verified by hand against the source before filing.

Start in plugins/security-guidance/hooks/security_reminder_hook.py at _agentic_review_with_race and trace the agentic and fallback thread paths, including the queue and completion checks. Review any existing security-guidance hook tests before making changes. Done means an agentic review that finishes before the delay does not trigger a second analyze_code_security call, while the fallback still works when needed.

由索引模型根据 Issue 内容生成。