fix(bench): write config.json atomically to prevent silent 404s

[regdocs]

Problem

A bench that clearly exists on disk was returning 404 for every request:

agent.exceptions.BenchNotExistsException: Bench bench-9052-000094-f16j does not exist
  File ".../agent/bench.py", line 61, in __init__
    self.docker_image = self.bench_config.get("docker_image")
  ...
json.decoder.JSONDecodeError: Expecting value: line 1 column 1 (char 0)

On the server, config.json was 0 bytes while a valid config.json.bak sat right next to it.

Root cause

set_bench_config claimed to avoid partial writes by writing to a temp file first, but its final step used shutil.copy2(temp, config.json). copy2 opens the destination with open(dst, "wb"), which truncates config.json to 0 bytes and then streams into it. If the process dies in that window (OOM kill, worker/supervisor restart, deploy), you're left with a 0-byte config.json + a valid .bak.

Bench.__init__ then raises on json.load (at bench.py:61, before the existence guard), and Server.benches swallowed the exception via suppress(Exception) — so the bench silently disappeared from the dict and get_bench raised BenchNotExistsException → 404, with no hint of the real cause.

The temp file lived in /tmp (a different filesystem), which is why copy2 was used instead of an atomic os.replace — the latter fails with EXDEV across filesystems.

Fix

• set_bench_config: write the temp file in the bench directory via tempfile.mkstemp(dir=self.directory), fsync, then os.replace onto config.json — an atomic same-filesystem rename with no truncation window. Removes the copy2/.bak dance.
• bench_init: the initial config.json write for a new bench used open("w") (truncates immediately) — made atomic the same way, so an interrupted New Bench job can't leave a 0-byte stub.
• Server.benches: replace suppress(Exception) with a try/except that logger.exceptions the skipped bench, so a corrupt config is diagnosable from logs instead of an opaque 404.

Testing

• Both files compile.
• Smoke-tested the atomic-write logic standalone: correct content, clean overwrite, no zero-byte window, no leftover temp files.

Recovering affected benches

For any bench already in this state, the good config is in the backup:

cd /home/frappe/benches/<bench-name>
install -o frappe -g frappe -m 600 config.json.bak config.json

[mergify]

Tick the box to add this pull request to the merge queue (same as @mergifyio queue).

• Queue this pull request

[greptile-apps]

Confidence Score: 4/5

The atomic-write fix is correct and removes the truncation window; the only rough edge is log verbosity for broken benches.

The core changes are solid. The logger.exception call inside an uncached property is the one thing worth a second look in production — a single bad bench directory will generate a log line on every property access across all callers.

agent/server.py — specifically the benches property and its many call sites.

_{Reviews (1): Last reviewed commit: "fix(bench): write config.json atomically..." | Re-trigger Greptile}

[greptile-apps]

Exception logged on every benches access for a broken bench

self.benches is an uncached @property called from ~15 sites (including hot paths like dump() and get_bench). A permanently corrupt config.json will fire logger.exception on every one of those accesses, flooding the log until the bench is manually repaired.