Transport retries replay non-idempotent POSTs (run_many, create, tables.upload) #85

RoeRetryTransport retries 5xx, 408, 429 and transport errors on every method, including POST. Only the multipart agent-run helpers opt out (x-roe-skip-retry in _dynamic_call.py).

• roe-python/src/roe/utils/transport.py

  Lines 42 to 82 in 0881424

  	def handle_request(self, request: httpx.Request) -> httpx.Response:
  	if request.headers.get(_BYPASS_HEADER):
  	return super().handle_request(request)
  	for attempt in range(self.max_retries + 1):
  	try:
  	response = super().handle_request(request)
  	except httpx.TransportError as exc:
  	if attempt >= self.max_retries:
  	raise
  	wait_time = min(2**attempt, 10)
  	logger.warning(
  	"Transport error on %s %s (%s), retrying in %ds (attempt %d/%d)",
  	request.method,
  	request.url,
  	exc,
  	wait_time,
  	attempt + 1,
  	self.max_retries,
  	)
  	time.sleep(wait_time)
  	continue
  	if (
  	not _should_retry_status(response.status_code)
  	or attempt >= self.max_retries
  	):
  	return response
  	wait_time = min(2**attempt, 10)
  	logger.warning(
  	"Roe API returned %d for %s %s, retrying in %ds (attempt %d/%d)",
  	response.status_code,
  	request.method,
  	request.url,
  	wait_time,
  	attempt + 1,
  	self.max_retries,
  	)
  	response.close()
  	time.sleep(wait_time)

• roe-python/src/roe/utils/_dynamic_call.py

  Lines 48 to 50 in 0881424

  	request_headers["x-roe-skip-retry"] = (
  	"1" # multipart POST — do not replay (aligns with TS)
  	)

So these POSTs are sent up to max_retries + 1 (default 4) times when the response is a 5xx or the read times out:

• agents.run_many (no idempotency_key parameter, unlike run / run_version)
• agents.create, agents.versions.create, policies.create
• tables.upload (multipart, but not opted out)

If the server committed the request and a gateway then returns 502, or the response read times out, each replay creates another batch of jobs / agent / policy / table. For run_many that's up to 4x the billed jobs.

Reproduced with a fake transport that always returns 503: run_many, agents.create, policies.create and tables.upload each sent 4 POSTs; agents.run sent 1.

The README documents that JSON POST bodies may be replayed, so this may be intentional, but the comment in client.py (L72-74) says retries are for "502/503/504 (idempotent methods only)", which doesn't match.

Possible fixes, happy to send a PR for whichever you prefer:

1. Don't retry POST on 5xx / read timeouts; only on connect errors (request never sent) and 429/503 with Retry-After.
2. Generate an Idempotency-Key per logical call and reuse it across retries, if the API honours it on these endpoints. Add idempotency_key to run_many.

roe-golang has the same behaviour (all POSTs, including single agent runs).

Go counterpart: roe-ai/roe-golang#55

Note: testing this end to end requires a Roe API key, so I checked it against the code and a local mock server rather than the live API.