[eliteprox]

Summary

• call_runner now writes every attempt to a RunnerCallRecord as it happens. Each RunnerAttempt records the payment challenge it answered (manifest_id), the signer auth_id it opened, whether the request went out with Livepeer-Payment headers (payment_sent), the runner's HTTP status, and how the attempt ended (outcome).
• The result, the stream and any error from the call all point to the same record. A caller can pass in its own record= so it can still read it after cancelling the call (for example with asyncio.wait_for).
• payment_sent and auth_ids on LivepeerGatewayError, LiveRunnerCallResult and LiveRunnerCallStream are read-only properties computed from the record. payment_sent is true once any attempt sent payment, so a gateway may switch to another runner only while it is false. auth_ids lists every signer session the call opened. The signer charges when it signs, so an auth id can appear before payment_sent is true.
• Errors that are not a LivepeerGatewayError (for example a ClientPayloadError while reading an error body) are re-raised as LivepeerGatewayError with the record attached.
• Attempt outcomes:
  • unreachable only when the connection itself failed, so the request never got to the runner;
  • timeout for a plain timeout, since the runner may already have the request;
  • also succeeded, payment_required, refused (503), http, payment (signer failure), disconnected, invalid_response and cancelled.
• LivePaymentSession.auth_id reads the signer's AuthID from the payment state, decoded once per payment response.
• open_stream now closes its HTTP session when reading an error body fails.
• LiveRunnerCallStream.session_id carries the paid challenge's manifest_id, as LiveRunnerCallResult.session_id already does.

Runner selection and failover (RunnerSelectionCursor, Scope startup, RunnerRejection) are unchanged and match main. Making them use the record, and stop failing over once a payment is sent, is planned for a follow-up PR.

Test plan

• uv run --frozen --group lint ruff check: no new findings on changed lines
• uv run --frozen --group test pytest -W error::ResourceWarning: 267 passed
• New tests run call_runner against a real aiohttp runner on a local port, with a fake signer, and cover:
  • a signer failure after a paid attempt;
  • an unpaid retry after a signer refresh, which still reports paid;
  • a refresh on the final attempt, which keeps the earlier auth ids;
  • a truncated 500 on a stream, which is wrapped and reports paid;
  • a cancelled call, whose record stays readable;
  • a refused connection (unreachable) versus a slow runner (timeout);
  • the outcomes for 503 and 400.
• livepeer_core test suite passes against this branch

🤖 Generated with Claude Code

[eliteprox]

Should we avoid storing _auth_id separately and keep it within state for downstream parsing by auth_id_from_signer_state

[eliteprox]

TYPE_CHECKING can usually be removed

[eliteprox]

Moved to eliteprox#5, which targets releases/m2. Both review comments are addressed there: the call record types moved into errors.py, so the TYPE_CHECKING import is gone, and LivePaymentSession.auth_id now reads from the stored signer state.