Enrich retry log messages with task/sample/model context

[sjawhar]

Summary

When inspect retries failed model requests, log messages like Retrying request to /responses in 0.396765 seconds lack context about which task, sample, and provider triggered the retry. In a concurrent runner processing many samples across many tasks, this makes debugging difficult.

This PR enriches all retry log messages with a compact context prefix and error summary:

[EeDA74nwfs3uirgyprfa4b hello_world/1/1 openai/gpt-4o-mini] Retrying request to /chat/completions in 1.000000 seconds

[Abc12xY mmlu/42/1 openai/gpt-4o] -> openai/gpt-4o retry 2 (retrying in 6 seconds) [RateLimitError 429 rate_limit_exceeded]

Format: [{sample_uuid} {task}/{sample_id}/{epoch} {model}]

What changed

New helpers in src/inspect_ai/_util/retry.py:

• sample_context_prefix() — builds the compact prefix from sample_active() ContextVar
• retry_error_summary() — extracts exception type/status/code without leaking message content
• SampleContextFilter — logging.Filter for SDK loggers, adds both inline prefix and structured fields
• install_sample_context_logging() — attaches filter at eval startup

Enriched existing loggers:

• log_model_retry() in model/_model.py — prefix + error summary
• log_httpx_retry_attempt() in _util/httpx.py — prefix

Wired into startup:

• init_eval_context() in _eval/context.py calls install_sample_context_logging()

Evidence of working

E2E: Real inspect eval against mock 429 server

Ran inspect eval examples/hello_world.py against a local mock server returning 429s:

[02/13/26 20:24:32] HTTP     POST                                   hooks.py:123
                             http://localhost:8765/v1/chat/completi
                             ons "HTTP/1.0 429 Too Many Requests"
                    INFO     [EeDA74nwfs3uirgyprfa4b        _base_client.py:1693
                             hello_world/1/1
                             openai/gpt-4o-mini] Retrying
                             request to /chat/completions
                             in 1.000000 seconds

The prefix [EeDA74nwfs3uirgyprfa4b hello_world/1/1 openai/gpt-4o-mini] appears on the SDK's own retry message — confirming the SampleContextFilter on openai._base_client works.

Structured JSON logging

When using a JSON log formatter (e.g. python-json-logger), the structured fields appear as top-level keys:

{
  "message": "[Abc12xY mmlu/42/1 openai/gpt-4o] Retrying request to /responses in 0.396765 seconds",
  "name": "openai._base_client",
  "levelname": "INFO",
  "sample_uuid": "Abc12xY",
  "sample_task": "mmlu",
  "sample_id": 42,
  "sample_epoch": 1,
  "sample_model": "openai/gpt-4o"
}

Unit tests: 24 passing

tests/util/test_retry_logging.py  24 passed
tests/test_retry.py               12 passed, 1 skipped
tests/test_retry_on_error.py       5 passed

Review-driven fixes

Code review caught three issues, all fixed with regression tests added before the fix:

Issue	Fix
Filter on wrong logger — installed on openai but SDK logs from openai._base_client; parent logger filters don't run for child records during propagation	Target openai._base_client directly
TypeError on non-string .code — getattr(ex, "code") can return int, crashing ' '.join()	str(raw_code) cast
% in prefix breaks formatting — mutating record.msg with % chars would corrupt msg % args	Call record.getMessage() first, then set resolved msg and clear args

Linear: ENG-594