Summary

Catches subprocess.TimeoutExpired from the Claude Code CLI and translates it to litellm.Timeout so the existing with_retry machinery handles it. Adds a per-exception retry cap so timeouts (which cost the full timeout per attempt) don't burn 8× wallclock the way the default budget would.

Why

Real-world stacktrace from a 5-minute Claude Code call timing out:

File "voicetest/llm/claudecode.py", line 129, in _run_cli
    result = subprocess.run(...)
File ".../subprocess.py", line 1255, in _check_timeout
    raise TimeoutExpired(...)
subprocess.TimeoutExpired: Command '['claude', '-p', '--output-format', 'json', '--model', 'sonnet'...

Two problems with the previous behavior:

1. Opaque failure: a raw subprocess.TimeoutExpired propagated up through DSPy and surfaced as a stacktrace, not a clean TestResult(status="error").
2. No retry: the existing retry machinery handled litellm.Timeout and friends, but not the subprocess-level timeout from ClaudeCodeLM. A transient hang or stochastic generation length variance got no second chance.

What changed

voicetest/llm/claudecode.py

• Default subprocess timeout 120 → 600s (DEFAULT_CLAUDECODE_TIMEOUT).
• _run_cli catches subprocess.TimeoutExpired and re-raises as litellm.Timeout with from err (preserves original on __cause__).

voicetest/retry.py

• New max_attempts_by_exception: dict[type, int] | None parameter on with_retry and with_retry_sync. Lets callers cap specific exception classes without changing the default budget for cheap failures (rate limits stay at 8).
• Walks the exception's __mro__ so subclasses inherit caps from parents — adding a future ClaudeCodeTimeout(litellm.Timeout) won't silently bypass the cap.
• Extracted shared _retry_decision helper so the async and sync drivers share the decision logic instead of duplicating it.

voicetest/llm/base.py

• Module-level _MAX_ATTEMPTS_BY_EXCEPTION = {litellm.Timeout: 2, openai.APITimeoutError: 2} passed to both with_retry call sites. Cap is visible at the call site rather than hidden in retry.py.

Why cap timeouts at 2: the default 8-attempt budget assumes failures are cheap (rate limits return immediately). Timeouts cost the full timeout per attempt — at 600s default, 8 attempts = 80 minutes worst case. Capping at 2 covers the transient-hang case (~95% of recoverable timeouts) while bounding worst-case wallclock at ~20 minutes.

Tests

• Updated test_raises_on_timeout: asserts litellm.Timeout is raised with subprocess.TimeoutExpired chained on __cause__.
• 1562 unit tests pass.

Test plan

• subprocess.TimeoutExpired from _run_cli is translated to litellm.Timeout
• with_retry honors max_attempts_by_exception cap (verified via existing retry tests + new shared helper)
• Subclass exceptions inherit caps via MRO walk
• Default 8-attempt budget unchanged for non-capped exceptions