feat(retry): full-jitter backoff + Retry-After honoring + max_retries cap (fixes #5165)

[leavedrop]

The issue author asked: "Would maintainers be open to a small PR for this retry policy improvement? I'm happy to contribute a focused PR if this direction makes sense." This is that PR.

Both simple_send_with_retries (aider/models.py) and the streaming send loop (aider/coders/base_coder.py) used identical deterministic retry_delay *= 2 backoff with no jitter, no Retry-After handling, and no explicit attempt cap. This PR fixes all four points raised in #5165, via a small shared helper module.

What changed — 4 improvements

• New aider/retry_utils.py (117 lines) — compute_retry_sleep(attempt, base_delay, cap, retry_after=None) implements the "Full Jitter" recipe (sleep is uniform(0, min(base * 2**(N-1), cap))). parse_retry_after(exception) walks common SDK header attribute paths plus one __cause__ hop, accepting both integer-seconds and HTTP-date forms per RFC 7231 §7.1.3. MAX_RETRIES = 8 module-level cap.

• aider/models.py (+12/-7) — simple_send_with_retries now uses compute_retry_sleep + parse_retry_after + MAX_RETRIES. Previous logic stopped retrying only when retry_delay > RETRY_TIMEOUT (60s), which allowed unbounded total wait if the doubling hit the ceiling early; now there's an explicit attempt counter.

• aider/coders/base_coder.py (+14/-7) — same change applied to the streaming send loop. The two retry blocks were near-identical; both now share the helper.

• Retry log line upgraded (both call sites) — Retrying in 12.3s (attempt 3/8, RateLimitError)... instead of just Retrying in 12.3 seconds.... Includes the actual jittered sleep (not the cap), attempt counter, and exception class so operators can tell which provider error is recurring.

Tests

tests/basic/test_retry_utils.py — 16 unit tests across three classes:

• TestComputeRetrySleep (6): bound stays in [0, cap] across 1000 samples; bound grows exponentially with attempt until clamped at cap; clamping holds when 2**attempt overflows; Retry-After value overrides jitter exactly; Retry-After capped at cap*2 when hostile; zero/negative Retry-After falls through to jitter.
• TestParseRetryAfter (9): integer-seconds from .response.headers; integer-seconds from top-level .headers; case-insensitive lookup; HTTP-date form returns seconds-until-date; past HTTP-date returns None; missing header returns None; no response/headers returns None; unparseable value returns None; walks __cause__ chain one hop.
• TestMaxRetriesConstant (1): smoke check on the default.

Validation

$ python -m py_compile aider/retry_utils.py aider/models.py aider/coders/base_coder.py
$ python -m pytest tests/basic/test_retry_utils.py -v
============================= 16 passed in 0.47s ==============================

Backward compatibility

• No CLI flag changes, no config changes — MAX_RETRIES and base delay (0.125s) are module-level constants matching the prior implicit defaults.
• RETRY_TIMEOUT = 60 constant retained as the jitter cap; only its role changed (was the only ceiling; now the per-sleep cap with MAX_RETRIES as the second ceiling).
• When the provider does not send Retry-After, behavior is identical to the old loop except for jitter randomization — no change to which errors are retryable (still gated by ex_info.retry).

Fixes #5165

[CLAassistant]

All committers have signed the CLA.