fix(agent): classify 429 'overloaded' bodies as overloaded, not rate_limit (#14038)

[teknium1]

Summary

A provider returning HTTP 429 with an "overloaded" body no longer exhausts the credential pool. The 429 status path classified every 429 as rate_limit with should_rotate_credential=True, so a busy (but valid) provider got its single key marked exhausted after two errors — locking out users with one key, who have nothing to rotate to.

Closes #14038.

Root cause

_classify_by_status returned rate_limit unconditionally for HTTP 429 — it never inspected the body. Z.AI / Zhipu reuse 429 for server-wide overload. The credential is valid; the server is just busy, so rotation is both useless and destructive (it burns the pool while the endpoint is still busy).

This is the credential-pool-exhaustion vector the issue describes ("after 2 errors the single API key is marked exhausted"). The companion message-only path was fixed separately in #52890 (salvage of #14261 by @ms-alan); this PR covers the 429-body vector that #52890 didn't touch. Approach informed by the closed #15414 by @briandevans, which first identified this half.

Changes

• agent/error_classifier.py: 429 path now matches the body against _OVERLOADED_PATTERNS and routes overload language to FailoverReason.overloaded (retryable, no credential rotation), matching the existing 503/529 and message-only paths. Genuine rate limits (no overload language) still rotate.
• Extracted the inline overloaded tuple fix(agent): classify message-only 'overloaded' as server overload (salvage of #14261 by @ms-alan) #52890 added into a shared _OVERLOADED_PATTERNS module constant so the status-code and message paths use one list.
• Tests: added a 429-overloaded regression test and a guard that a genuine 429 rate limit still rotates.

Validation

Error	Before	After
HTTP 429 + "temporarily overloaded"	rate_limit + rotate (exhausts pool)	overloaded · backoff+retry · no rotation
HTTP 429 "rate limit exceeded"	rate_limit + rotate	rate_limit + rotate (unchanged)
HTTP 503/529 overloaded	overloaded	overloaded (unchanged)
message-only overloaded	overloaded (#52890)	overloaded (unchanged)

scripts/run_tests.sh tests/agent/test_error_classifier.py → 164 passed. tests/agent/test_credential_pool_routing.py → 10 passed.

Infographic

[github-actions]

🔎 Lint report: hermes/hermes-97a90860 vs origin/main

ruff

Total: 0 on HEAD, 0 on base (➖ 0)

🆕 New issues: none

✅ Fixed issues: none

Unchanged: 0 pre-existing issues carried over.

ty (type checker)

Total: 11502 on HEAD, 11500 on base (🆕 +2)

🆕 New issues (2):

Rule	Count
unresolved-attribute	2

First entries

tests/run_agent/test_credits_notices_toggle.py:76: [unresolved-attribute] unresolved-attribute: Unresolved attribute `_credits_session_start_micros` on type `AIAgent`
run_agent.py:3014: [unresolved-attribute] unresolved-attribute: Object of type `Self@get_credits_spent_micros` has no attribute `_credits_session_start_micros`

✅ Fixed issues (1):

Rule	Count
invalid-assignment	1

First entries

tests/run_agent/test_credits_notices_toggle.py:76: [invalid-assignment] invalid-assignment: Object of type `None` is not assignable to attribute `_credits_session_start_micros` of type `int`

Unchanged: 6049 pre-existing issues carried over.

Diagnostics are surfaced as warnings — this check never fails the build.