fix(errors): classify throttle messages before token-overflow patterns; add new overflow shapes (port opencode#37848)

[teknium1]

Summary

Throttle messages that mention tokens no longer route a healthy session into the compression loop, and two previously-unclassified overflow/payload error shapes now trigger compression recovery. Port of anomalyco/opencode#37848 (and its dev-branch twin #37840).

Root cause: "too many tokens" sits in _CONTEXT_OVERFLOW_PATTERNS, and the AWS Bedrock / proxy throttle shape "Throttling error: Too many tokens, please wait before trying again." matched no rate-limit pattern — so the message-only path fell through to the overflow list and compressed on every throttle instead of backing off.

All three gaps were proven live on current main before the fix (probe output in the validation table).

Changes

• agent/error_classifier.py:
  • _RATE_LIMIT_PATTERNS += "throttling" — checked BEFORE the overflow list in the message-only path, so token-mentioning throttles now classify as rate_limit.
  • _CONTEXT_OVERFLOW_PATTERNS += "maximum allowed input length" (Together/Fireworks-style wording: "Input length N exceeds the maximum allowed input length of M tokens.").
  • _PAYLOAD_TOO_LARGE_PATTERNS += "request_too_large", "request exceeds the maximum size" (Anthropic's structured 413 type re-wrapped without a status code by aggregators/proxies).
• tests/agent/test_error_classifier.py: new TestThrottleVsOverflowDisambiguation + TestExpandedOverflowPatterns classes (positive + regression guards, including plain "Too many tokens" still classifying as overflow).

Adaptation notes: OpenCode implemented this as a global exclusions list gating isContextOverflow. Hermes already orders rate-limit checks before overflow in _classify_by_message, so the correct adaptation is one rate-limit pattern, not a new exclusion mechanism. OpenCode's "token limit exceeded" overflow pattern was deliberately NOT adopted — hermes routes "limit exceeded" through the usage-limit billing/rate-limit disambiguation on purpose. Their request_too_large → overflow mapping was adapted to hermes's payload_too_large taxonomy (same compression recovery, more precise reason).

Validation

Input (no status code)	Before	After
"Throttling error: Too many tokens, please wait before trying again."	context_overflow, compress=True	rate_limit, compress=False
"Input length 131393 exceeds the maximum allowed input length of 131040 tokens."	unknown	context_overflow, compress=True
{"error":{"type":"request_too_large","message":"Request exceeds the maximum size"}}	unknown	payload_too_large, compress=True
"Rate limit exceeded, please retry after 30 seconds."	rate_limit	rate_limit (unchanged)
"Too many tokens" (plain, Z.AI/GLM)	context_overflow	context_overflow (unchanged)

Tests: 265 passed (test_error_classifier.py + test_bedrock_integration.py), 238 passed sibling (test_bedrock_adapter.py, test_nous_rate_guard.py, test_24996_fallback_exhaustion_cooldown.py, test_413_compression.py).

Infographic

[github-actions]

૮ >ﻌ< ა ci review

_{ran on 372e94b}

all good!