Bug: `/too many tokens/i` overflow regex misclassifies rate limit errors as context overflow

[mrm007]

Problem

The regex /too many tokens/i in OVERFLOW_PATTERNS (packages/opencode/src/provider/error.ts, line 22) is too broad. It matches rate limit / daily quota errors that contain "too many tokens" but are not context overflow errors.

Example: AWS Bedrock returns this when the daily token quota is exhausted:

Too many tokens per day, please wait before trying again.

This matches /too many tokens/i → classified as ContextOverflowError → triggers compaction → compaction also fails (same quota error) → user is stuck. The error never reaches the retry loop or plugin event hooks (session.status with type: "retry").

Cerebras has the same issue: "Tokens per minute limit exceeded - too many tokens processed" (see #7463).

Expected Behavior

Daily/minute token quota errors should be classified as retryable API errors (or at least non-overflow errors), so that:

1. The retry loop can handle them with backoff
2. Plugins listening for session.status retry events can trigger fallback logic
3. Compaction is NOT triggered (context size isn't the problem)

Root Cause

The classification chain:

1. ProviderError.parseAPICallError() calls isOverflow(message)
2. isOverflow() checks OVERFLOW_PATTERNS — /too many tokens/i matches
3. Returns { type: "context_overflow" } → becomes ContextOverflowError
4. SessionRetry.retryable() line 63: ContextOverflowError.isInstance(error) → returns undefined (not retryable)
5. SessionProcessor.process() sets error on message, emits session.error, never emits session.status with type: "retry"
6. Processor returns "compact" → compaction triggered on a quota error

Suggested Fix

Make the regex exclude temporal qualifiers:

// Before
/too many tokens/i, // Generic fallback

// After — exclude "per day", "per minute", "per hour", "processed"
/too many tokens(?! per (day|minute|hour|second))(?! processed)/i,

Or add a pre-check for rate limit patterns before checking overflow patterns, since rate limits should take priority.

Affected Error Messages

Error Message	Source	Should Be
"Too many tokens per day, please wait before trying again."	AWS Bedrock	Rate limit (retryable)
"Tokens per minute limit exceeded - too many tokens processed"	Cerebras	Rate limit (retryable)
"prompt is too long: 285744 tokens > 200000 maximum"	Anthropic/Bedrock	Context overflow ✅ (correct)

Environment

• OpenCode version: 1.1.53+
• Provider: amazon-bedrock (Claude Opus 4.6)
• OS: macOS

Related Issues

• Beta header context-1m-2025-08-07 not sent for Claude Opus 4.6, causing 200k hard limit #12507 — Beta header not sent for Opus 4.6 (separate issue causing 200K limit)
• Prompt too long when using Claude Opus 4.6 (Global) Amazon Bedrock #12688 — Prompt too long on Bedrock Opus 4.6
• [FEATURE]: support anthropic beta 1m token on bedrock #11267 — Feature request: 1M token support on Bedrock
• When using the cerebras api i am frequently getting: "Tokens per minute limit exceeded - too many tokens processed" #7463 — Cerebras "too many tokens processed" (also affected by this regex)
• [FEATURE]: Native Model Fallback / Failover Support #7602 — Native model fallback support
• Quota Limit Exceeded Error is not handled properly. #3525 — Quota limit exceeded error handling (partially fixed, this is a regression path)

[github-actions]

This issue might be a duplicate of or closely related to existing issues. Please check:

• When using the cerebras api i am frequently getting: "Tokens per minute limit exceeded - too many tokens processed" #7463: Cerebras "Tokens per minute limit exceeded - too many tokens processed" (explicitly mentioned, same regex misclassification)
• Quota Limit Exceeded Error is not handled properly. #3525: "Quota Limit Exceeded Error is not handled properly" (older quota handling issue)
• Beta header context-1m-2025-08-07 not sent for Claude Opus 4.6, causing 200k hard limit #12507: "Beta header not sent for Claude Opus 4.6, causing 200k hard limit" (related AWS Bedrock context)
• Prompt too long when using Claude Opus 4.6 (Global) Amazon Bedrock #12688: "Prompt too long when using Claude Opus 4.6 (Global) Amazon Bedrock" (related context)
• Implement exponential back-off when hitting rate limits #1712: "Implement exponential back-off when hitting rate limits" (retry handling this bug prevents)
• LLM provider SDK lacks retry logic for transient API errors #11705: "LLM provider SDK lacks retry logic for transient API errors" (related retry logic needed)
• [FEATURE]: Native Model Fallback / Failover Support #7602: "Native Model Fallback / Failover Support" (mentioned in this issue for fallback logic)

The core fix needed here (distinguishing rate limit errors from context overflow) will likely help resolve or unblock several of these related issues.

[rekram1-node]

ill just drop the generic fallbacks then