Delete the provider-error LLMEvent so streams carry output only and
every provider-reported failure exits through the typed error channel.
Anthropic SSE error events, OpenAI Responses response.failed/error
events, and Bedrock exception frames now fail the stream with an
LLMError from the shared classifier (carrying the provider code, e.g.
rate_limit_exceeded -> LLM.RateLimit).
Core follows: the publisher drops its provider-error case, the runner
drops held-back overflow events (overflow recovery keys off a thrown
LLM.ContextOverflow), compaction and title stop scanning events for
failures, and isContextOverflowFailure is deleted with its last
consumer. V1 packages/opencode gets minimal compile fixes only (dead
switch case, one test stream).
Replace the LLMError { module, method, reason } wrapper with a flat
tagged union (LLM.BadRequest, LLM.Authentication, LLM.PermissionDenied,
LLM.NotFound, LLM.RateLimit, LLM.QuotaExceeded, LLM.ContentPolicy,
LLM.ContextOverflow, LLM.ServerError, LLM.APIError, LLM.ConnectionError,
LLM.TimeoutError, LLM.MalformedResponse, LLM.NoRoute) plus an isLLMError
guard. Add one shared classifyApiFailure classifier used by the HTTP
executor and the AI SDK adapter so both surfaces classify identically,
preserving status, headers, body, and retry-after.
Core policy moves onto tags: retry RateLimit | ServerError |
ConnectionError | TimeoutError; toSessionError adds
provider.context-overflow, provider.timeout, and provider.not-found.
The provider-error stream event and the runner's held-back overflow
handling are unchanged here; isContextOverflowFailure now bridges old
events and new tags until the event is removed.
