Dashboard › opencode-lore › Distillation
5bd201a9-4938-44a9-a6a5-4da67afbe959["lore_tm_v1_Pp5hLD3_XuJW38pLvRIZpYklXUIP7-_xhaLBOh9OYts","lore_tm_v1_vXroIRrrms_S6-JB1ajgYSKb8NLTTYD_dRwuXObU0i4","lore_tm_v1_cbD9KoMfhORfohrcBE8qKNKnF08Nwoq7pi_6UBgbOpk","lore_tm_v1_Oecm-qcvVrP2bic2OfGChEidutWVL_FNXA6ctZjWfMM","lore_tm_v1_UFrwi_TvAzrh_rAaTw183pc3qZbldXpPC0T2IHNrRn0","lore_tm_v1_dgiEQfYYXn3xx6cD-lHaF4R_pEw68L630zL80EmNF0Y"]
Date: Sep 16, 2026
0 files changed, 0 insertions(+), 0 deletions(-) with SHA-256 e3b0c44298fc1c149afbf4c8996fb92427ae41e4649b934ca495991b7852b855.commit_id does not exist in commit_id ++ " " ++ description.first_line() ++ "\n"; the tool suggested commit.1512d486 with description fix(responses): unify provider failure handling, authored by ben@byk.im at 2026-09-16 11:49:37, on top of hidden empty commit e19ee1c7; associated operation IDs were 58a2b9a6dcc9 and 9c44ffd239ad.packages/gateway/src/stream/openai-responses.ts around lines 12255-12262: when content parsing fails before usage validation, parsing may be retried on only usage and usageMetadata to retain numeric/cache consistency checks; invalid usage still throws, and the projection is βnever returned as success.βparseResponsesNonStreamEnvelope() accepts statuses completed and incomplete for completion-shape validation, but accumulateNonStreamResponse() throws ResponsesTerminalError whenever the parsed status is not completed.accumulateResponsesNonStreamJSON() excludes item_reference entries from replayable output, rejects duplicate or missing response-item identities with malformed Responses response item identity, and rejects duplicate or missing tool identities with malformed Responses response tool identity.incomplete_details.reason === "content_filter" to stop reason content_filter; other incomplete responses map to max_tokens; a response containing tool_use changes end_turn to tool_use.input_tokens_details and falls back to prompt_tokens_details; it reports cached_tokens as cacheReadInputTokens and cache_write_tokens as cacheCreationInputTokens.accountUnsuccessfulResponse() records validated nonzero provider usage through accountConversationUsage(), optionally marks state dirty, adds GenAI usage attributes, then sets span status code 2 with message upstream response did not complete and ends the span.captureUnsuccessfulResponses() converts a ResponsesTerminalError with status incomplete into { response, successful: false }; other terminal statuses yield undefined. The non-stream path returns HTTP 502 for no captured response or when an unsuccessful incomplete response contains recall tool use; otherwise it emits the incomplete response without treating it as successful.preserveUpstreamErrorResponse() rather than running 4xx/429 bodies through an SSE validator that could turn them into status 200 or a synthetic stream failure.vertex β anthropic), allowing same-wire Anthropic clients to stream unchanged; cross-protocol buffering/re-emission is reserved for genuinely different wire protocols.sanitizedUpstreamErrorResponse().400 bodies containing prompt is too long, context_length_exceeded, or maximum context length; it computes overshootRatio = actual/max, defaults unparseable errors to 1.3, and escalates to compression layer 3 when the ratio is at least 1.2, otherwise layer 2.finalizeWithRecall() intercepts gateway-injected recall tool use before client delivery and supports multiple recall calls, including drill-down into t:<id> source citations; resource, progress, and time budgets normally stop the chain before its emergency ceiling.