The WS v2 passthrough relay only read output_tokens_details.image_tokens,
so image generation over WebSocket billed image input tokens at the text
input rate and ignored the hosted response.tool_usage.image_gen breakdown
that the HTTP Responses path already merges (extractOpenAIUsageFromJSONBytes
+ mergeHostedImageGenToolUsage).
The relay now reads image input tokens from input_tokens_details or
prompt_tokens_details, fills missing image input/output counters from
response.tool_usage.image_gen (usage values keep precedence), and carries
ImageInputTokens per turn and in the session aggregate into the
OpenAIForwardResult used for billing.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
A Responses refusal arrives as a {"type":"refusal","refusal":...} content
part (streamed as response.refusal.delta), not as output_text. The
Responses → Chat Completions converters only read output_text, and
ResponsesContentPart had no refusal field, so Chat clients received an
empty assistant message when the model refused.
Map refusal parts to choices[0].message.refusal, response.refusal.delta
to delta.refusal, and let BufferedResponseAccumulator keep refusal deltas
so non-streaming replies rebuilt from deltas keep them too.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
The buffered Anthropic aggregators index finalResp.Content with the
upstream content_block_delta index after checking only idx < len. A
malformed or third-party Anthropic-compatible upstream that sends a
negative index makes the request goroutine panic, so the client gets no
response and the forward never returns a result to record usage.
Add the lower bound in all four buffered aggregators (Chat Completions
and Responses, on both the Anthropic gateway and the OpenAI native
Anthropic paths); the out-of-range delta is skipped like an index past
the end.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Legacy Chat function calling has no call IDs: the assistant turn carries
function_call and the role=function result refers to it only by name.
ChatCompletionsToResponses ignored the assistant function_call and used
the function name as the result's call_id, so the Responses upstream got
a function_call_output with no matching function_call and rejected the
replayed history.
Emit a function_call item for each legacy assistant call and give the
following same-name result the same call_id. IDs follow message order
(call_legacy_N), so the same history converts to the same input across
turns, and skip IDs already used by tool_calls. A result with no
preceding call keeps the previous name-based call_id.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
chatMessageToResponsesItems only mapped "system" explicitly and sent every
other unknown role through the user converter, so Chat Completions
developer messages reached the Responses upstream as user turns and lost
their instruction priority.
The Responses API accepts role "developer" in input messages; convert it
like a system message and keep the original role. Downstream converters
already understand it (ResponsesToAnthropicRequest folds developer into
the system prompt, the Codex transform forwards it unchanged).
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Chat Completions names a forced function as
{"type":"function","function":{"name":"X"}}, while the Responses API
expects {"type":"function","name":"X"}. ChatCompletionsToResponses passed
the Chat shape through unchanged, so OpenAI API-key accounts on
/v1/chat/completions forwarded the nested shape to /v1/responses and the
upstream rejected the request.
Flatten the named Chat choice; strings and objects already in the
Responses shape are kept as-is.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
The Responses WebSocket handler reads the API key auth snapshot once at
connect and reuses it for every turn: RecordUsage takes the group pricing
from that snapshot, and the profit gate reinstalled in BeforeTurn takes
its selling price from the connect-time ctxkey.Group. A group price
change therefore never reaches an open connection until the client
reconnects, while HTTP requests pick it up once the auth cache entry is
invalidated.
Each later turn now re-reads the key through APIKeyService.GetByKey (the
auth cache the middleware uses) in BeforeTurn. When the key ID, group ID,
group platform and subscription type are unchanged, the fresh group is
swapped into ctxkey.Group for the turn's profit gate and passed to that
turn's RecordUsage. A failed lookup or a key moved to another group keeps
the connection snapshot.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Adds two regression cases mirroring the #5009 OpenAI ones for Grok
platform accounts (issue #7774):
- TestOpenAIResponsesEmptyCompletedFailsOverGrok: an empty
response.completed on the forwardGrokResponses streaming path must
produce UpstreamFailoverError with an empty client body. Verified to
FAIL on main without the guard extension and PASS with it.
- TestOpenAIResponsesEmptyCompletedWithOutputSucceedsGrok: normal Grok
streams with semantic output and usage are untouched.
Fixes#7774
Issue #5009 added a failover guard for OpenAI Responses streams that
terminate with an empty response.completed (no output, no usage, no
error), but the guard is gated on PlatformOpenAI. Grok platform
accounts reach the same handler (forwardGrokResponses ->
handleStreamingResponseWithReasoning) and exhibit the same upstream
silent-refusal shape, so their empty completions are relayed to the
client as successful empty turns: Responses SDK clients map the empty
completed to finishReason stop and accept the turn with no retry.
Extend the guard to PlatformGrok, matching the unconditional guard on
the passthrough path (openai_gateway_passthrough.go).
Fixes#7774
- R1-1: an explicit upstream "unavailable" answer now fences the organization
for 15 minutes (CLAUDE_RESET_UPSTREAM_UNAVAILABLE) instead of 24h; ambiguous
outcomes keep the 24h fence. Frontend tells the operator to retry later.
- R1-2: rotate the idempotency key after definite pre-claim refusals (busy,
not available, unresolved, upstream unavailable); map IDEMPOTENCY_IN_PROGRESS
and IDEMPOTENCY_RETRY_BACKOFF to clear zh/en messages.
- R1-3: redeem reason and cleared windows pass only through explicit
allowlists, so upstream values can never echo a grant ID to clients.
Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
POST /admin/accounts/:id/claude/reset-credits/redeem (Idempotency-Key
required, same middleware chain as the Codex reset-quota route). The server
re-queries eligibility and claims only the upstream next_grant_id when it is
redeemable; grant/org IDs never reach the client. Claims are serialized by
account and organization Redis leases, and a durable per-organization fence
in the idempotency table blocks further redemption after an unconfirmed
outcome (settles after 24h). Ported from upstream PR #7591.
Co-authored-by: korkin25 <eugeny.kuzakov@gmail.com>
Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
- Drop selection_token (unsalted hash of a short grant ID, reversible) and
claudeGrantSelection from the read-only DTO; assert it never serializes.
- query() returns only the projected result; no raw upstream block escapes.
- Clear a cooldown_until that is already past, using the same `now` as the
redeemable check; the cell also shows the cooldown only when in the future.
- Invalidate in-flight queries when account id, platform or type changes.
- Sort the earliest-expiry fallback by parsed timestamp, invalid/missing last.
Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
- only consult account mappings after primary/channel candidates fail to price (or composite)
- source mappings from the scheduler snapshot (same pool as account selection), no direct DB query and no per-model cache
- ungrouped API keys use the scheduler's ungrouped pool
- no negative caching: new mappings take effect on the next request
Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
Wesley Liddick