Keep catalog membership on schedulable accounts and intersect advertised
capabilities across all active group members that map an alias. Preserve
upstream plugin, service-tier, and models-list body-limit changes.
Co-authored-by: Cursor <cursoragent@cursor.com>
Intersect advertised alias capabilities across all active group members
that map the model, not only currently schedulable accounts, so rate-limits
cannot widen image input or context windows.
Co-authored-by: Cursor <cursoragent@cursor.com>
Users only see the reasoning effort they requested. Admins still see
the requested value plus the forwarded mapped value, matching the
model column's requested vs upstream split.
The raw Chat Completions passthrough returned a nil error whenever the upstream
body ended, so an upstream that cut the SSE stream mid-generation surfaced as
HTTP 200 with usage 0/0: the client kept a half-finished answer, the request was
not counted as an SLA failure, and ops saw nothing. Observed against
opencode.ai/zen/go (deepseek-v4-pro): 11 of 1346 streams in 24h recorded a first
token but no final usage, alongside a concurrent `connection reset by peer`
against the upstream edge.
Track whether the upstream ever emitted a terminal signal — [DONE], a usage
chunk, or a finish_reason — and treat a stream that ends without one as an
upstream failure. Matching only [DONE] would misfire on compatible upstreams
that EOF right after their last chunk and would drop billing for tokens that
were actually delivered, so all three count.
- nothing written to the client yet: UpstreamFailoverError, retried on another
account with no client-visible half stream
- bytes already on the wire: typed openAIUpstreamStreamReadError, so the
handler appends an SSE error frame and marks the request an SLA failure
Client cancel/disconnect keeps the previous semantics — there the upstream read
error is indistinguishable from a truncation, so partial usage is still billed
rather than blamed on the account.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
normalizeOpenAIParallelToolCallsWithoutTools only looked at the top-level
"tools" array, but normalizeOpenAIResponsesLiteTools moves namespace tools
into an input item of type "additional_tools" and deletes the top-level key.
A Responses Lite request that carries tools therefore looks like it has none,
and the parallel_tool_calls:false that ensureOpenAIResponsesLiteParallelToolCalls
had just pinned gets deleted on the way out.
OpenAI then applies its default of true and rejects the request:
400 unsupported_value: "X-OpenAI-Internal-Codex-Responses-Lite requires
`parallel_tool_calls` to be false."
Note the field cannot simply be pinned to false unconditionally: without tools
OpenAI rejects it with "'parallel_tool_calls' is only allowed when 'tools' are
specified", so the two constraints have to be honoured together.
Reuse the same tool-detection standard the Lite path already uses by adding
openAIRequestBodyHasTools, the []byte counterpart of openAIResponsesLiteHasTools,
so both sides of the repo agree on what "has tools" means.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
The upstream names one offending index per response, but a replayed
conversation routinely carries dozens of items of the same type, each with a
status the upstream schema does not accept. Clearing a single index per round
trip needs one retry per item, so a conversation with more than
maxOpenAIResponsesRejectedFieldRetries such items exhausts the bounded budget
and the 400 reaches the client. Reported against tool_search_output items,
where the rejection surfaced as "Unknown parameter: 'input[60].status'".
Clear the status of every input item sharing the rejected item's type in the
same pass. Items of other types keep theirs: the rejection only proves that
the rejected item's type has no status field. When the rejected item carries
no type to match on, fall back to clearing the named index alone.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Wesley Liddick