Anthropic count_tokens rejects generation-only fields such as temperature, top_p, top_k, stream, and stop sequences. Passing the original messages payload through unchanged can turn otherwise valid requests into upstream 400 errors.
Sanitize only the count_tokens upstream payload after the gateway's existing request normalization, preserving fields that existing compatibility paths rely on while removing parameters the count_tokens endpoint does not accept.
Fixes#2764
Co-authored-by: Cursor <cursoragent@cursor.com>
Follow-up to #2816 (already merged): the same long-context pricing
exemption that affected cache_read also applies to all three
cache_creation price fields (standard, 5m ephemeral, 1h ephemeral).
computeCacheCreationCost reads these prices directly from pricing and
never sees the LongContextInputMultiplier that computeTokenBreakdown
applies to inputPrice / outputPrice / cacheReadPrice.
For GPT-5.4 / 5.5 above the 272k threshold, this causes the cache_write
portion of long sessions to be billed at roughly half what it should
be (default multiplier 2.0). Cache writes are conceptually input-side
operations and should share the same long-context treatment as input /
cache_read.
This patch threads an explicit multiplier into computeCacheCreationCost
so the function can be unit-tested in isolation and matches the existing
pattern used for cache_read. computeTokenBreakdown captures the long
context decision once and passes LongContextInputMultiplier when it
applies, 1.0 otherwise.
Adds three regression tests mirroring the #2816 cache_read tests:
- positive: long-context triggered -> cache_creation scaled by 2.0x
- negative: below threshold -> cache_creation stays at base price
- breakdown: 5m + 1h ephemeral prices both scaled when applicable
Refs #2816
Co-authored-by: Cursor <cursoragent@cursor.com>
The antigravity upstream-passthrough path (account.Type == AccountTypeUpstream
forwarding to a Claude-format upstream) drains the SSE stream via
streamUpstreamResponse + extractSSEUsage. The extractor only reads top-level
event["usage"], which matches Anthropic's message_delta but misses
message_start where usage is nested under event.message.usage.
As a result, every streaming /v1/messages request through this path drops
the input-side fields (input_tokens, cache_read_input_tokens, cache_creation_*)
and writes a usage_logs row with input_tokens=0 + output_tokens>0. The user
in #2332 observed 2,728 such rows attributed to claude-opus-4-6 / haiku-4-5
streaming requests; their billing on output is correct but the input-side
accounting is missing. (Their "duplicate write from message_delta" hypothesis
isn't borne out by the code — RecordUsage is invoked once per request and
writeUsageLogBestEffort dedupes by request_id; what they're seeing is
single records produced by this buggy extractor.)
Branch on event.type so message_start reads from event.message.usage and
other events keep using event.usage, matching how parseSSEUsagePassthrough
already handles both shapes for the Anthropic OAuth / API-key / Bedrock paths.
Adds two extractSSEUsage table cases plus a TestExtractSSEUsage_StreamingSequence
that drives the message_start → message_delta sequence end-to-end; both fail
on main and pass with this change.
Fixes#2332
Co-authored-by: Cursor <cursoragent@cursor.com>
When session long-context pricing is triggered in computeTokenBreakdown
(e.g. GPT-5.4 / GPT-5.5 above the 272k token threshold), the multiplier
was only being applied to InputPricePerToken and OutputPricePerToken.
The cache_read price was left at its base value, so CacheReadCost was
silently undercharged whenever a long-context session also had cache
hits — which is essentially every long Codex / Claude Code session.
Concretely for gpt-5.4 with 300k cache_read tokens, the bug
under-billed the request by exactly 1x the LongContextInputMultiplier
on the cache portion (e.g. 0.075 instead of 0.150 in the regression
test).
Cache reads are conceptually input-side replays, so they should scale
with LongContextInputMultiplier, matching the treatment of
InputPricePerToken.
Adds two regression tests:
- positive: long-context triggered -> cache_read scaled by 2.0x
- negative: below threshold -> cache_read stays at base price
Fixes#2293
Co-authored-by: Cursor <cursoragent@cursor.com>
Pool mode currently retries the same account for a fixed set of
upstream HTTP statuses: 401, 403, 429. Some upstream pool deployments
also need same-account retry for transient provider/proxy statuses
such as 502, 503, 520, 529, but hard-coding more statuses changes
behavior for everyone.
Add a per-account credentials option `pool_mode_retry_status_codes`
that lets admins choose which upstream HTTP status codes trigger
same-account retry in pool mode:
- Unset (default): preserve the current 401/403/429 default
- Explicit list: override the defaults with the configured codes
- Codes normalized to the 100-599 range, deduplicated, sorted
The standalone `isPoolModeRetryableStatus` helper is kept as the
default-only fallback. All 15 gateway call sites switch to the new
`Account.IsPoolModeRetryableStatus` method so behavior is preserved
for accounts that do not configure the new field.
Frontend admin UI gains a "Retry Status Codes" comma-separated input
under the pool-mode section in both Create/Edit account modals
(en + zh i18n).
Fixes#2731
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
Errcheck flagged three unchecked strings.Builder.WriteString calls and
gofmt rejected over-aligned trailing comment in the route table.
Rewrite writeResponsesFailedSSE with json.Marshal on typed structs
instead of Builder+strconv.Quote. Same wire format, but:
- no unchecked Write returns to silence
- strict JSON escaping (strconv.Quote emits \a and \v which are not
valid JSON; Marshal handles all runes correctly)
- omitempty model field via struct tag instead of conditional Builder
- consistent with the json.Marshal style used elsewhere in handler/
Collapse trailing comment whitespace in stream_error_event_test.go to
satisfy gofmt.
All 30+ subtests in the package still pass.
Case B: when a slot wait flushes SSE ping comments first (Writer.Written
becomes true), the previous ensureForwardErrorResponse short-circuited
on `c.Writer.Written()` and returned false without notifying the client.
Subsequent upstream errors (http2 timeout, stream INTERNAL_ERROR, etc.)
produced silent EOF; Codex CLI reported "stream closed before
response.completed" just like the user-slot timeout case.
Remove the Written() early return; coerce streamStarted to true when
Writer has already been written to, and let handleStreamingAwareError
walk the existing logic — which now (thanks to the previous commits)
emits a protocol-compliant response.failed for /responses paths and the
legacy `event: error` for others.
Update tests that previously asserted "do not override written response":
the new contract is to *append* an SSE terminal frame so the client sees
a clean close instead of EOF. recoverResponsesPanic inherits this fix.
Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>
The first revision compared GetInboundEndpoint(c) against EndpointResponses
("/v1/responses"). NormalizeInboundEndpoint only recognizes paths that
contain the literal "/v1/responses" substring, but the project actually
registers six /responses routes — three of which (top-level
r.POST("/responses", ...) and codexDirect's "/backend-api/codex/responses")
have FullPath values without the "/v1" prefix and therefore fall through
to the default branch.
Codex CLI users targeting the bare /responses route at the production
deployment (observed 2026-05-24 ~11:05 UTC, user 16) never reached the
new writeResponsesFailedSSE path: the endpoint check was false, the
legacy `event: error` frame fired, and the strict SDK kept reporting
"stream closed before response.completed".
Replace the strict equality check with inboundIsResponses(c), which
uses suffix detection on FullPath (falling back to URL.Path when
FullPath is empty in test fixtures) and covers all six route variants:
/v1/responses[/...]
/responses[/...]
/backend-api/codex/responses[/...]
Add test table covering all routes plus negative cases.
Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>
When /v1/responses streaming hits the user/account concurrency wait, the
wait loop sends SSE ping comments to keep the connection alive, which
flushes HTTP 200 + headers. If the wait then times out (or any other
post-flush error fires), handleStreamingAwareError previously emitted a
generic `event: error` frame. Codex CLI requires the stream to end with
a Responses terminal event (response.completed/failed/incomplete/cancelled),
so it reports "stream closed before response.completed" and the user-facing
rate-limit intent is lost.
This change detects inbound = /v1/responses in both handleStreamingAwareError
implementations and emits a protocol-compliant response.failed event whose
field set mirrors apicompat.makeResponsesCompletedEvent
(id/object/model/status/output/error). The synthetic id reuses
ctxkey.RequestID so client errors can be grepped against server logs.
sequence_number is intentionally omitted to preserve monotonicity on streams
that already emitted real events.
Other inbound endpoints (/v1/chat/completions, /v1/messages) keep their
legacy formats untouched.
Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>
