DeepSeek native Responses rejects Codex view_image results when function_call_output.output contains input_image parts. Move media into a following user message and keep the tool output as a string so call_id pairing remains valid.
The Responses to Chat Completions bridge mapped both the instructions field and role developer items to system messages. Codex always sends both, and also injects developer notices mid-conversation when switching models, so the forwarded request contained consecutive leading system messages and mid-conversation system messages. Strict Qwen-family upstreams reject this with HTTP 400 "System message must be at the beginning."
Leading system/developer messages are now merged into a single leading system message, and later system/developer messages are downgraded to user so the content is preserved without violating upstream role ordering.
The streaming Responses-to-Anthropic converter emitted assistant text only
from output_text.delta, so a stream carrying its answer on
response.output_text.done or in the terminal response.output produced a
protocol-valid message with no text.
Record emitted text per output_text part in a ledger that survives content
block closure, and recover only the suffix a finished payload adds, as #6629
does for function-call arguments. Where the part a payload finishes cannot be
established, decline to recover rather than repeat delivered text: terminal
recovery runs only when nothing was emitted, and a done payload indexed
differently from every delta is ignored.
Give output_text.done its own handler so the shared block-closing helper and
its tool-path caller stay unchanged, and so a payload arriving after
message_stop cannot open a content block past the end of the message.
Keep always-on toolConfig from main (reasoning models without tools) while
retaining the PR behavior that drops mixed built-in tools when client
function tools are present, so includeServerSideToolInvocations stays unused.
Co-authored-by: Cursor <cursoragent@cursor.com>
Update the leftover buildTools unit test that still expected
googleSearch to coexist with client function declarations.
Co-authored-by: Cursor <cursoragent@cursor.com>
Resolve three conflicts from main's #6680 (supplementUnmappedOpenAIModels):
1. gateway_handler.go — keep PR's allowlist refactoring (ModelAllowlistEnabled,
writeAllowlistedModelsList, modelListingSource); main's OpenAI branch in
writeModelsList was auto-merged.
2. group_models_list.go (modify/delete) — delete file per PR intent; relocate
supplementUnmappedOpenAIModels to group_model_allowlist.go where gateway_service
and openai_codex_models_service can still call it.
3. openai_codex_models_service.go — combine main's supplementUnmappedOpenAIModels
wrapping with PR's ModelAllowlistEnabled() guard.
Also fix two auto-merged test files that still referenced the old
GroupModelsListConfig / ModelsListConfig types.
- Resolve conflict in openai_codex_models_service_test.go: keep both
GPT-6 Astra instructions template assertion (from this PR) and ultra
reasoning level + multi_agent fields (from #6718 on main)
- Fix compilation error on main: add missing `group *Group` parameter
to buildCodexModelsManifestForAccounts calls in
openai_codex_model_metadata_test.go (lines 24, 32) — caused by #6688
changing the function signature while #6718 added new callers with
the old signature
Upstream returns 400 INVALID_ARGUMENT when toolConfig is absent. The guard
`if !isReasoning || len(tools) > 0` skipped it for reasoning models called
without tools, contradicting the intent stated in the comment above it.
Set toolConfig unconditionally and cover the three cases with a unit test.
The Antigravity service fingerprints clients by the version embedded in the User-Agent header. The built-in default (1.23.2) had drifted far behind current real client builds, so requests using the default UA started failing OAuth and probe flows.
Bump DefaultUserAgentVersion from 1.23.2 to 2.9.1 (current client build) and update the constant test accordingly. The default remains overridable via ANTIGRAVITY_USER_AGENT_VERSION and the admin setting.
Official OpenAI hosts and models.dev entries without an api field left
Astra falling back to reasoning effort none and empty Fast tiers. Persist
complete per-model metadata, advertise Astra reasoning/Fast locally, and
surface partial sync warnings when image models remain incomplete.
Co-authored-by: Cursor <cursoragent@cursor.com>
Anthropic gates newer models on a minimum Claude Code client version. With
CLICurrentVersion pinned at 2.1.220, claude-fable-5-1 fails every request with
"Claude Code 2.1.220 does not support this model; version 2.1.251 or newer is
required", and the only fix is a release that bumps one constant.
Add SUB2API_CLAUDE_CLI_VERSION, mirroring what internal/pkg/xai already does
with XAI_GROK_CLI_VERSION. CLICurrentVersion remains the built-in pin; the new
CLIVersion() is what every consumer reads. With no env var set the resolved
value equals the pin, so existing deployments are unaffected.
The value is resolved once at package init rather than per call: the User-Agent
header and the request body's cc_version are written by different code paths for
the same request, and reading different values would make that request
contradict itself.
Validation is strict (three numeric segments, not below the pin) because
identity_service.go documents what a malformed version does once it reaches an
account's persistent fingerprint: sustained 429s without reset headers, 5s
fallback cooldown, pool collapse into a 503 storm, and no reset entry point.
Invalid values warn and fall back rather than being silently dropped.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_0168uDMAj9anaJrnH8CeFze6
Google released Gemini 3.7 Flash on 2026-08-13. Register it alongside 3.8
using the same shape as the existing 3.6 entries: bare id plus the
Antigravity -high/-low/-medium/-tiered thinking tiers, the model catalog,
the custom-mapping passthroughs, fallback pricing and the tier alias
normalizer.
Pricing matches 3.8 at $0.75 input / $3.75 output / $0.075 cached input
per MTok. Both are promotional through 2026-12-31 and double on
2027-01-01; the constants note this.
Also corrects the 3.8 catalog timestamps to the documented GA dates.
Google released Gemini 3.8 Flash on 2026-09-02. Register the model and its
Antigravity thinking tiers the same way Gemini 3.6 Flash is registered:
- DefaultAntigravityModelMapping: bare id plus -high/-low/-medium/-tiered
- geminiModels catalog exposed through /antigravity/v1/models
- ensureAntigravityDefaultPassthroughs so accounts with a custom
model_mapping still resolve the new ids
- fallback pricing and the thinking-tier alias normalizer, so the tiers
share the base model price card instead of failing pricing lookup
Fallback rates are Google's published prices: $0.75 input / $3.75 output /
$0.075 cached input per MTok. These are promotional through 2026-12-31 and
double on 2027-01-01; the constant carries a note to that effect.
Refs #6523
Wesley Liddick