Image generation and media understanding both thread the
sanitized models.providers.google.request config (including
allowPrivateNetwork) into resolveGoogleGenerativeAiHttpRequestConfig.
Speech synthesis omitted that arg, so TTS always saw
allowPrivateNetwork: false regardless of config — silently falling
back to a different speech provider when the configured Google TTS
endpoint resolved to a private/internal IP (proxies, custom backends,
test mocks).
Mirror the image-generation-provider pattern: thread request through
synthesizeGoogleTtsPcm at both call sites (synthesize and
synthesizeTelephony).
Follow-up to #67216.
* fix(github-copilot): preserve all reasoning IDs and add gpt-5.3-codex support
The existing guard (8fd15ed0e5) only skipped rewriting reasoning item IDs
when encrypted_content was a non-null string. When gpt-5.3-codex is used
via GitHub Copilot, the model falls through to the forward-compat catch-all
with reasoning:false, so encrypted_content is never requested and arrives
as null — bypassing the guard and causing a rewrite. Copilot validates
reasoning item IDs server-side regardless of whether the client includes
encrypted_content, so the rewritten id triggers the 400 error.
Two changes:
1. connection-bound-ids.ts: skip ALL reasoning items unconditionally.
Reasoning items always reference server-side state bound to their
original ID; rewriting any of them breaks Copilot's lookup.
2. models.ts + index.ts: extend the forward-compat cloning logic to
cover gpt-5.3-codex (adds it to the template-target set and to
CODEX_TEMPLATE_MODEL_IDS so it can also serve as a template source
for gpt-5.4). Adds gpt-5.3-codex to COPILOT_XHIGH_MODEL_IDS for
the thinking profile.
Thanks @InvalidPandaa.
* docs(github-copilot): clarify gpt-5.3-codex is a no-op template for itself
https://claude.ai/code/session_01EAFmq4WyKkiUkVAqRXp4Bm
* fix(github-copilot): remove dead reasoning prefix branch in deriveReplacementId
https://claude.ai/code/session_01EAFmq4WyKkiUkVAqRXp4Bm
* fix(github-copilot): align reasoning id replay tests
* test(plugin-sdk): use cjs sidecar for require fast path
---------
Co-authored-by: Claude <noreply@anthropic.com>
Co-authored-by: Peter Steinberger <steipete@gmail.com>
* fix(minimax): use dedicated image generation endpoint
MiniMax image generation uses a dedicated API endpoint
(api.minimax.io/v1/image_generation) that is separate from the
text/chat API endpoint (api.minimax.io/anthropic).
Previously, the resolveMinimaxImageBaseUrl function would extract
the origin from the provider's configured baseUrl. If a user had
configured their baseUrl to the chat endpoint (e.g.,
api.minimax.chat/anthropic), the image generation would incorrectly
use that endpoint, resulting in "invalid api key" errors.
This fix always uses the dedicated image generation endpoint,
ignoring the provider's baseUrl configuration for image generation.
Fixes#61149
* fix(minimax): support CN endpoint for image generation
Respect MINIMAX_API_HOST environment variable to determine whether
to use the global (api.minimax.io) or CN (api.minimaxi.com) endpoint
for image generation.
This ensures that CN users who configure MINIMAX_API_HOST to use
api.minimaxi.com will continue to use the CN endpoint for image
generation, while global users continue to use api.minimax.io.
The original bug was caused by the code extracting the origin from
the provider's configured baseUrl, which could be set to incorrect
endpoints like api.minimax.chat. This fix uses the dedicated image
generation endpoints instead.
Fixes#61149
* fix(minimax): infer CN endpoint from provider config when env is unset
When MINIMAX_API_HOST is not set, fall back to checking the provider's
configured baseUrl to determine whether to use the CN or global image
endpoint. This ensures CN users who went through onboarding (which sets
models.providers.minimax.baseUrl to https://api.minimaxi.com/anthropic)
are correctly routed to the CN image endpoint.
The isMinimaxCnHost check ensures we only use the baseUrl origin for
CN detection - invalid endpoints like api.minimax.chat would not match
minimaxi.com and would correctly fall through to the global default.
Fixes#61149
* test(minimax): cover dedicated image endpoints
* fix(logging): handle context assembly diagnostics
* Revert "fix(logging): handle context assembly diagnostics"
This reverts commit f51d2f7d67f8193268dd37553ac77e80a0423390.
* test(minimax): isolate image endpoint env
* docs(changelog): credit minimax image fix
---------
Co-authored-by: Peter Steinberger <steipete@gmail.com>
Fix the bundled Bonjour gateway discovery crash-loop caused by ciao probe cancellation rejections after the Bonjour plugin migration.
The plugin entry now wires the existing rejection handler into the advertiser, and the unhandled-rejection handler registry is anchored on globalThis so staged plugin SDK module copies register into the same process-level handler set used by the host.
Verification:
- pnpm test:serial extensions/bonjour/src/advertiser.test.ts src/infra/unhandled-rejections.fatal-detection.test.ts
- OPENCLAW_LOCAL_CHECK_MODE=throttled pnpm check:changed partially completed: conflict markers plus core/core-test/extensions/extension-test typecheck passed; local lint lane hit a self-lock and was stopped.
* feat(litellm): add image generation provider
Registers litellm as an image-generation provider so model refs like
litellm/gpt-image-2 route through the LiteLLM proxy, and
agents.defaults.imageGenerationModel.fallbacks entries of the form
litellm/... resolve without "No image-generation provider registered
for litellm" errors.
Implementation uses the OpenAI-compatible /images/generations and
/images/edits endpoints that LiteLLM proxies for. BaseUrl resolves from
models.providers.litellm.baseUrl (default http://localhost:4000). Private
network is auto-allowed when baseUrl is a loopback/RFC1918 address, which
covers the common self-hosted LiteLLM proxy case without needing
OPENCLAW_PROVIDER_ALLOW_PRIVATE_NETWORK. Public baseUrls keep normal SSRF
defaults.
Default model is gpt-image-2 (matching upstream 4.21+ OpenAI default).
Advertises the same 2K/4K sizes OpenAI now exposes, plus legacy
256/512/1024 for dall-e-3. Supports both generate and edit.
Local patch. LiteLLM has no upstream image-generation support yet; revisit
if upstream adds one.
* ci: rerun after upstream main hot-fix
* fix(litellm): harden image generation provider
---------
Co-authored-by: Chris Zhang <chris@ChrisdeMac-mini.local>
Co-authored-by: Peter Steinberger <steipete@gmail.com>
