Repository navigation
chore(release): bring main (v0.124.2) into release/v0.125.0 - #7804
Conversation
… thoughts Vertex's shared global pool leaves a few percent of gemini-3.7-flash calls silent for minutes before the first byte. The gateway waited 300 s per attempt and Pi retried once, so a stall cost the person up to 10 minutes. - A streaming attempt on a route with first_byte_timeout_seconds is cut when it sends nothing for that long and sent again at once, from the existing retry budget and inside the route deadline. The built-in agenta route sets 60 s. - The built-in Gemini stream asks for thought summaries, so a model that is thinking sends bytes and only a stall hits the limit. reasoning_effort moves into thinking_config (Vertex refuses both), minimal and none map to low (gemini-3.7-flash refuses minimal). Thoughts reach the caller as reasoning_content when it asked for reasoning, and lose their text when it did not. - The runner reads the OpenAI client's 'Request timed out.' as the model provider not answering instead of an unclassified run error.
…extra_body Review follow-ups: the limit bounds the wait for response headers (Vertex sends them with the first output), a null extra_body still gets thought summaries, and a reasoning_effort that is not a string is left for Vertex instead of raising.
…mini route Off by default. When on, builtin/agenta calls ask for Vertex's Priority PayGo tier. Google bills it above Standard and the wallet still charges Standard, so the platform pays the difference.
….8 flash default on built-in Gemini - The built-in route's response-headers limit is 90 s. - Each retry of a stalled call ends the system prompt with one more newline. Vertex can stay stuck on one request and stall again on the identical body; any change the model sees answers in seconds. - A built-in Gemini call that names no reasoning effort thinks at low instead of Vertex's default medium (streaming: thinking_level low; non-streaming: reasoning_effort low). - gemini-3.8-flash is the first built-in model, so new agents default to it. Existing agents keep their saved model.
docs(docs): changelog entries for v0.123.0, v0.124.0 and v0.124.1
fix(api): stop built-in Gemini turns from hanging on stalled Vertex calls, and default to gemini-3.8-flash at low thinking
fix(helm): give api and web-mobile enough time to start on a fresh node
[release] v0.124.2
# Conflicts: # api/pyproject.toml # api/uv.lock # clients/python/pyproject.toml # clients/python/uv.lock # hosting/kubernetes/helm/Chart.yaml # sdks/python/pyproject.toml # sdks/python/uv.lock # services/pyproject.toml # services/uv.lock # web/ee/package.json # web/mobile/package.json # web/oss/package.json # web/package.json # web/packages/agenta-api-client/package.json
📘 Docs preview
Pages changed in this pull request
This comment updates in place on every push. |
📝 Summary
Merge Risk:
|
There was a problem hiding this comment.
Actionable comments posted: 3

Review info
-
Run configuration-
- Source mode: Auto
- Repository: Agenta-AI/agenta
- Branch: sync/main-into-v0.125.0
- File: .coderabbit.yaml
- Commit: f558897
- Layers: Repository YAML (base), Organization UI (inherited)
- Review profile: CHILL
- Plan: Advanced
- Run ID:
be1e4ffd-1aff-4ee1-8916-5a5b90d80fa0
Files selected for processing: 21-
api/oss/src/core/gateways/llms/catalog.pyapi/oss/src/core/gateways/llms/dtos.pyapi/oss/src/core/gateways/llms/providers/passthrough/adapter.pyapi/oss/src/core/gateways/llms/providers/passthrough/static_fields.pyapi/oss/src/core/gateways/llms/service.pyapi/oss/src/core/gateways/llms/types.pyapi/oss/src/utils/env.pyapi/oss/tests/pytest/unit/gateways/test_gateways_builtin_agenta.pyapi/oss/tests/pytest/unit/gateways/test_gateways_llm_retries.pydocs/blog/entries/chat-redesign-and-office-previews.mdxdocs/blog/entries/project-access-and-permissions.mdxdocs/blog/entries/project-provider-keys-only.mdxhosting/docker-compose/ee/env.ee.dev.examplehosting/docker-compose/ee/env.ee.gh.examplehosting/docker-compose/oss/env.oss.dev.examplehosting/docker-compose/oss/env.oss.gh.examplehosting/kubernetes/helm/templates/api-deployment.yamlhosting/kubernetes/helm/templates/web-deployment.yamlhosting/kubernetes/helm/templates/web-mobile-deployment.yamlservices/runner/src/engines/sandbox_agent/errors.tsservices/runner/tests/unit/model-request-timeout.test.ts
Included review availability: This review used your included allowance. Your plan provides up to 8 included reviews per hour; 7 remain after this review.
-
| with pytest.raises(LLMUpstreamTimeoutError): | ||
| await _relay( | ||
| _service( | ||
| adapter, timeout_seconds=0.15, response_headers_timeout_seconds=0.1 |
There was a problem hiding this comment.
Widen the timing margin in test_a_stall_retry_never_runs_past_the_routes_deadline.
The test passes only if the first stall ends before the 0.15 s route deadline. After the 0.1 s cut, _may_retry checks loop.time() < deadline. The margin is 50 ms. A loaded CI runner can delay the event loop by more than that. _may_retry then returns false and the test sees one call instead of two. The test becomes flaky.
Use a larger margin. Keep the ratio so the second attempt still has less time left than response_headers_timeout_seconds.
Proposed fix
-
--- "a/api/oss/tests/pytest/unit/gateways/test_gateways_llm_retries.py" +++ "b/api/oss/tests/pytest/unit/gateways/test_gateways_llm_retries.py" @@ -463,9 +463,9 @@ with pytest.raises(LLMUpstreamTimeoutError): await _relay( - _service( - adapter, timeout_seconds=0.15, response_headers_timeout_seconds=0.1 - ), + _service( + adapter, timeout_seconds=0.4, response_headers_timeout_seconds=0.25 + ), stream=True, )
Committable suggestion
-
‼️ IMPORTANT
Carefully review the code before committing. Ensure that it accurately replaces the highlighted code, contains no missing lines, and has no issues with indentation. Thoroughly test & benchmark the code to ensure it meets the requirements.Suggested changeadapter, timeout_seconds=0.15, response_headers_timeout_seconds=0.1 adapter, timeout_seconds=0.4, response_headers_timeout_seconds=0.25
|
|
||
| ## Removal takes effect at the next request | ||
|
|
||
| When you remove a member, change a role or delete a key, the API refuses the next request. A run that started just before the change ends at its first checked call with `403 access_denied`. A removed member who calls an agent gets this error instead of a generic failure. |
There was a problem hiding this comment.
Qualify the API-key revocation timing.
Line 36 says a deleted key can remain usable for about five seconds. This sentence and the summary instead promise that deletion takes effect at the next request. State the delay in both places so readers do not treat key deletion as immediate revocation.
|
|
||
| <Summary> | ||
|
|
||
| Prompts and evaluators now run on the provider keys saved in the project. A project without a key for a provider gets the error "No API key found for model". Add the key in **Settings > Providers & Models**. |
There was a problem hiding this comment.
State when a missing project key causes an error.
In OSS, a project without a saved key can use a matching services-container key by default. Lines 28–35 describe that fallback. Qualify the error claim here and in Line 23: the run fails when neither a project key nor an enabled fallback key is available. (docs.agenta.ai)
Railway Preview Environment
PR closed or converted to draft. Environment absence verified; retained storage requires separate provider accounting. Post |
Brings main into release/v0.125.0 after the v0.124.2 release merged (#7800). It adds the Gemini stall fix (#7795), the start-up probe fix (#7801) and the changelog entries.
The merge had conflicts only on version strings (0.125.0 on the release branch, 0.124.2 on main). The merge keeps 0.125.0 and takes every other change from main unchanged.