Cause (Documented platform behavior): required_variables defaults to "*" and an empty list is rejected when user_prompt contains variables, to guarantee a required input socket for scheduling. Fix status: documented_behavior Limitations: - The upgrade note also says user_prompt m
Cause (Documented platform behavior): Without ClickHouse, runs cannot be located by ID alone; the SDK requires project (session) id and start time, and does not support loading child runs. Fix status: documented_behavior Other error fragments: - read_run requires start_time on
Cause (Documented platform behavior): The client pauses the batch loop on server OOM/shutdown/hangup signals and raises if the stream is not re-established in time. Fix status: documented_behavior Other error fragments: - Batch stream failed to re-establish: Evidence (public sour
Cause (Documented platform behavior): Flow runtime counts calls per listener and raises when the count exceeds max_method_calls. Fix status: documented_behavior Evidence (public sources, summarized; not reproduced by this contributor): - https://raw.githubusercontent.com/crewAIIn
Cause (Documented platform behavior): Image/audio parts are only accepted when a multimodal projector is loaded; the projector is a separate GGUF. Fix status: documented_behavior Other error fragments: - audio input is not supported - hint: if this is unexpected, you
Cause (Documented platform behavior): Previously 0 silently fell back to the constructor window_size; now it is rejected. Fix status: documented_behavior Evidence (public sources, summarized; not reproduced by this contributor): - https://raw.githubusercontent.com/deepset-ai/hays
Cause (Documented platform behavior): Security guard (advisory GHSA-f4xh-w4cj-qxq8 referenced in source) stops the SDK from reading arbitrary files from paths supplied in data. Fix status: documented_behavior Evidence (public sources, summarized; not reproduced by this contributo
Cause (Documented platform behavior): The client checks each object against the server gRPC max message size before streaming it. Fix status: documented_behavior Limitations: - Server env var name for the limit is not confirmed in the client source. Evidence (public
Cause (Documented platform behavior): Query and documents were embedded with models of different dimensions (or the store contains documents from multiple models). Fix status: documented_behavior Other error fragments: - The embedding size of all Documents should be the same. Evi
Cause (Documented platform behavior): Embeddings must be enabled explicitly (--embedding/--embeddings restricts the server to the embedding use case); the OAI route needs pooled vectors, while pooling none returns per-token unnormalized embeddings (only via the non-OAI /embedding
Cause (Documented platform behavior): Tool support is now validated against the tools resolved for each run, not only init-time tools. Fix status: documented_behavior Evidence (public sources, summarized; not reproduced by this contributor): - https://raw.githubusercontent.com/de
Cause (Documented platform behavior): Since 0.5.1 secrets are no longer read from env by default when deserializing manifests. Fix status: documented_behavior Evidence (public sources, summarized; not reproduced by this contributor): - https://raw.githubusercontent.com/langchain-
Cause (Documented platform behavior): Pydantic validators generated from the schema check list-vector length and integer ranges when validation is enabled. Fix status: documented_behavior Other error fragments: - Vector field '{fname}' contains values outside the INT8 range (-128
Cause (Documented platform behavior): PromptBuilder and ChatPromptBuilder now treat every Jinja2 template variable as required by default (previously optional); a required input that never arrives leaves the builder blocked, so the scheduler cannot make progress. Fix status: docu
Cause (Documented platform behavior): Parallel tool handlers do not implement streaming. Fix status: documented_behavior Evidence (public sources, summarized; not reproduced by this contributor): - https://raw.githubusercontent.com/567-labs/instructor/e12f8b49203b0c1f253d27c1e709
Cause (Documented platform behavior): Security hardening: deserializing public prompts can instantiate arbitrary LangChain objects, so it is opt-in. Fix status: documented_behavior Limitations: - The exact release that introduced the guard is not determined from the pinned source
Cause (Documented platform behavior): Guard in set_wired_limit; raising the OS limit requires sysctl iogpu.wired_limit_mb first (per mlx-lm docs). Fix status: documented_behavior Limitations: - That sysctl raises max_recommended_working_set_size is an inference linking the two so
Cause (Documented platform behavior): The 2026-07 backend rejects index creation from a collection or backup in create(); SDK 10 does not expose those kwargs and raises a guided error. Fix status: documented_behavior Other error fragments: - the 2026-07 API rejects
Cause (Documented platform behavior): Instructor pre-validates the schema against Bedrock native structured-output limits: minimum/maximum/multipleOf/minLength/maxLength unsupported, minItems only 0 or 1, no free-form mappings (additionalProperties must be false). Fix status: doc