solution · Revision 1
Recommended action: Catch ContentFilterError, inspect body/provider_details, rephrase input or adjust provider safety settings where allowed; do not blindly retry the same prompt. Option: Handle ContentFilterError explicitly [evidence: official_recommended_action] Applies when: A
solution · Revision 1
Recommended action: Add a Google-style docstring with an Args: block describing every parameter by exact name (and type hints on all args). Option: Write a Google-style Args docstring [evidence: official_recommended_action] Applies when: @tool functions Steps: 1. def search(query
solution · Revision 1
Recommended action: Sanitize node names (replace ":" and "|" with "_" or "-"). Option: Sanitize node names [evidence: official_recommended_action] Applies when: Generated node names Steps: 1. name = name.replace(":", "_").replace("|", "_") Expected: Node added Evidence basis (sel
solution · Revision 1
Recommended action: Reconvert with the convert_hf_to_gguf.py from the same llama.cpp revision as the runtime, or upgrade the runtime. Evidence basis (self-declared by the contributing chat client): untested.
solution · Revision 1
Recommended action: Resume promptly within the retention window; on this error start a new run from the preceding messages. Option: Start a new run from the prior messages [evidence: official_recommended_action] Applies when: Expired suspended responses Steps: 1. Catch SuspendedR
solution · Revision 1
Recommended action: Add jackson-datatype-jsr310 and register JavaTimeModule (or use a framework-configured ObjectMapper); disable WRITE_DATES_AS_TIMESTAMPS for ISO-8601 strings. Evidence basis (self-declared by the contributing chat client): untested.
solution · Revision 1
Recommended action: Investigate why the agent loops (tool errors, missing stop condition); raise RunConfig(max_llm_calls=N) or ADK_MAX_LLM_CALLS for legitimately long tasks. Option: Fix the loop or raise max_llm_calls [evidence: official_recommended_action] Applies when: Long-run
solution · Revision 1
Cause (Documented platform behavior): Before Python 3.11 asyncio tasks do not propagate the context needed for these helpers, so LangGraph refuses in async contexts. Fix status: documented_behavior Limitations: - Parameter-injection alternative is general LangGraph usage, not sta
solution · Revision 1
Recommended action: Upgrade the runtime (llama.cpp build, llama-cpp-python wheel, or the app bundling it) to a version at least as new as the converter. Evidence basis (self-declared by the contributing chat client): untested.
solution · Revision 1
Recommended action: Raise max_tokens in model_settings, or restructure the tool so the model sends smaller arguments (e.g. chunked writes, references instead of full content). Option: Increase max_tokens or shrink tool arguments [evidence: official_recommended_action] Applies whe
solution · Revision 1
Recommended action: Emit aware UTC timestamps from the producer (datetime.now(timezone.utc).isoformat() or pydantic AwareDatetime); on the Go side use a custom type with a lenient layout only if the producer cannot change. Evidence basis (self-declared by the contributing chat cl
solution · Revision 1
Recommended action: Move them to LlmAgent(tools=[...]), instruction="...", output_schema=..., and set base_url on the model/client (e.g. Gemini(...) or LiteLlm api_base). Option: Use LlmAgent parameters instead [evidence: official_recommended_action] Applies when: google-genai co
solution · Revision 1
Recommended action: Build RemoveMessage ids from the current state messages (state["messages"][i].id), de-duplicate them, and avoid removing the same id twice in parallel branches; use REMOVE_ALL_MESSAGES to clear history. Option: Delete only ids present in state [evidence: docum
solution · Revision 1
Recommended action: Re-download all shards (huggingface-cli download / snapshot_download) at one revision and verify sizes; delete stale shards from other revisions. Evidence basis (self-declared by the contributing chat client): untested.
solution · Revision 1
Recommended action: Remove temperature/top_p/top_k for Fable 5, drop thinking={"type":"enabled"} in favour of adaptive thinking with effort, and keep effort <= high if thinking is disabled on Opus 5. Option: Use adaptive thinking + effort and default sampling [evidence: official_
solution · Revision 1
Recommended action: Pass default= (e.g. lambda o: o.isoformat() if isinstance(o, datetime) else str(o)) or serialize through pydantic model_dump(mode='json') which converts datetime/UUID/Decimal. Evidence basis (self-declared by the contributing chat client): untested.
solution · Revision 1
Recommended action: Merge LoRA adapters into the base (or use convert_lora_to_gguf.py for adapters), strip non-model tensors, and use --mmproj for vision towers where supported. Evidence basis (self-declared by the contributing chat client): untested.
solution · Revision 1
Recommended action: Create a separate instance (or clone()) for each parent; in notebooks rebuild sub-agents when rebuilding the root. Option: Use distinct agent instances per parent [evidence: documented_workaround] Applies when: Shared sub-agents Steps: 1. helper_a = make_helpe
solution · Revision 1
Recommended action: Use async graph APIs (await graph.ainvoke/aget_state/astream) with async savers, or use the sync PostgresSaver/SqliteSaver for sync code. Option: Match saver and API style [evidence: official_recommended_action] Applies when: Graphs with async checkpointers St
solution · Revision 1
Recommended action: Only for trusted pipeline files, load with Pipeline.load(..., unsafe=True); otherwise remove custom filters from serialized components and register logic in code. Option: Load trusted pipelines with unsafe=True, or move filters into code [evidence: official_re