You signed in with another tab or window. Reload to refresh your session.You signed out in another tab or window. Reload to refresh your session.You switched accounts on another tab or window. Reload to refresh your session.Dismiss alert
Периодически вылетает вот такая ошибка, приостанавливая работу агентов. Причина неизвестна.
Steps to reproduce
Возникает сама собой в ходе работы агентов. Если после этой ошибки, когда полностью устанавливается работа, сказать ему продолжай, тогда он продолжит и будет какое-то время работать корректно, а потом снова вылетит эта ошибка, и так до бесконечности.
Signal 5d2096f660d136a9 (mode_config x4, 8 High/Crit) breaks into TWO distinct config clusters, neither an engine bug:
CLUSTER A — context overflow (4x, category=InvalidRequest, recoverable=false, source=chat/generation.rs + generic):
"LLM error (400 Bad Request): request (100380/100393 tokens) exceeds the available context size (100096 tokens)". Request is only ~284-313 tokens over the model's hard 100096 cap. This is a provider-side 400: the auto-compaction in chat/generation.rs (apply_segment_summarization, maybe_compact_after_high_pressure_length_stop) only runs AFTER a length-stop, not on a hard 400 before the request is sent — so it cannot prevent this. It is NOT a retryable StreamCorrupted (recoverable=false, category=InvalidRequest): retrying the SAME request fails identically. Fix = shrink the prompt (ctx_probe/ctx_apply, new chat) or pick a larger-context model. Not config drift, not an engine defect.
CLUSTER B — search_semantic not registered (4x, category=ModelUnavailable, source=yaml_configs/mode_validation.rs) in modes: agent, setup, setup_mcp, buddy.
search_semantic IS declared in these modes' YAML (crates/refact-yaml-configs/src/defaults/modes/agent.yaml:255, setup.yaml:255, setup_mcp.yaml:192, buddy.yaml:86) and the tool IS implemented in src/tools/tool_search.rs:340, but registration is GATED ON VECDB (see src/tools/review_agents/stages.rs:356-358 VECDB_GATED_STAGE_TOOLS = ["search_semantic"]). So: vector DB not initialized/available => tool not registered => mode_validation.rs warn_unknown_mode_tools fires for each mode that lists it. It's a WARN + Buddy notify (deduped per mode via claim_first_report), not a hard failure, but the model will hit "tool not found" if it calls it. Fix = enable/initialize vecdb for the project so search_semantic registers (preferred, keeps the feature), OR remove the line from those mode YAMLs if semantic search isn't needed.
RISK: A = every long chat dies at the context wall (high friction). B = silent capability gap; agent/setup/setup_mcp/buddy modes advertise a tool that 400s on call.
SMALLEST SAFE NEXT CHECKS: (1) confirm whether the 400s are one chat or several; if one, trim context or switch model. (2) Check if vecdb/embedding backend is configured & running for this workspace; if yes, why is the registry missing the tool; if no, that explains all 4 mode_config warns at once.
Activity
RaySinner commentedon Sep 6, 2026
Signal 5d2096f660d136a9 (mode_config x4, 8 High/Crit) breaks into TWO distinct config clusters, neither an engine bug:
CLUSTER A — context overflow (4x, category=InvalidRequest, recoverable=false, source=chat/generation.rs + generic):
"LLM error (400 Bad Request): request (100380/100393 tokens) exceeds the available context size (100096 tokens)". Request is only ~284-313 tokens over the model's hard 100096 cap. This is a provider-side 400: the auto-compaction in chat/generation.rs (apply_segment_summarization, maybe_compact_after_high_pressure_length_stop) only runs AFTER a length-stop, not on a hard 400 before the request is sent — so it cannot prevent this. It is NOT a retryable StreamCorrupted (recoverable=false, category=InvalidRequest): retrying the SAME request fails identically. Fix = shrink the prompt (ctx_probe/ctx_apply, new chat) or pick a larger-context model. Not config drift, not an engine defect.
CLUSTER B — search_semantic not registered (4x, category=ModelUnavailable, source=yaml_configs/mode_validation.rs) in modes: agent, setup, setup_mcp, buddy.
search_semantic IS declared in these modes' YAML (crates/refact-yaml-configs/src/defaults/modes/agent.yaml:255, setup.yaml:255, setup_mcp.yaml:192, buddy.yaml:86) and the tool IS implemented in src/tools/tool_search.rs:340, but registration is GATED ON VECDB (see src/tools/review_agents/stages.rs:356-358 VECDB_GATED_STAGE_TOOLS = ["search_semantic"]). So: vector DB not initialized/available => tool not registered => mode_validation.rs warn_unknown_mode_tools fires for each mode that lists it. It's a WARN + Buddy notify (deduped per mode via claim_first_report), not a hard failure, but the model will hit "tool not found" if it calls it. Fix = enable/initialize vecdb for the project so search_semantic registers (preferred, keeps the feature), OR remove the line from those mode YAMLs if semantic search isn't needed.
RISK: A = every long chat dies at the context wall (high friction). B = silent capability gap; agent/setup/setup_mcp/buddy modes advertise a tool that 400s on call.
SMALLEST SAFE NEXT CHECKS: (1) confirm whether the 400s are one chat or several; if one, trim context or switch model. (2) Check if vecdb/embedding backend is configured & running for this workspace; if yes, why is the registry missing the tool; if no, that explains all 4 mode_config warns at once.
Что бы это ни значило
[-][Bug]:[/-][+][Bug]: Неизвестная ошибка[/+]