Repository Issue Activity (beta)

#3432 【研究分享】Self-Bootstrapping Speed Limit: A Quantitative Model from a 32-Round Self-Evolving AI Experiment

#3431 Pre-publication review: our reading of Letta's memory write path (corrections welcome by 2026-08-30)

#3362 System Optimization Protocol: Cross-Session Variable Stabilization via Phase 6 Axis Rotation

#3391 Pre-execution gate as a regression-test pattern for sandbox-to-server boundary classes

#3425 Sharing a crash-proof, EDR-safe memory layer for MemGPT-style memory

#3428 [Proposal] Hosted memory for Letta users — 97.5% fewer tokens (ViBo Cloud API)

#3280 MCP SSRF protection needs allow-list for Docker/local development

#3390 [Bug]: Missing ClientTimeout in AsyncComposioToolSet leads to severe coroutine starvation

#3310 Error wrapper mislabels upstream provider rate limits as 'Rate limited by OpenAI' (affects GLM-5 / z.ai and likely others)

#3346 OpenAI-compatible /v1/chat/completions endpoint rejects standard image_url content blocks

#3427 Tool execution sandbox: NameError 'DynamicModel' when args_json_schema has no 'title' field

#3426 [test] permission check

#3270 sliding_window compaction ignores percentage, performs full context wipe

#3242 Post-step compaction uses inflated context_token_estimate (total_tokens includes output), causing redundant compaction passes

#3247 OpenAIProvider ignores API-reported context_length, defaults all unknown models to 30k tokens

#3279 GLOBAL_MAX_CONTEXT_WINDOW_LIMIT caps summarizer model, preventing compaction of overflowed context

#3309 Feature request: delete_protected flag to prevent agent self-deletion

#3342 Security: OWASP Agent Memory Guard for memory poisoning defense (ASI06)

#3352 Feature request: bring scheduled messages to self-hosted (parity with cloud API)

#3357 Feature: Pluggable type-aware output-compressor pipeline for tool results

#3424 Memory Safety Net: crash-proof + EDR-safe recovery layer for MemGPT-style memory

#3302 PATCH /tools silently disables parallel execution when the field is omitted

#3379 Local LLM completion settings are mutated across requests

#3383 LM Studio provider stores /api/v0 as the inference endpoint

#3385 VLLM provider ignores encrypted API keys during model discovery