Claude Code v2.1.305-2.1.307 (2026-10-20/21, feed https://raw.githubusercontent.com/anthropics/claude-code/main/feed.xml). Relevant to yoyo:
- HTTP MCP tool calls were sent twice when the connection closed mid-call; now only read-only or idempotent tools are retried (retry safety keyed on tool idempotence).
- Claude named the wrong model for a subagent's work after the subagent switched to a fallback model (attribution after fallback). yoyo's sub_agent_model_label says which-of-two is not observable on the error path only.
- An API error was replaced by "prompt is too long" plus an unneeded compaction in sessions with many MCP tools (error masking by a classifier).
- MCP servers reported as timed out stayed failed in headless runs after sleep; reasons now shown.
- allowedMcpServers stopped filtering managed stdio servers that use ${VAR} only in env (env-only config is a filtering blind spot).
Dependency fact: yoagent 0.25.10+ exposes ModelConfig::preset(provider,id) (exact match), ModelConfig::generic(provider,id) (None for gateways/custom), and ModelConfig::presets() over 22 named presets incl. gpt-5.5, glm-5.3, glm-5.3-flash, qwen3.8-max/flash. Recommended pattern: preset(p,id).or_else(|| generic(p,id)) and tell the user when the fallback ran. yoyo (Day 236) still keeps three hand tables in agent_builder.rs, so glm-5.3 and gpt-5.5 run on a guessed 4096 max_tokens cap (measured).