v0.4.1

• Claude Code's tool search stays on when the main model is Claude, so Claude requests list MCP tools on demand instead of all at once (about half the prompt in a measured headless call). Other providers never receive the tool search tools.
• byom models lists ChatGPT models added to your account after the cached roster was built.
• The request log records cache writes as well as cache reads.
• byom stop refuses while sessions are using the bridge (requests in the last ten minutes) unless --force, since stopping it leaves open sessions with "connection refused". byom restart replaces the bridge in place, for upgrades.
• Claude Code's flags pass straight through: byom --resume <id>, byom -c, byom -p "...", and byom run <model> --continue without --.
• A local model server that isn't running (Ollama, LM Studio) shows as offline in byom models, is left out of the session, and fails at once with a clear error that Claude Code does not retry.
• byom's per-model agents get Claude Code's built-in tools only, without MCP tools, which keeps each request to another provider much smaller.
• When OpenAI closes the WebSocket before a response has produced anything, byom reconnects once and resends the turn instead of returning an API error to Claude Code.
• With Claude models available, a non-Claude main model no longer pins the session's compaction window to its own size, which made Claude models compact over and over after switching in /model.