@caeliq/claude-code-router 2.0.3 → 2.1.0

package/README.md

@@ -507,11 +507,12 @@ Example provider:
 ```
 
 - **bridge** (default): Claude Code hosts tools; Cursor built-ins are denied in an isolated workspace under `~/.claude-code-router/cursor-sdk-workspaces/`
+- Cursor builds its harness prompt server-side from the SDK workspace root, so bridge mode grounds the model in the host environment instead: the project root and platform from the incoming `<env>` block are injected through the workspace `AGENTS.md`, the head and tail of the agent prompt, and the built-in denial message. Host tool calls whose arguments reference the isolated workspace are intercepted, answered with a correction rather than forwarded to Claude Code, and counted in the session metrics. Isolated workspaces are removed with their session and swept when orphaned.
 - Discover models with `ccr model get cursor` (lists via `Cursor.models.list`, not REST `/models`)
 - Docker Compose passes `CURSOR_API_KEY` into the container when set; local Cursor sandboxing is forced off in Docker
 - Cursor prompt caching is native to the held-open SDK agent session. CCR reports per-request token estimates to Claude Code and maps SDK cache-read deltas back as bounded Anthropic cache-read usage.
 - Cursor thinking is forwarded from both SDK stream `thinking` messages and token-level `Agent.send({ onDelta })` `thinking-delta` updates, then closed with a synthetic signature so Claude Code can render it as Anthropic extended thinking. Claude Code 2.1.89+ hides interactive thinking summaries by default; enable `"showThinkingSummaries": true` in the settings file passed to Claude Code to display them. This is a client rendering setting: CCR still transports the thinking block when it is disabled.
-- Stopping a Cursor response cancels the owned SDK run with bounded cleanup and invalidates unsafe SDK sessions; active-run send failures use Cursor's native `local.force` retry before CCR falls back to a fresh full-transcript session.
+- Cursor turns are coordinated per Claude conversation. Identical overlapping retries share one bounded response producer (and recent completed result), so only one consumer reads the Cursor iterator and only one `agent.send` is submitted. Stopping the last subscriber awaits bounded run/iterator retirement before re-entry. Pure, exactly matched tool results resume a live parked run; a rejected result plus meaningful replacement text, an unmatched/dead run, or divergent transcript retires the old agent and replays the full conversation. Idle agents receive slim follow-ups only when the incoming transcript is exactly the committed host-visible transcript plus one supported user message; larger suffixes are fully replayed. CCR does not use `local.force` to bypass Cursor's active-run guard.
 
 > **See also**: Full Cursor setup is documented in `docs/docs/server/guides/cursor.md`.