v0.23.1

Patch release: smarter tool drops, correct behavior in restricted-tool sessions, and a prompt-cache fix for multi-session setups.

Improvements

Dropped tool calls keep a skeleton while recent. When a tool output is dropped (by the agent or automatic cleanup) and the call is among the 20 most recent, Magic Context now keeps a compact [truncated] placeholder — the call and its result stay paired on the wire instead of disappearing. This keeps less capable models from losing their bearings (or inventing tool calls) right after a cleanup, and replaces what the removed drop_tool_structure option used to do — automatically, and without cache cost. Older drops are still removed entirely.
Context-reduction surface now follows tool availability. Sessions whose tool set doesn't include ctx_reduce (common for restricted subagents) no longer get §N§ tags, reduction guidance, or reduction nudges they can't act on. The decision is made once per session, so the system prompt stays cache-stable.

Removed a false-positive prompt-cache invalidation trigger. Tool-definition fingerprints are recorded process-wide, so in multi-session setups one session re-recording tool definitions (e.g. after a restart) could force other sessions to rebuild their cached history baseline and re-bill large prompt prefixes. The tool-set-change trigger is gone; genuine tool changes are reconciled by the existing rebuild paths.

Documentation site: https://docs.cortexkit.io — getting started, concepts, full configuration reference (generated from the schema), tools/commands reference, and troubleshooting.
Regression harness for openai-compatible provider message-adjacency invariants (groundwork for #135).