The API rejects an invalid tool schema by position, and Claude Code cannot pre-filter those tools because Klaus disables feature-flag fetching with CLAUDE_CODE_DISABLE_NONESSENTIAL_TRAFFIC.
Klaus knows the order it sent, so it resolves tools.N.custom.input_schema to the Pi tool name in both the result-error and transport-error paths.
5 5 - Expose Klaus tools to Claude under their Pi names instead of the `mcp__klaus__` prefix, while still accepting prefixed names from resumed sessions and older runtimes
6 6 - Stop Klaus from replacing large Pi tool results with a `<persisted-output>` pointer into a deleted temp directory, and isolate the Klaus child from ambient Claude Code context, deferred tool loading, tool backgrounding, and non-streaming retries
7 7 - Capture Klaus tool identities from Claude's `PreToolUse` hook so tool calls keep working when a permission rule answers them without asking Klaus, and truncate tool results at Claude's result-size ceiling instead of losing them to a persisted file
8 +- Name the Klaus tool behind an Anthropic input-schema rejection instead of leaving the API's positional `tools.N` reference unexplained
8 9 - Reuse the running Claude child for the next turn in a Klaus session instead of spawning one per turn, swap its tool list in place when Pi changes tools, and drop it after a two-minute idle lease or any lineage change, saving about 1.7 seconds per follow-up turn
9 10 - Run every Klaus query in streaming input mode, interrupt the child instead of killing it so an abandoned turn no longer spends an extra request, flush mirrored transcripts eagerly, and surface child API retries, self-decided permission denials, warnings, and worker shutdowns to the human
10 11 - Add Bak for asynchronous S3-compatible session backup and conflict-safe restore, with host initialization, manual backup, refresh, and status commands, durable retries, safe partial-failure handling, and no filesystem or network work during Pi startup