MCP Desktop Tools

# Performance and Profiling MCP Desktop Tools B2 introduces persistent caching and bounded concurrency to accelerate repeated filesystem traversals and text searches. ## Disk cache for `repo_map` * Cached under `~/.mcpdt/cache` (override with `MCPDT_CACHE_DIR`). * Keys include workspace id, traversal parameters, include/exclude globs, and a tree signature derived from directory metadata. * Entries expire after **1 hour** by default (`MCPDT_CACHE_TTL_SEC`). * Disable globally via `MCPDT_DISABLE_CACHE=1` or per-invocation with `mcp-tools --no-cache`. * Metrics include `cache_hit`, `cache_key`, `fs_walk_count`, `bytes_scanned`, and `max_workers`. ## In-memory cache for `search_text` * TTL cache scoped to the running process with a 120 second expiry. * Cache key factors in workspace, query, glob filters, depth, and regex flags. * Disable with `--no-cache` / `MCPDT_DISABLE_CACHE=1`. * Metrics include `cache_hit`, `rg_elapsed_ms`, and `cache_key`. ## Concurrency controls * Filesystem traversal uses a bounded thread pool with the default size `min(32, max(4, 2*CPU))`. * Override with `mcp-tools --max-workers <N>` or `MCPDT_MAX_WORKERS`. * The traversal batches file stat operations (`fs_batch_size=256`) to balance throughput with memory usage. ## Subprocess timeouts * ripgrep and git invocations default to **30s** (`MCPDT_SUBPROC_TIMEOUT_MS`). * Timeouts raise `TimeoutError` surfaced in CLI warnings/metrics. ## Profiling * Use `mcp-tools --profile ...` to capture stage timings. Tables are printed to `stderr`; structured data is included in the `me trics.profile` array. * Stages: * `signature`, `cache_lookup`, `walk` for `repo_map`. * `build_args`, `run_rg`, `parse` for `search_text`. * git operations contribute to `git_cmd_ms` in metrics. ## Recommendations * Warm caches on large repositories to benefit from disk reuse. * Tune `--max-workers` downward on slow disks or networked filesystems to reduce contention. * Combine `--include/--exclude` globs to reduce `bytes_scanned` and pressure on caches. * Record representative timings in this file after running benchmarks (e.g., 40%+ speed-up on cached repo_map traversals).