Changelog

Unreleased

add a managed local runtime to sourceweave-search-mcp so the MCP server can discover, reuse, or start its Docker-backed dependencies automatically when explicit endpoint variables are absent
coordinate multi-process local MCP usage by discovering SourceWeave-managed stacks via Docker Compose project identity, persisting runtime state, and only tearing containers down after the last managed process exits
fall back to dynamically allocated host ports when the canonical local ports are already occupied, while still reusing a healthy externally managed stack on the canonical defaults
package the managed runtime assets and refresh docs, client setup guidance, server metadata, and deterministic release coverage for the new launch behavior

add effort guidance to search_web so MCP clients can choose quick, normal, or deep research passes with matching search depth and pagination
emit richer search status updates while fetching and refining results so clients can see what the tool is doing during longer research runs
render compose-friendly SearXNG, Crawl4AI, and Redis endpoints into the generated OpenWebUI artifact while keeping host-side package defaults unchanged
standardize the release quality gate on pyright alongside lint, packaging, and deterministic runtime checks

simplify the implementation back to the minimal public search/read behavior after reviewing the hidden internal feature paths with a KISS/YAGNI lens
keep the default fit_markdown-first cleaned output path and compact search_web surface aligned across code, tests, docs, and the OpenWebUI artifact
preserve Crawl4AI-extracted tables by default through crawl and cache so read_pages returns them without adding new public flags
preserve Crawl4AI image desc values by default alongside url and alt, and document the verified CLI behavior in README and tool descriptions

sharpen search_web and read_pages descriptions so MCP clients understand when SourceWeave is a better fit than generic fetch tools for search-first, batched web research
remove the in-process page store and make Redis or Valkey the canonical persisted page cache with richer stored crawl representations
integrate Crawl4AI CacheMode intentionally so direct reads bypass upstream cache while fresh searches force a new fetch and still warm Crawl4AI's cache
fix direct URL and document reads to return immediately from the freshly fetched record even when cache persistence is unavailable in the same call
refresh README, contributing notes, and repo skills to reflect the Redis-only cache design, current quality gate, and release workflow expectations

fix MCP Registry publishing by aligning the registry namespace with the GitHub OIDC publisher namespace
simplify server.json to the supported PyPI package metadata path for registry publication
make the release metadata sync script dependency-free so lightweight release workflows can run it before installing runtime dependencies
align the publishable MCP container name and OCI labels across Dockerfile, release automation, docs, and local compose

add direct URL support to read_pages, including explicit per-URL document conversion for direct reads
make focus explicitly optional for read_pages, with empty focus performing a normal cleaned read
improve MCP and OpenWebUI tool descriptions so agents understand when to search first versus read directly by URL
add MCP Registry metadata and publishing workflow, including server.json and registry verification markers
add optional release publishing to PyPI, GHCR, and Docker Hub from the manual GitHub release workflow
improve README deployment guidance with uvx, local service containers, and container-compose examples
tighten release hygiene with explicit local ignore rules, stronger packaging metadata, and release metadata sync checks