v1.0.1

Libre Box v1.1.0

Second release: pinned and trimmed LibreChat configuration, a smaller and reproducible
sandbox image, safer log handling, and a reworked agent prompt.

LibreChat

• Pinned image. The stack no longer tracks librechat:latest — it runs the published
  ghcr.io/danny-avila/librechat:v0.8.8-rc4 tag, so redeployments are reproducible.
• Config migrated to schema 1.3.16. Added fileContextSizeLimit (16),
  fileContextCharLimit (1M) and an agents fileLimit, plus 30s refresh intervals for
  the tools and status panels.
• Share limits lifted. SHARE_IP_*, SHARE_USER_* and SHARE_VIOLATION_SCORE are
  set so self-hosted sharing is no longer throttled.
• Single-process scheduler. SCHEDULES_SINGLE_PROCESS=true keeps scheduled runs from
  being duplicated across workers.
• Leaner MCP roster. The Serper MCP server is gone (LibreChat's built-in web search
  covers it), Composio moved from sse to streamable-http, and its allowed address now
  carries the explicit :443 port that the MCP address check requires.
• File search disabled. Removed from the agent capabilities and from the default
  pinned tools.
• Summarization through OpenRouter. The provider is now OpenRouter with
  anthropic/claude-sonnet-5.
• Deployment extensions. librechat/plugins and librechat/skill are mounted
  read-only into the container, so plugins and skills can be shipped with a deployment
  without rebuilding the image.

Sandbox

• Ghidra 12.1.4 (build 20260921), with the toolchain table in the README synced.
• Pinned to linux/amd64 in both the Dockerfile and compose, so builds on arm64 hosts
  produce the same image as CI.
• Smaller image. npm, cargo, pip, uv and Go module caches are purged in every
  provisioning layer instead of only at the end, and the build scripts and versions.env
  are removed from the final image.

Operations

• make clear-logs reworked. It now handles every file under logs/, not just
  *.log: files are deleted when the stack is down and truncated in place while it runs,
  with explicit output and a hint to retry with sudo on permission errors.

Documentation

• The example agent prompt was restructured around the agent's capabilities, gained an
  autonomy section and an explicit refusal of empty replies, and dropped the
  response-language section.
• The README documents the prompt's sections and safety boundaries, the optional agent
  capabilities, deployment plugins and skills, and a fuller operations reference.

Full changelog: v1.0.0...v1.1.0

v1.0.0

ci: pin setup-uv to v10.1.0