Skip to content

CHANGELOG.md

Latest commit

 

History

History
112 lines (91 loc) · 5.98 KB

CHANGELOG.md

File metadata and controls

112 lines (91 loc) · 5.98 KB
Raw

Changelog

All notable changes to this project will be documented in this file.

The format is based on Keep a Changelog, and this project adheres to Semantic Versioning.

[Unreleased]

Added

  • Four GitHub Actions workflows (test.yml, lint.yml, security.yml, release.yml), each with every uses: action pinned to a live-verified 40-character commit SHA and each proven hosted-green with a real run on origin/main (CI-01, CI-02, CI-03, CI-04)
  • A tag-safe release/publish workflow (release.yml) that gates GHCR's :latest tag on the actual triggering event (flavor: latest=auto only on real tag pushes, never on workflow_dispatch), proven via a hosted dry run on both trigger paths before any real semver tag was ever pushed (CI-05)
  • Four HTTP-200-verified README badges (Test, Lint, Security, Release), added only after each referenced workflow had a real hosted-green (or hosted-verified dry-run, for release.yml) run, with a hygiene test guarding that every badge references a workflow file that actually exists on disk (CI-06)

Changed

  • ruff added as a pinned lint dependency and calibrated clean against its built-in default rule set across src/ and tests/; mypy src/ now exits 0 under the pre-existing strict [tool.mypy] config, with zero strictness dialed back (CI-02)
  • Dockerfile now installs via uv sync --locked --no-dev --no-editable instead of an unlocked uv pip install --system -e ., so the built image's dependency graph is byte-identical to the audited uv.lock (CI-05)

Security

  • fastmcp dependency ceiling raised from >=2.11,<3 to >=3.2.0 (resolved 3.4.0), remediating CVE-2026-32871 (GHSA-vv7q-7jx5-f767, OpenAPIProvider SSRF/path-traversal). The full local test suite (469 tests) and the byte-frozen golden MCP tool-schema snapshot both pass unchanged on the new version (CI-03)

[1.0.0] - Pending (not yet tagged)

Prepared for the first public release. This version has been assembled and tested on main but no v1.0.0 git tag has been pushed yet, and no image has been published to GHCR under this version -- tagging is a separate, deliberate operator decision (a v* tag triggers the production release workflow). The content below reflects what is ready to ship once that tag is pushed.

Added

  • Canonical per-tool risk classification (RISK_CLASSIFICATION) and a dispatch-layer read-only gate for write/destructive MCP tools (SEC-01/SEC-02)
  • Shared MCP tool registry (registry.py) consolidating stdio and HTTP tool registration into a single source of truth (CONS-01)
  • respx-based mock HTTP test harness for FortiGateAPI boundary tests (CONS-03)
  • Live end-to-end test harness exercising the real ASGI app over a pre-bound socket (CONS-02)
  • Golden MCP tool-surface schema snapshot freezing tool name and parameter schema across both transports (STAB-04)
  • Pure-ASGI Bearer-token authentication middleware for the HTTP transport (SEC-05)
  • Centralized host/port/vdom/identifier validators applied at every REST call site
  • SECURITY.md, CONTRIBUTING.md, CODE_OF_CONDUCT.md, and GitHub issue/PR templates
  • Four verified MCP client configuration examples: Claude Desktop (stdio and HTTP), Claude Code, and Cursor

Changed

  • FortiGateDeviceConfig.verify_ssl now defaults to true for devices loaded from a config file (SEC-04)
  • mcp and fastmcp dependencies pinned to verified PyPI release bands (mcp>=1.23.0,<2.0, fastmcp>=2.11,<3)
  • The stdio server now shares one tool registry with the HTTP server instead of duplicating ~30 hand-written tool registrations
  • README.md and HTTP_MCP_GUIDE.md rewritten to describe only verified, evidence-based behavior (uv-first install, loopback-bind guidance for unauthenticated HTTP, no unverified claims)
  • pytest configuration consolidated entirely into pyproject.toml; coverage gate calibrated to a real measured baseline

Removed

  • The undocumented optional nginx reverse-proxy profile in docker-compose.yml (and its nginx/nginx.conf): when activated, it published ports 80/443 on all host interfaces and proxied straight to the default-unauthenticated MCP backend over the internal Docker network, bypassing the compose file's loopback-only publish; its TLS server block was entirely commented out and no release document referenced the profile. A repository hygiene test now requires every compose publish spec to bind 127.0.0.1.

Fixed

  • The stdio server's async/sync tool-registration mismatch, where awaiting synchronous tool methods raised TypeError at call time
  • Dict-vs-individual-field parameter signature drift on the three create-tools (create_address_object, create_service_object, create_static_route)
  • A broken package entry point (pyproject.toml referenced a non-existent src.main:main)
  • The Dockerfile's stale COPY pytest.ini . line, which broke docker build after pytest.ini was deleted
  • A .dockerignore gap that allowed the gitignored, secrets-bearing config/config.json to be copied into built Docker images

Security

  • Bearer-token HTTP authentication enforcement with constant-time token comparison (SEC-05)
  • Token and password redaction in logs via a handler-level log filter (CONF-03)
  • TLS certificate verification enabled by default for config-loaded devices (SEC-04)
  • Dispatch-layer read-only gate: write and destructive MCP tools are rejected unless explicitly enabled via server.allow_writes or FORTIGATE_MCP_ALLOW_WRITES (SEC-01)
  • FortiGate API token rotated on the live device (Plan 05-01, primary mitigation); the dangling pre-rotation commits, kept alive only by leftover filter-branch backup refs, were removed from local git history via refs/original cleanup and git gc (Plan 05-04, secondary hygiene), without ever fetching from or altering origin/main (REL-06). GitHub's server-side cache of the pre-rotation commit may still resolve it by direct SHA indefinitely (documented, accepted residual risk).