MCPTap v2.4.0

[pczerkas]

[2.4.0]

Highlights

• Hot-reload for configuration files

Added

• Hot-reload for configuration files — the proxy now polls mtime of
  configuration files (proxy.env, openrouter.env, requesty.env,
  mcp-intercept.yaml, per-model.yaml, use_tool_hook.py) every 2 seconds
  and triggers a selective reload cascade when any of them changes:

  • env files (proxy.env, openrouter.env, requesty.env) → full
    Settings reload + propagation to all dependent components (interceptor,
    per-model config, tool hook);
  • mcp-intercept.yaml → stop old MCP subprocess + start a new
    MCPInterceptor instance;
  • per-model.yaml → reload the per-model config dict;
  • use_tool_hook.py → reload the tool-hook enabled flag and Settings
    (path may change).

  The reloader runs as a background asyncio task inside the aiohttp event
  loop, wired into on_startup/on_cleanup lifecycle in app.py.

• _SettingsProxy with reload_settings() — settings.py was refactored
  so the module-level settings object is a transparent proxy that delegates
  __getattr__/__setattr__ to the current Settings instance. Calling
  reload_settings() rebuilds the instance from env files and swaps the
  proxy target, so all callers automatically see reloaded values without
  re-importing.

• Stale provider env key cleanup — before loading a new provider env
  file, known provider keys (MCP_TAP_API_KEY, MCP_TAP_MODEL,
  MCP_TAP_PLAN_MODE_MODEL, MCP_TAP_OPENROUTER_PROVIDER,
  MCP_TAP_OPENROUTER_DISABLE_PROVIDER_FALLBACKS) are removed from
  os.environ to prevent stale values leaking across provider switches.

• 25 new config-reload tests — tests/test_config_reloader.py (616 lines)
  covering _SettingsProxy delegation/swap semantics, stale env key cleanup,
  mtime detection, selective reload cascade routing, callback failures,
  lifecycle wiring, and application-level reload callbacks
  (reload_per_model_config, reload_tool_hook, reload_intercept,
  reload_env_and_propagate).

• markdownlint pre-commit hook — added a markdownlint hook to
  .pre-commit-config.yaml, wrapped by the shared wrap_hook.sh wrapper.
  The hook runs markdownlint-cli2 on committed Markdown files at the
  pre-commit stage (fail_fast: true, pass_filenames: false).

• .markdownlint-cli2.jsonc — markdownlint CLI configuration that
  disables the MD013 (line length) rule and enables auto-fix mode.

Changed

• settings.py refactored for hot-reload — the monolithic
  Settings.from_env() classmethod was split into _load_env_files() (env
  loading + provider selection + stale key cleanup) and
  _build_settings_from_env() (reads os.environ into the dataclass).
  from_env() now delegates to both, preserving the original entry point.
  The Settings dataclass docstring and module docstring were updated to
  document the hot-reload mechanism.

• Markdownlint directive in changelog-creator skill — added a
  <!-- markdownlint-disable-next-line MD034 --> comment to the example
  invocation blockquote in .agents/skills/changelog-creator/SKILL.md.

Full Changelog

v2.3.0...v2.4.0

---

This discussion was created from the release MCPTap v2.4.0.