v0.2.0

What's Changed

Added

• ai-config sync command: writes the unified configuration back to each
  provider's native files (export direction; migrate is the import direction)
  • --dry-run previews every planned write
  • --provider <name> (repeatable) targets specific providers
  • --all syncs every known provider, even undetected ones
  • --project syncs project-level targets (.cursor/, .vscode/,
    .github/copilot-instructions.md, CLAUDE.md, AGENTS.md, .rules)
  • --prune makes provider MCP lists an exact mirror of the unified config,
    backing up files before removing servers
• Per-provider MCP scoping via providers.<name>.mcp:
  {"include": [...]} or {"exclude": [...]}
• Provider coverage expanded from 3 to 9: added Cursor, Codex CLI, Gemini CLI,
  VS Code (Copilot), Zed, and Continue alongside Devin, Windsurf, and Claude
• Multi-format config support: JSON, TOML (config.toml), and YAML
  (config.yaml) are read and written where providers use them
• Per-provider MCP shape translation: mcpServers (most), servers with
  type (VS Code), context_servers (Zed, stdio only), TOML
  [mcp_servers.*] (Codex), YAML list (Continue)
• Rules sync: unified AGENTS.md lands as a managed block
  (<!-- BEGIN ai-config managed -->) in shared files, or as dedicated rule
  files (.cursor/rules/*.mdc, .continue/rules/*.md, Windsurf global rules)
• Windsurf coverage spans the Codeium-era paths (~/.codeium/) and current
  ~/.windsurf/ paths; Devin CLI covers ~/.config/devin/ and ~/.devin/
• MCP migration dedupes identical servers and keeps conflicting same-name
  servers under a <name>.<provider> alias instead of silently overwriting
• "disabled": true flag on unified MCP server entries: disabled servers
  stay in the store as dormant inventory and are never exported by sync,
  even when named in a provider's include list
• tests/ suite: 43 tests covering formats, provider translation, sync,
  migration, and project-local config

Fixed

• AgentEnv.project_config was declared @property but accepts a cwd
  argument, so get_merged_config(cwd=) (the documented provider integration
  call) always raised TypeError
• ProjectConfig.enable_feature overwrote the enabled_features list with
  each call, and disable_feature wrote to a disabled_features key that
  nothing read - both now correctly maintain enabled_features
• All outstanding mypy errors in project.py, config.py, and migration.py
• Project local MCP overrides were unusable: save_mcp_config(local=True)
  wrote .ai/mcp-config.local.json while get_mcp_config read
  .ai/mcp_config.local.json (and init-project gitignored the latter).
  Everything now uses the hyphenated mcp-config.local.json
• ai-config validate counted skills in ~/.agents/config/skills/ instead
  of the actual ~/.agents/skills/ directory

Changed

• BREAKING: ai-config sync no longer writes to every detected provider
  by default. Export is opt-in via "providers.<name>.sync": true, or
  explicit --provider / --all flags. Global config is now the
  store-of-truth; providers read it directly or receive explicit exports.
  This prevents credentials in MCP env from propagating to every installed
  tool
• Migration dedupes across provider naming schemes (e.g. VS Code registry
  names like io.github.X/foo-mcp match foo), normalizes empty injected
  fields on compare, and reuses existing aliases instead of minting .2/.3
  suffixes on re-import
• ProviderMigrator.PROVIDER_PATHS now derives from the provider registry in
  providers.py (single source for read paths and sync targets)
• Unified MCP store key renamed to context_servers in
  config/mcp-config.json and the loaded unified config; the legacy
  mcpServers key is still accepted on read
• New dependencies: tomli (Python <3.11), tomli-w, pyyaml