BREAKING CHANGE: The [mcp.servers] format was completely incorrect and not
any official Codex format. The only correct format is [mcp_servers] at the
top level of config.toml.
Changes:
- Remove incorrect [mcp.servers] nested table support
- Always use [mcp_servers] top-level table (official Codex format)
- Auto-migrate and cleanup erroneous [mcp.servers] entries on write
- Preserve error-tolerant import for migrating old incorrect configs
- Simplify sync logic by removing format selection branches (~60 lines)
- Update all documentation and tests to reflect correct format
- Add warning logs when detecting and cleaning incorrect format
Backend (Rust):
- mcp.rs: Simplify sync_enabled_to_codex by removing Target enum
- mcp.rs: sync_single_server_to_codex now always uses [mcp_servers]
- mcp.rs: remove_server_from_codex cleans both locations
- mcp.rs: Update import_from_codex comments to clarify format status
- tests: Rename test to sync_enabled_to_codex_migrates_erroneous_*
- tests: Update assertions to verify migration behavior
Frontend (TypeScript):
- tomlUtils.ts: Prioritize [mcp_servers] format in parsing
- tomlUtils.ts: Update error messages to guide correct format
Documentation:
- README.md: Correct MCP format reference to [mcp_servers]
- CLAUDE.md: Add comprehensive format specification with examples
All 79 tests pass. This ensures backward compatibility while enforcing
the correct Codex official standard going forward.
Refs: https://github.com/openai/codex/issues/3441
Fixed an issue where custom/extension fields (e.g., timeout_ms, retry_count)
were silently dropped when editing Codex MCP server configurations in TOML format.
Root cause: The TOML parser functions only extracted known fields (type, command,
args, env, cwd, url, headers), discarding any additional fields during normalization.
Changes:
- mcpServerToToml: Now uses spread operator to copy all fields before stringification
- normalizeServerConfig: Added logic to preserve unknown fields after processing known ones
- Both stdio and http server types now retain custom configuration fields
This fix enables forward compatibility with future MCP protocol extensions and
allows users to add custom configurations without code changes.
- Separate MCP server metadata from connection spec for cleaner architecture
- Add comprehensive server entry fields: name, description, tags, homepage, docs
- Remove legacy format compatibility logic from extract_server_spec
- Implement data validation and filtering in get_servers_snapshot_for
- Add strict id consistency check in upsert_in_config_for
- Enhance import logic with defensive programming for corrupted data
- Simplify frontend by removing normalization logic (moved to backend)
- Improve error messages with contextual information
- Add comprehensive i18n support for new metadata fields
- chore(deps): switch TOML dependency from @iarna/toml to smol-toml and update lockfile
- feat(mcp): add TOML editing/validation for Codex while keeping JSON for Claude; support auto ID extraction from TOML and JSON->TOML conversion for wizard output; add pre-submit required checks (stdio.command / http.url)
- refactor(mcp): unify JSON/TOML validation errors via i18n; add formatTomlError for consistent, localized messages; consolidate state into formConfig/configError
- feat(i18n): add TOML labels/placeholders and error keys (tomlConfig, tomlPlaceholder, tomlInvalid)
- feat(utils): introduce tomlUtils with parse/stringify/validate/convert helpers using smol-toml; provide tomlToMcpServer, mcpServerToToml, extractIdFromToml, validateToml
- build: confirm Vite no longer externalizes Node builtins during build; renderer builds without 'Module 'stream' has been externalized' warning
Jason