refactor: AI features cleanup wave 1 (providers, ViewModel, MCP, Copilot)

[datlechin]

Summary

Wave 1 of the AI features audit (see AI_FEATURES_AUDIT.md). 13 commits across 8 focused refactors. Production builds clean, AI test target runs.

What changed

• JSONValue dedup: deleted Core/AI/Chat/JSONValue.swift and ChatToolJSONFormatter.swift. All chat tools, providers, and the MCP wire now share JsonValue. Encoding helpers live as native methods on the type.
• Composer focus border: removed the animated multi-layer angular gradient (IntelligenceFocusBorder + IntelligenceShimmer). Replaced with a single 1pt Color.accentColor stroke, matching macOS HIG.
• AIChatViewModel split: 1249-line god class decomposed into +Streaming, +SchemaContext, +ToolApproval, +MessageEditing, +SlashCommands, +Persistence. Main file now 293 lines. Replaced 3-boolean state machine (isStreaming/lastMessageFailed/pendingApproval) with a single StreamingState enum. Added re-entry guard on startStreaming(). Removed errorMessage didSet side effect, added explicit clearError().
• Provider extracts: new String+AIEndpoint.swift (normalizedEndpoint()), new JSONValue+Encoding.swift. Anthropic, OpenAI-compatible, and Gemini all share endpoint normalization, JSON encoding, and a unified mapHTTPError (Gemini's 403 special case folded into the base via parameter). fetchAvailableModels standardized at 5s timeout.
• Inline suggestion: replaced Timer debounce with Task<Void, Never>? + Task.sleep. Added sourceIdentity: ObjectIdentifier to InlineSuggestionSource; manager verifies identity before applying so source switches (Copilot ↔ AI Chat) cannot apply stale results. Lifted acceptCommand out of the shared InlineSuggestion struct entirely; Copilot tracks LSP commands in its own private [UUID: LSPCommand] map keyed by suggestion id. Debounce duration moved to AISettings.inlineSuggestionDebounceMs (default 500ms, clamped 50-5000).
• MCP transport: split MCPHttpServerTransport.swift into MCPHttpConnectionContext, MCPHttpRequestRouter, MCPSseWriter. Migrated PairingExchangeStore from @unchecked Sendable + NSLock to actor. Tightened MCPAuthPolicy.sessionApprovals to private. Added : keep-alive SSE heartbeat every 30s.
• Storage versioning: AIConversation gained schemaVersion: Int with backward-compatible decode. AISettings.includeSchema / includeCurrentQuery defaults are now consistent across regular init, init(from:), and default. CustomSlashCommandStorage.add() rejects duplicate names with a typed error surfaced via alert.
• Copilot lifecycle: CopilotService.stop() wraps shutdown() RPC in a 10s withTaskGroup race. LSPTransport.stop() is now async, cancels the read Task, and explicitly closes stdin/stdout FileHandles before terminating the process. The read loop uses FileHandle.bytes AsyncSequence. Renamed CopilotSchemaContext → CopilotPreambleBuilder. CopilotBinaryManager strips the com.apple.quarantine xattr after SHA-512 verify.

Files

• 64 files changed, +2575 / -2291
• Main file shrinkage: AIChatViewModel.swift 1249 → 293 lines

Test plan

• xcodebuild build green for the production target
• swiftlint lint --strict clean on every touched file
• AI test classes pass (Conversation, Settings, ChatToolRegistry, ChatToolArgumentDecoder, AIChatViewModel Slash/Mention/Action subsets)
• Full test target run (3 pre-existing test infra failures: fixError* for MongoDB/Redis assume plugin metadata loads in test target; OpenTableTabTests hits a pre-existing assertion in MainContentCoordinator+FilterState.swift:311. Out of scope for this PR.)
• Manual smoke: open AI chat, send a message, verify streaming, accept inline suggestion in editor, exercise tool approval

Out of scope (Wave 2)

• ChatToolSchemaBuilder (feat: implement tab persistence and last query memory #5 in roadmap)
• Chat view HIG fixes (Fix/bug #6 in roadmap)
• Provider polish: refresh known model lists, log Ollama detection error category (Add table delete/truncate confirmation dialog #11 in roadmap)
• MentionDetector Unicode + ChatToolRegistry per-tool mode (feat: Native macOS inline error banners (replace blocking dialogs) #12 in roadmap)

See section 4 of AI_FEATURES_AUDIT.md for the full backlog.

[chatgpt-codex-connector]

You have reached your Codex usage limits for code reviews. You can see your limits in the Codex usage dashboard.