feat: run Electron chat through local Claude and Codex

[NarwhalChen]

Summary

• replace the Electron app Foxychat service dependency with a local Main-process AI runtime
• stream chat through a context-isolated preload and IPC bridge to authenticated Claude Code and Codex CLI subscriptions
• expose Main-process MCP and builtin tools to local providers with request-scoped user approval
• add builtin:computer_control for atomic screenshot, pointer, keyboard, scroll, drag, and wait actions
• preserve screenshot image content through a loopback-only authenticated Codex MCP transport
• add a General Settings selector for Claude Code and Codex; the selection uses the existing local persisted model state consumed by new chats
• keep chat history and provider selection local; ~/proj/foxychat-server remains unchanged
• retain the atomic WDIO/MCP automation as an AI-operated local testing method; no CI workflow is added by this work

Runtime and safety boundaries

• renderer input is validated and size-bounded at IPC; events are isolated by sender and request ID
• Main owns provider discovery, trusted cwd selection, lifecycle, abort, and disposal
• all state-changing computer_control calls require the existing in-app approval flow
• screenshot checks macOS Screen Recording permission; pointer and keyboard actions check Accessibility permission and return actionable errors when denied
• Codex MCP listens only on loopback, uses a random bearer token, and preserves MCP image blocks rather than serializing them to text
• background automation profiles remain hidden and isolated unless show_window is explicitly enabled

Verification

• pnpm --filter @convera/app test — 20 files / 60 tests passed
• pnpm --filter @convera/app lint — 0 errors (7 existing React Hook warnings)
• pnpm --filter @convera/app exec tsc --noEmit --skipLibCheck
• pnpm --filter @convera/app automation:typecheck
• pnpm automation:prepare — Electron/Vite package and matching native automation dependencies prepared
• real hidden Electron profile selected Codex in Settings and retained aria-pressed=true after relaunch
• real Codex subscription selected builtin:computer_control, displayed the in-app approval prompt, completed wait(1ms), and returned CODEX_CUA_PASS
• Codex MCP integration test verifies PNG image blocks survive the transport and unauthenticated callers are rejected

Deliberate limitations

• no new Claude model or Claude CUA run was performed for this update
• this Mac currently denies Screen Recording to the Electron development host, so a real screenshot cannot be captured until the user grants that OS permission; the image transport itself is covered by integration tests
• attachments remain hidden and unsupported in this local streaming slice
• Codex uses the Main-process working directory for now
• full TypeScript checking without skipLibCheck still reports upstream third-party declaration errors; app source and automation typechecks pass
• packages/website and ~/proj/foxychat-server are unchanged

[vercel]

The latest updates on your projects. Learn more about Vercel for GitHub.

2 Skipped Deployments

Project	Deployment	Actions	Updated (UTC)
foxyfox-docs	Skipped		Jul 30, 2026 1:19pm
foxyfox-website	Skipped		Jul 30, 2026 1:19pm

[vercel]

@NarwhalChen must be a member of the codefox-repo team on Vercel to deploy.
- Click here to add @NarwhalChen to the team.
- If you initiated this build, request access.

Learn more about collaboration on Vercel and other options here.

[NarwhalChen]

Goal update after rebasing onto origin/main f751ac0:

• Remote head: ad928d2
• Electron local-first agent path now exposes builtin and connected MCP tools to Claude/Codex while keeping execution and interaction in the main process.
• Codex root cause fixed: removed the obsolete features.rmcp_client flag and kept native mcp_servers.convera HTTP configuration; MCP tool-call elicitation now returns structured empty content.
• Regression coverage: Vitest 15 files / 49 tests passed; lint 0 errors (7 existing hook warnings); Prettier, automation typecheck, and git diff check passed.
• Real Electron E2E passed with a fresh isolated profile and hidden/unfocused window: Codex invoked builtin ask_user_input, the UI presented Alpha/Beta, automation selected Alpha, and Codex completed with CODEX_TOOL_OK: Alpha. The session and Chromedriver were closed cleanly.

No merge performed.

[NarwhalChen]

Added commit bf2c1c3 to complete Codex tool-call display through the AI SDK UI stream.

What changed:

• Electron Main now forwards AI SDK UIMessageChunk events instead of rebuilding a parallel text/tool protocol.
• Renderer aggregates chunks with readUIMessageStream, renders native/dynamic tool states through the existing tool UI, and persists parts in IndexedDB.
• Tool name qualification remains intact for injected Convera tools while native Codex tools such as exec pass through unchanged.

Verification:

• pnpm --filter app test: 16 files / 50 tests passed
• pnpm --filter app automation:typecheck: passed
• full Prettier and ESLint checks passed (7 existing hook warnings, no errors)
• pnpm automation:prepare: packaged Electron and driver successfully
• real hidden Electron run with gpt-5.3-codex-spark: Codex called exec, UI displayed arguments/result, output was CODEX_TOOL_UI_OK, exit code 0, final reply DONE; persisted part was dynamic-tool / output-available with input and output

[NarwhalChen]

Follow-up commit 4d0c160 removes the legacy tool-call compatibility layer entirely.

• deleted legacy ToolInvocation, tool-invocation, and toolInvocations types/fields/rendering/persistence
• generic and custom tool renderers now consume AI SDK ToolUIPart | DynamicToolUIPart directly
• repository search confirms no legacy tool-call identifiers remain
• 17 test files / 52 tests passed; lint, format, automation typecheck, and Electron package passed
• real hidden Codex run passed after the removal: native exec rendered AI_SDK_ONLY_OK, exit code 0, final DONE; IndexedDB stored only AI SDK parts and no legacy field

[NarwhalChen]

Superseded by the reviewable split merged through #208, #209, and #210. The complete #206 tree is now on main through those three PRs.

[NarwhalChen]

Implementation refresh for the local-first Electron target is available on kobe/local-first-electron-backend at ad928d278d204b854dae27471bb77417a20999cb (rebased onto origin/main f751ac093aebafd3a95c8a08f3126238582133a4).

Scope:

• Removes Electron renderer dependencies on hosted auth/chat service paths while preserving local data.
• Runs Claude Code and Codex subscriptions through an Electron main-process AI SDK runtime and IPC bridge.
• Exposes controlled builtin and configured MCP tools to local agents, with renderer-mediated input/approval interactions.
• Keeps background automation parameterized and isolated per agent/session.

Root cause fixed in the final Codex tool path:

• Codex 0.144.6/0.146.0 rejects the obsolete features.rmcp_client override.
• The adapter now uses native HTTP MCP server configuration without rmcpClient, and returns structured empty content for accepted MCP elicitation.
• The automation click driver now returns the pre-click element snapshot so successful clicks remain successful when a popover option immediately unmounts.

Verification:

• pnpm --filter @convera/app test: 15 files, 49 tests passed.
• pnpm --filter @convera/app lint: 0 errors; 7 pre-existing React Hook warnings.
• pnpm --filter @convera/app format: passed.
• pnpm --filter @convera/app automation:typecheck: passed.
• git diff --check origin/main...HEAD: passed.
• Tracked-file audit found no .env, secrets, automation screenshots/logs, or generated build artifacts.

Real hidden Electron E2E evidence (fresh isolated profile goal-codex-ui-final-20260730-1758, session 302d3e8204cda34c3178c5caf889a3b6):

• Window remained visible=false and focused=false before and after the flow.
• gpt-5.3-codex-spark called builtin:ask_user_input through the local MCP bridge.
• Renderer displayed Choose one with Alpha / Beta; selecting Alpha completed and produced CODEX_TOOL_OK: Alpha.
• Driver cleanup returned running:false; the Chromedriver listener was gone.

PR #206 was already closed when this refresh was published, so this comment records the verified branch update without reopening or merging it.