AI Chat Todo

AI Chat Todo is a preview VS Code extension that turns action items from Codex and Claude Code chats into a focused, workspace-private todo list.

It is built for the moments when an assistant asks you to do something outside the chat, such as "go update the Play Store Console release settings", "create a GitHub secret", "configure Firebase", or "authenticate the deployment CLI". The extension captures those user-action todos, explains the service and tooling involved, checks setup with read-only commands, and builds a handoff prompt you can give back to Codex or Claude Code when the tool is ready.

What It Does

• Captures user-action todos from recent Codex sessions in trusted workspaces.
• Lets you import copied assistant text from the clipboard, including Claude Code clipboard imports that preserve Claude as the source provider.
• Shows pending candidates in the AI Chat Todo Activity Bar view.
• Opens a Todo Workbench with source context, setup guidance, tool recommendations, and a generated assistant prompt.
• Explains "Why Captured?" for pending-review candidates with deterministic rule ids, source metadata, workspace evidence, and tool-match reasons.
• Lets you add workspace-private ignore rules for recurring false-positive candidate text.
• Shows when todos were added using your local timezone and preserves nearby Codex conversation context when available.
• Recommends CLIs such as gh, vsce, fastlane supply, firebase, gcloud, clasp, GAM/Google Workspace admin tooling, wrangler, flyctl, Rails/Bundler, kamal, caddy, npm, twine, gem, dotnet, Gradle/Maven, docker, vercel, stripe, and aws.
• Uses read-only workspace evidence such as firebase.json, .github/workflows, vercel.json, .xcodeproj, *.gradle.kts, wrangler.toml, Cloudflare Pages _redirects, fly.toml, appsscript.json, Rails binstubs/config, Kamal config, Caddyfiles, package.json, pyproject.toml, .gemspec, .sln/project files, Dockerfiles, Compose files, Google service files, and package metadata to explain recommendations.
• Contributes read-only VS Code agent-mode tools so compatible assistants can list AI Chat Todo records, recommend tooling for a todo, and fetch the centralized handoff prompt.
• Contributes a bundled read-only MCP stdio server with list_todos, recommend_tooling, and get_handoff_prompt tools for MCP-capable assistants.
• Contributes a lightweight local @todo chat participant for /blocked, /review, /handoff, and /tooling questions.
• Offers explicit clipboard language-model analysis for better titles, ambiguous snippets, and missing-registry suggestions.
• Tracks multi-step user-action processes from ordered lists, checklists, and assistant "remaining steps" sections.
• Runs read-only setup checks and marks accepted todos Ready or Blocked Setup.
• Can run a selected read-only setup check as a VS Code task using process arguments instead of shell-interpolated commands.
• Exports/imports workspace todo state and opens registry diagnostics for troubleshooting.
• Shows a Todo view badge and status bar indicator when captured todos need review.
• Opens Codex or Claude Code where supported and copies the generated handoff prompt.

Getting Started

1. Install the OpenAI Codex VS Code extension (openai.chatgpt) or Anthropic Claude Code extension (anthropic.claude-code).
2. Open a trusted workspace in VS Code.
3. Open the AI Chat Todo Activity Bar view.
4. Choose Scan recent Codex sessions or Import todo from clipboard from the empty Todo view.
5. Review captured candidates and accept or dismiss them.
6. For accepted todos, run setup checks and follow the install/auth guidance.
7. Use Ask Assistant to Help to copy a handoff prompt and open the matching assistant.

Automatic trusted-workspace session scanning is enabled by default. You can tune it with:

• codexTodoCompanion.autoScanCodexSessions
• codexTodoCompanion.autoScanIntervalSeconds
• codexTodoCompanion.autoScanRecentSessionLimit

Manual commands are also available:

• AI Chat Todo: Scan Recent Codex Sessions
• AI Chat Todo: Import Todo From Clipboard
• AI Chat Todo: Import Claude Todo From Clipboard
• AI Chat Todo: Review Captured Todos
• AI Chat Todo: Clear Pending Review
• AI Chat Todo: Ask Assistant to Help
• AI Chat Todo: Ignore Similar Candidates
• AI Chat Todo: Run Setup Check as Task
• AI Chat Todo: Export Todo State
• AI Chat Todo: Import Todo State
• AI Chat Todo: Show Registry Diagnostics
• AI Chat Todo: Analyze Clipboard With Language Model

Preview Limits

This preview focuses on the OpenAI Codex VS Code extension for automatic session capture, with initial Anthropic Claude Code support for clipboard import, opening Claude, and handoff prompt staging.

Not built yet:

• Codex lifecycle hook capture beyond trusted-workspace session polling.
• Automatic Claude Code transcript/session scanning.
• Exact assistant source-thread navigation beyond Codex/sidebar and line-aware transcript fallback.
• Additional assistant adapters for Gemini Code Assist, GitHub Copilot Chat, Cursor, Windsurf, Cline, Kilo Code, Continue, Augment, or similar VS Code AI workflows.

Assistant-specific behavior lives behind adapter interfaces so future Anthropic, Google, GitHub, Cursor, Windsurf, or other VS Code AI add-ons can be supported without rewriting the todo, registry, prompt, or UI core. The buildout list is tracked in docs/assistant-addons.md.

Privacy And Trust

AI Chat Todo stores todo state locally in workspace-private VS Code extension storage, backed by a JSON state file when available so the bundled MCP server can read the same store. It saves todo titles, Markdown bodies, parsed user-action steps, short assistant excerpts, nearby Codex conversation excerpts for captured todos, source metadata, setup-check results, and tool recommendation metadata. Handoff prompts are generated on demand from stored todo metadata.

It does not store service tokens, API keys, service-account files, refresh tokens, passwords, cookies, hidden CLI auth files, or full transcripts.

Automatic Codex session scanning only runs in trusted workspaces, filters sessions to the active workspace, and uses local transcript excerpts for todo detection. The extension does not upload transcript content, workspace evidence, or command output to external services.

The extension declares limited untrusted-workspace support. Existing todo UI can load in untrusted workspaces, but Codex session scanning, workspace evidence collection, and setup checks require workspace trust.

Generated install commands are shown as user-reviewed data. The extension does not execute installs automatically. Setup checks are read-only by policy, and state-changing service actions such as deploys, publishing, secret writes, billing changes, and production mutations require explicit user confirmation outside this extension.

The VS Code language model tools and bundled MCP tools are also read-only. They expose todo summaries, registry recommendations, and generated handoff prompts to compatible assistants, but they do not create todos, run setup checks, execute commands, mutate todo status, or send prompts automatically.

The Analyze Clipboard With Language Model command is different from background capture: it only runs after you explicitly choose it, warns that the clipboard excerpt will be sent to the selected VS Code language model provider, and opens the analysis in a temporary Markdown document. It is not used by automatic scanning.

Development

Useful commands:

npm run compile
npm run test:unit
npm run test:vscode
npm run check
npm audit --omit=optional
npm run package
npm run release:check

Local development and release-candidate testing are documented in docs/local-development.md. The security model is documented in docs/security.md, and release planning is documented in docs/release.md.

Release Status

This extension is marked preview: true while capture formats, registry scoring, and Codex linkage stabilize. Publishing should only happen from a clean, verified tree after packaging and testing a local VSIX.