gpilot

AI-powered git workflow inside VS Code. Generate commit messages, create pull requests with AI descriptions, run AI code reviews, and produce module specifications — all from a task-oriented sidebar.

The sidebar uses a transparent background and VS Code theme tokens, so it adapts cleanly to any active theme (dark, light, high-contrast). Most actions shell out to the gpilot CLI; spec generation calls the configured AI provider directly using keys read from the OS keychain.

Layout

┌─────────────────────────────────────────┐
│ ● Ready          AI ▢ ▣   Claude … ▾   │  ← Header
├─────────────────────────────────────────┤
│ Commit  Pull Request  PR Review  Spec MD│  ← Tabs
├─────────────────────────────────────────┤
│                                         │
│            (active tab content)         │
│                                         │
├─────────────────────────────────────────┤
│ ● Anthropic connected     Manage Keys   │  ← Footer (pinned)
└─────────────────────────────────────────┘

• Header — live status dot, AI on/off toggle, model selector.
• Tabs — four equal-width tabs; the active tab gets a blue underline.
• Footer — provider connection state on the left, Manage Keys on the right, always pinned to the bottom of the panel.

When AI is off, the tabs and content area are dimmed and a "Enable AI Mode" lock banner is shown above them — every CTA is non-interactive in this state.

Tabs

Commit

• Generate Commit Message — runs gpilot commit --dry-run and drops the suggested Conventional Commit message into an editable textarea.
• Commit — runs git commit -m "<edited>". The extension never auto-commits the AI draft.
• Staged / Changes sections below the CTA, each collapsible.
  • Rows render like VS Code's SCM view: filename in the primary color followed by the dim parent directory that truncates from the start (preserving the leaf folder) when the row is too narrow.
  • The checkbox on each row toggles stage / unstage.
  • Clicking the row opens the matching diff:
    • Staged rows → HEAD ↔ index (git diff --staged), opened via vscode.diff with git: URIs on both sides.
    • Changes rows → HEAD ↔ working tree, opened via vscode.diff with a git: URI on the left and the file URI on the right.
  • Each diff is pinned with workbench.action.keepEditor so multiple diffs can stay open at once.
  • Clicking the same row again closes that specific diff via vscode.window.tabGroups.close(tab).
  • Closing a diff tab in the editor — staged or working — fires tabGroups.onDidChangeTabs, which clears the row's blue highlight in the panel.

Pull Request

When the current branch isn't pushed yet:

• Branch name + "not pushed" pill.
• Push Branch CTA runs git push -u origin HEAD.

After push:

• Generate PR title + description — runs gpilot pr create --dry-run and fills two editable fields.
• Create PR — opens the PR via gh pr create.

PR Review

Header row matches the Commit and Pull Request tabs — ⎇ <branch-name> [PR open] ↗ — the ↗ opens the PR in the browser via gh pr view --web.

Toggle between Manual and Auto modes:

• Manual — Generate Review produces a list of issues. Each card shows severity (blocker / warning / info), the comment text, and a file:line reference. Preview Fix opens the AI's proposed change before applying. Publish Review → posts all comments to the remote.
• Auto — Generate & Publish Review runs the review and publishes in one step, no preview.

Spec MD

Generate a module specification for any file in the repo:

1. Pick a file via VS Code's Quick Pick (the picker truncates long paths and shows the full path on hover).
2. Choose which sections to include — all four (Purpose, API Surface, Usage, Edge cases & errors) are checked by default.
3. Click Generate spec.md to produce <basename>.spec.md next to the source file.
4. The generated file appears inline with an Open ↗ button.

Spec generation calls your configured AI provider directly:

• Anthropic — https://api.anthropic.com/v1/messages
• OpenAI — https://api.openai.com/v1/chat/completions
• Gemini — https://generativelanguage.googleapis.com/v1beta
• Ollama — http://localhost:11434/api/generate

If no key is configured, a deterministic structural parse runs as a fallback so spec generation never silently fails.

AI mode toggle

The header has an AI on/off toggle:

• On — the four tabs are interactive, AI generation is available.
• Off — every tab and CTA is locked behind a dim overlay with a "Enable AI Mode" banner. The mode is persisted across sessions.

Model switcher

Pick the active provider + model from the header dropdown. Supports:

• Claude (Sonnet 4.6, Opus 4.7)
• GPT-4o
• Gemini 2.5 Pro / 2.0 Flash
• Any local Ollama model auto-discovered at http://localhost:11434/api/tags

The selection is written into gpilot.config.yml. The dropdown is width-capped and truncates long labels; the full label is shown on hover.

Live Ollama health

When the active provider is Ollama, the extension polls http://localhost:11434/api/tags and updates the footer accordingly:

• Ollama (local) running — daemon is reachable.
• Ollama not running — daemon is down or unreachable.

Status refreshes every ~2.5 seconds while the panel is open, plus immediately whenever the panel becomes visible.

API key management

• Manage Keys in the footer (and the Configure button on the setup screen) opens a quick-pick that writes secrets to the OS keychain via keytar. No .env files, no plaintext, no VS Code SecretStorage.
• The setup screen is shown until at least one AI key (skip if using Ollama) and one platform token are saved.

Commands

Settings

Requirements

• Node.js 24+
• VS Code 1.85+
• A git repository in the workspace
• gh CLI on PATH if you plan to create PRs from the panel
• API keys for whichever AI provider you select (skip if using Ollama)

Configuration file

The extension expects a gpilot.config.yml at the workspace root. Minimal example:

ai:
  provider: claude
  model: claude-sonnet-4-6
platform:
  type: github
  owner: your-org
  repo: your-repo
mode:
  commit: interactive
  pr_create: interactive
  pr_description: auto
  pr_review: auto
  comment_fix: interactive
review:
  rules:
    - id: no-console-logs
      description: No console.log in production code
      severity: warning
    - id: no-hardcoded-secrets
      description: No hardcoded API keys or tokens
      severity: blocker

How it works

The extension host never imports CLI modules directly. Each user action follows one of three paths:

1. One-shot stdout capture — npx gpilot … --dry-run / --json for commit, PR draft, review, and status. The extension hydrates process.env from the keychain before invoking, so the CLI sees ANTHROPIC_API_KEY / OPENAI_API_KEY / GEMINI_API_KEY / GITHUB_TOKEN etc. without any additional setup.
2. Direct AI call — Spec MD generation calls the configured provider's HTTP API directly, with a structural fallback when no key is available.
3. Terminal spawn — when the user should see streaming output or answer prompts (palette commands, fix flows, gh pr create, git push).

State storage:

• Mode (AI on/off) and first-launch flag — context.globalState
• Secrets — OS keychain via keytar (service: gpilot)
• Model selection — gpilot.config.yml
• Open diff tabs — tracked via vscode.window.tabGroups.onDidChangeTabs so the file-row highlight in the panel stays in sync with VS Code's editor state.

License

MIT