Plannotator embeds an AI chat sidebar directly in the code review UI. You can select lines in a diff, ask questions, and get streaming responses. The AI sees the full diff context automatically, so you can ask things like “explain this change” or “is this safe?” without copy-pasting code.

Supported providers

Claude (via Claude Agent SDK)

Requires the claude CLI installed and authenticated. Uses Claude Code’s full system prompt, so the AI has the same capabilities as a Claude Code session — file reading, search, web access — plus the diff context.

Models:

• Sonnet 4.6 (default)
• Opus 4.6
• Haiku 4.5

Codex (via Codex SDK)

Requires the codex CLI installed and authenticated. The AI operates in a sandboxed read-only mode with the diff context injected as a system prompt prefix.

Models:

• GPT-5.4 (default)
• GPT-5.4 Mini
• GPT-5.3 Codex
• GPT-5.3 Codex Spark
• GPT-5.2 Codex
• GPT-5.2

Pi (via RPC subprocess)

Requires the pi CLI installed and configured. Plannotator spawns pi --mode rpc and communicates over JSONL/stdio. Models are discovered dynamically from your Pi installation — whatever models you’ve configured in Pi are available here.

No API keys are managed by Plannotator — Pi uses its own local configuration.

OpenCode (via OpenCode SDK)

Requires the opencode CLI installed and authenticated. Plannotator spawns opencode serve and communicates via HTTP + SSE. Models are discovered dynamically from your connected providers.

OpenCode supports session forking, resuming, and runtime permission approvals — the richest capability set of all four providers.

Configuration

Provider and model selection is available in Settings > AI. These persist via cookies across sessions.

You can also override the provider and model per-session using the config bar at the bottom of the AI sidebar. Changing the provider or model starts a new session — old messages stay visible but the conversation resets.

How it works

A session is created lazily on your first question. Until then, no resources are used.

Claude sessions use { preset: "claude_code" } with the review context appended. This means the AI has full Claude Code capabilities (tool use, file reading, search) plus the diff. If the code review was launched from a Claude Code session, the AI can fork from the parent session, preserving conversation history.

Codex sessions inject the review context as a system prompt prefix. The AI has Codex’s built-in capabilities plus the diff. Codex sessions are always standalone — fork support is not available.

Pi sessions inject the review context as a system prompt prefix, similar to Codex. Pi uses its full default toolset (read, bash, edit, write). Pi sessions are always standalone — fork and resume are not available.

OpenCode sessions pass the review context via the system field on the prompt API. OpenCode supports forking from a parent session and resuming previous sessions. Permission requests work the same as Claude — approval cards appear inline.

Diff context handling: Large diffs are truncated at roughly 40k characters to stay within context limits. However, when you select specific lines and ask a question, the selected code is always sent alongside the question regardless of truncation.

Permission requests

When using Claude, the AI may request permission to use tools like Read, Glob, Grep, or WebSearch. When this happens, an approval card appears inline in the chat. You can approve or deny each request individually.

Codex sessions run in a sandboxed read-only mode, so permission requests do not apply.

OpenCode supports the same permission approval flow as Claude — tool calls that need approval appear as inline cards. You can approve or deny each request.

Pi does not expose a permission approval gate over RPC, so tool execution is handled entirely by Pi’s own runtime.

Reasoning effort

Codex supports a reasoning effort setting with four levels: Low, Medium, High, and Max. This is available in the config bar at the bottom of the AI sidebar. Higher effort means slower but more thorough responses.

This setting only applies to Codex — Claude, Pi, and OpenCode do not expose a reasoning effort control.

Available settings