Agent Support

Agent Note uses the same git-note format for every agent, but each agent exposes different hooks and transcript data. This page helps you choose what to expect before you enable it for a team.

Capability Matrix

Agent	Status	Prompt	Response	Files	AI Ratio	Line Estimate
Claude Code	Full support	Yes	Yes	Yes	Yes	Default
Codex CLI	Supported	Yes	Yes	Yes	Yes	When Codex patch history matches the final commit
Cursor	Supported	Yes	Yes	Yes	Yes	When edit counts match and the final file still matches the latest AI edit
Gemini CLI	Preview	Yes	Yes	Yes	Yes	Not yet

Note

“Files” means Agent Note can show which committed files were touched by the agent. “Line Estimate” means it can also estimate AI-written lines instead of only counting files. “Preview” means Agent Note supports the current public behavior, but the agent’s local hook or transcript format may still change.

Recommended Setup

npx agent-note init --agent claude

Use one --agent value for each agent your team actually uses. Multiple agents can be enabled in one repository:

npx agent-note init --agent claude codex cursor gemini

Each developer still needs to run init once after cloning because git hooks live inside .git/.

Choose By Expectation

You need…	Expect…
The most complete line attribution today	Claude Code
Codex CLI support with conservative attribution	Codex CLI
Cursor editor history and response recovery	Cursor
Gemini CLI prompt / response capture	Gemini CLI

How Attribution Works

Line-level

Agent Note can count AI-added lines when it has enough edit history to compare the AI edit, later human edits, and the final commit diff.

File-level

When line-level proof is not safe, Agent Note records whether the committed file was touched by the agent and calculates the ratio by files.

Line-level is intentionally conservative. If the evidence does not line up, Agent Note falls back to file-level or skips uncertain data instead of guessing.

Agent Notes

Claude Code

Claude Code provides hook-native prompt, response, and edit data. Agent Note records pre-edit and post-edit file states, then subtracts later human edits from the final diff.

Codex CLI

Codex CLI support is transcript-driven. Agent Note reads local transcripts, connects apply_patch operations to committed files, and upgrades to line-level only when patch counts match the final diff.

Shell-only edits do not create per-prompt files_touched from command text. When the current Codex transcript is trusted, Agent Note can still mark the commit files as AI-assisted at commit-level.

Cursor

Cursor support uses afterFileEdit / afterTabFileEdit hooks plus response hooks or local transcripts. Line-level attribution is possible only when edit counts match and the committed file still matches the latest AI edit.

Gemini CLI

Gemini CLI support uses BeforeTool / AfterTool for file edits and BeforeAgent / AfterAgent for prompt and response recovery. Attribution currently stays at file-level.

When Data Is Missing

Missing data usually means Agent Note chose safety over a guess.

• No note is written when Agent Note cannot find a committed file touched by the agent.
• A prompt can be omitted when it belongs to an older commit window.
• A response can be missing when the agent did not expose it through hooks or transcripts.
• Generated artifacts and paths listed in the repository-root .agentnoteignore stay visible in Notes, PR Report, and Dashboard; they are only excluded from the AI Ratio denominator.

If you need to inspect a specific case, start with Troubleshooting.