[ENHANCEMENT] Support Git AI line-level attribution of AI-generated code

[cesaregerbino]

Is your feature request related to a problem? Please describe.

Zoo Code has no way to record which lines of code it wrote, so teams cannot reliably measure how much AI-generated code ends up in their repositories. This matters for organizations that need AI-code attribution for governance, auditing or ROI measurement.

Git AI is an open-source Git extension that has become a common standard for this: it stores line-level attribution in Git Notes, linking each AI-written line to the agent, model and session that produced it. It already supports Claude Code, Codex, GitHub Copilot, Cursor, Continue, OpenCode, Windsurf, Gemini CLI and others, but not Zoo Code. Teams that standardized on Zoo Code are therefore left with switching agents or maintaining a private fork.

Describe the solution you'd like

When Git AI is installed on the user's machine, Zoo Code reports its file edits to it automatically, with no manual steps for the user:

• After a normal git commit, git ai blame and git ai stats show which lines were written by Zoo Code, with model and session.
• Lines edited by a human after the AI edit are attributed to the human.
• When Git AI is not installed, Zoo Code behaves exactly as today (no errors, prompts or performance changes).
• It works with any provider/model configured in Zoo Code (including Azure OpenAI and local OpenAI-compatible models), since Git AI only records the model name as metadata.

Git AI documents a generic integration path, the agent-v1 checkpoint preset (docs):

1. On startup, detect whether git-ai is available on PATH.
2. Right before a file-editing tool runs, call git ai checkpoint agent-v1 --hook-input stdin with type human (optionally with the paths about to be edited), so pending human changes are recorded as human.
3. Right after the edit, call it again with type ai_agent, passing agent name, model, conversation id, edited file paths and the session transcript (user/assistant/tool_use messages, without tool results).

Zoo Code already tracks tasks, models and file edits in its agent loop, so the required data should already be available. Git AI's guidelines ask integrations to work cross-platform and cross-shell and not to depend on extra tools such as jq or node. The Git AI maintainers also offer to help agent authors with integrations.

Describe alternatives you've considered

• Commit trailers (e.g. Co-authored-by) via custom rules: depends on model compliance and misses commits made manually by the user.
• Logging at the LLM gateway: measures usage and cost, not the code that actually lands in the repository.
• Switching to an agent already supported by Git AI: possible, but we would rather keep using Zoo Code.
• Maintaining a private fork: we would much rather see this upstream, so we can keep following Zoo Code releases.

Additional context

• Briefly discussed on the Zoo Code Discord (#general, 30 Sep 2026), where a maintainer replied: "We'd love a PR on this!"
• Possible trade-offs: an extra process call around each edit (passing edited file paths keeps it fast in large repos); the session transcript is sent to Git AI locally (Git AI redacts secrets and stores prompts outside the repository), so a setting to disable the integration may be desirable; files created via terminal commands may need separate handling, as in other agents' integrations.