dsh-git-ai

Once an AI agent starts editing your codebase, "who actually wrote this line?" becomes a real question. Which lines came from a model, which model, in which session — that context lives scattered across logs and terminal history, invisible the moment you look at a diff in the editor.

dsh-git-ai hooks into the Harness tools/pre-execute and tools/post-execute interception points and translates every write, edit, str_replace_editor call, as well as bash and pwsh invocations, into git-ai agent-v1 checkpoints written to git notes via stdin. Inside VS Code, hovering over any line now reveals the author (agent), model name, and session ID. git-ai additionally reports ai_additions and ai_accepted per tool-model pair, giving you a quantified picture of how much of the diff is AI-authored. Read-only str_replace_editor view operations are skipped automatically; failed edits emit no checkpoint while failed bash commands still close their before/after snapshot pair.

The plugin treats git-ai as a fully optional external dependency. A missing binary, a timeout, a non-zero exit, or an executor rejection is logged as a warning and never propagated — the tool call and the agent turn proceed exactly as if the plugin were not installed.

Built for engineers who need to audit AI-generated code, distinguish model output from human edits, and for teams that want to answer "did the agent write this or did a person?" at a glance during code review.