Skip to main content

Agent Identity for Git and GitHub

Date: 2026-03-01 Status: Research — no action taken yet

Problem​

Almost all SecurityV0 development is performed by Claude Code with a human in the loop. As autonomous bot development grows (e.g., via Telegram on Mac mini server, --dangerously-skip-permissions), the need to distinguish human vs agent activity increases.

Two dimensions:

  1. GitHub Issues/PRs — who created the task?
  2. Git commits — who authored the code?

Current Setup​

  • Same GitHub account (Ivan-Fn) for both human and agent work
  • claude-code label on agent-created issues (configured in AGENTS.md + UserPromptSubmit hook)
  • No commit-level distinction — git log shows all commits as the same author

This works for a small team where the developer knows which sessions are agentic.

Industry Approaches (as of March 2026)​

1. Separate Bot Account​

Create a dedicated GitHub account (e.g., sv0-bot):

  • Git commits use bot identity via env vars (no permanent config change): 
    GIT_AUTHOR_NAME="sv0-bot"
    GIT_AUTHOR_EMAIL="sv0-bot@users.noreply.github.com"
    GIT_COMMITTER_NAME="sv0-bot"
    GIT_COMMITTER_EMAIL="sv0-bot@users.noreply.github.com"
  • GitHub Issues/PRs created by the bot account — immediately visible as non-human
  • Dedicated SSH key scoped to specific repos
  • CODEOWNERS can require human review on bot commits
  • Filterable: git log --author="sv0-bot", gh search issues --author sv0-bot

Pros: Clean attribution, git-level filtering, CODEOWNERS integration, audit-friendly Cons: Extra account to manage, SSH key setup, free GitHub accounts have limited private repo access

Reference: Agent Identity for Git Commits (dev.to)

2. Same Account + Labels/Tags​

Keep personal account, use labels and conventions to distinguish:

  • claude-code label on issues
  • Draft PRs for agent work
  • Review gates enforce human sign-off

Pros: Simple, no extra account, works today Cons: No commit-level distinction, git log can't filter, relies on process discipline

Reference: GitHub Community Discussion

3. GitHub Copilot's Native Approach​

GitHub's Copilot coding agent runs as copilot-swe-agent[bot] — a first-class bot identity. All commits, PRs, and issues are visibly from the bot. This is the direction GitHub is heading for their own tooling.

Reference: GitHub Copilot Coding Agent

4. Git AI (Line-Level Attribution)​

git-ai is an open-source extension that provides line-by-line attribution of human vs AI code within the same commit. More granular than account separation but adds tooling complexity.

Proposed Tiered Model for SecurityV0​

ModeIdentitySupervisionExample
Interactive (human in the loop)Personal account + claude-code labelHuman approves each actionClaude Code terminal sessions
Autonomous (minimal supervision)Bot account (sv0-bot)Pre-approved, runs unattendedTelegram bot, --dangerously-skip-permissions, Mac mini server

Rationale​

  • Interactive sessions: the human IS the author — they're directing the work, reviewing output, approving commits. Labels are sufficient to note the tool used.
  • Autonomous sessions: the bot IS the author — no human reviewed before commit/push. A separate identity makes this visible in git history and GitHub UI, and allows CODEOWNERS rules to enforce post-hoc review.

Implementation (when ready)​

  1. Create GitHub account sv0-bot (or similar)
  2. Generate SSH key, add to bot account, grant collaborator access to SecurityV0 repos
  3. Configure autonomous agent environment: 
    export GIT_AUTHOR_NAME="sv0-bot"
    export GIT_AUTHOR_EMAIL="sv0-bot@users.noreply.github.com"
    export GIT_COMMITTER_NAME="sv0-bot"
    export GIT_COMMITTER_EMAIL="sv0-bot@users.noreply.github.com"
    export GIT_SSH_COMMAND="ssh -i ~/.ssh/sv0_bot_key -o IdentitiesOnly=yes"
  4. Add CODEOWNERS rule requiring human review on bot PRs
  5. Keep interactive sessions on personal account with claude-code label (no change)

Decision​

No action for now. Current label-based approach is sufficient for interactive development. Revisit when autonomous bot development (Telegram/Mac mini) reaches production use.

Sources​