assisted-by

assisted-by

Credit AI in git the way the Linux kernel says to — an Assisted-by: trailer, enforced.

A Claude Code plugin (and portable agent skill) that makes coding agents attribute their help correctly on every commit, following the kernel's coding-assistants policy:

Assisted-by: Claude:claude-opus-4-8

It ships two things:

• git-attribution skill — teaches the agent the rule, so it writes the right trailer on its own.
• attribution-guard.py hook — a Claude Code PreToolUse hook that blocks a git commit if the agent gets it wrong, and tells it how to fix the message.

The rule

• ✅ Credit AI with an Assisted-by: trailer: Assisted-by: AGENT_NAME:MODEL_VERSION [extra-analysis-tools] (e.g. Assisted-by: Claude:claude-opus-4-8).
• 🚫 No Signed-off-by: from the AI — the Signed-off-by certifies the Developer Certificate of Origin, and only a human can do that.
• 🚫 No Co-Authored-By: Claude — that older convention is replaced by Assisted-by:.
• 🚫 Don't list basic tools (git, gcc, editors) — only the AI agent and any specialized analysis tools (coccinelle, sparse, smatch, …).

Install

Claude Code — skill + enforcing hook (recommended)

claude plugin marketplace add bcmyguest/assisted-by
claude plugin install assisted-by@assisted-by

That registers this repo as a plugin marketplace and installs the plugin (skill + hook). Restart Claude Code to load the hook.

Codex / opencode / Cursor / any skills-aware agent — skill only

The skill is in the portable skills format, so any agent that reads SKILL.md can use the guidance:

npx skills add bcmyguest/assisted-by

The portable skill carries the guidance; the enforcing hook is Claude Code-specific for now. Native hook enforcement for opencode / Codex is on the roadmap.

How the hook decides

The guard inspects the git commit command line on the Bash PreToolUse event. It only acts when a message is being authored there — -m / --message, -F / --file, -C / --reuse-message, or --amend without --no-edit. It blocks (exit 2, reason on stderr, fed back to the agent) when the command:

• lacks an Assisted-by: trailer, or
• adds Co-Authored-By: Claude, or
• adds a Signed-off-by: naming Claude/Anthropic.

A bare git commit whose message is composed in the editor isn't visible to the hook — add the trailer yourself in that case. The guard never blocks on a parse failure or on commits that don't author a message (--amend --no-edit, plain git commit).

Why

The human submitter takes full responsibility for a contribution and its licensing, so the DCO chain (Signed-off-by:) must stay human. Assisted-by: records the assistance transparently without implying the AI can certify provenance.

Layout

.claude-plugin/
  plugin.json        # plugin manifest (name, version, keywords, skills)
  marketplace.json   # single-plugin marketplace (source ".") for `claude plugin marketplace add`
hooks/hooks.json     # registers the PreToolUse Bash guard
scripts/attribution-guard.py
skills/git-attribution/SKILL.md
skills.sh.json       # grouping for `npx skills`

Roadmap

• Native hook enforcement for opencode and Codex (the skill already works there; only the blocking hook is Claude Code-specific today).

License

See LICENSE. All rights reserved; in particular, the contents may not be used as training, fine-tuning, or evaluation data for machine-learning or AI systems.