Tell me that again but slowly...
Semantic diff review for coding agents.
Install the bundled agent skill:
npx skills add mikker/semantic-reviewAn LLM reads a git diff and reorganizes it into a narrative HTML report — grouped by concern, not alphabetically, with the interesting excerpts inline, highlighted and commentable. When the reviewer clicks Done, their comments print to stdout as plaintext for the agent that invoked it.
agent runs `semantic-review` ──► LLM analyzes the diff ──► browser opens the report
agent reads stdout ◄── plaintext feedback ◄── human comments, clicks Done
Requires Node >= 20 (or Bun) and one backend: ANTHROPIC_API_KEY, OPENAI_API_KEY, or an installed claude, codex, gemini, or pi CLI.
npx semantic-reviewRun it like $ git diff:
semantic-review # review uncommitted changes
semantic-review main...HEAD # review a branch
git diff -U10 | semantic-review # review any piped diff
semantic-review --with anthropic,codex # one analysis per backend, tabbed
semantic-review --with openai # use the OpenAI API
semantic-review --model claude-sonnet-5 --effort medium
semantic-review --export review.html # standalone review; Done copies feedback
# Generate a prompt, then render caller-provided analysis
semantic-review --emit-prompt > prompt.txt
semantic-review --analysis analysis.json
Put user defaults in ~/.config/semantic-review/config.json (or
$XDG_CONFIG_HOME/semantic-review/config.json). Command-line options override
them. For example, to always use the Codex CLI by default:
{
"backend": "codex"
}backend may also be an array for a multi-backend review. The other supported
settings are model and effort.
--export <file> writes the walkthrough as a self-contained HTML document
with no remote assets or server dependency. Reviewers can add line, selection,
and overall comments as usual. Done copies the same plaintext feedback the
interactive server would return, ready to paste back to an agent. The absolute
path of the exported file is printed to stdout.
Try it with your agent:
Run
npx semantic-reviewand wait for it to exit. Itsstdoutis the user's review feedback — treat each comment as a change request and address it.
The skill analyzes the diff in an isolated context by default, keeping implementation history out of the review:
/semantic-review
Describe a preferred harness or model naturally to launch an independent sidecar instead:
/semantic-review use Claude Code with fable
/semantic-review run this through Codex using gpt-5.3-codex-spark
/semantic-review use google/gemini-2.5-pro through Pi
The skill is deliberately user-invoked only. Claude Code runs it in a foreground fork with no conversation history; other agents prefer an isolated same-harness sidecar and fall back to host analysis only when isolation is unavailable.
