Review a pull request in CI

Run eVi headless in a pipeline and fail the build on real findings — either with the built-in reviewer or your own agent.

Uses: CLI + Python SDK · You'll need: eVi installed in the CI job and a model backend it can reach (a local server on the runner, or an OpenAI-compatible endpoint via evi backend add).

Option A — the built-in reviewer

evi review diffs your changes, has a model review them, and can turn the verdict into an exit code. That is all you need for a gate:

# Fail the job if the review verdict is FAIL. --multi runs several lenses.
evi review --multi --exit-code

Useful flags:

• --staged — review git diff --cached (pre-commit).
• --branch main — review this branch against main.
• --json — emit {verdict, exit_code, review} for a later step to parse.
• --diff-file patch.diff — review a saved patch instead of the working tree.

A drop-in GitHub Actions workflow is in examples/github/pr-review.yml.

Option B — your own agent, headless

When you want custom logic — post the summary as a PR comment, gate on a specific check — drive an agent yourself. run_headless returns a structured result and to_json serialises it; set the exit code from result.error:

from evi.sdk import build_agent, run_headless, to_json

# Scope the agent to read-only fs + code tools for an unattended run.
agent = build_agent(tool_categories=["fs", "code"])
result = run_headless(agent, "Does this diff introduce any obvious bug? Answer with findings.", max_turns=8)

print(to_json(result))          # {"text", "tools", "usage", "error"}
raise SystemExit(1 if result.error else 0)

Runnable version: examples/python/headless_ci.py

python examples/python/headless_ci.py "Does this repo have a README? Answer yes/no."

Keeping CI runs safe and cheap

• Scope the tools. tool_categories=["fs", "code"] gives the agent read-only file and search tools — no shell, no writes — which is what an unattended run should have.
• Bound the work. max_turns= caps how many tool round-trips a single run may take, so a confused model can't loop forever.
• Pick a small model on the runner. See Run a model on your own hardware — a 7B coder model is plenty for diff review and keeps the job fast.

Where to go next

• Guard an agent's inputs and outputs — add input/output safety on top of a headless run.