Skip to content
Cawdit

Frequently asked questions

Cawdit did not review a pull request

Cawdit skips a pull request in these cases, and logs the reason:

  • It is a draft, and review.skip_drafts is on, which is the default.
  • Its author is in review.skip_authors.
  • Cawdit reviewed its head commit already.
  • Every changed file matches paths.ignore.
  • With cawdit serve, the owner is not in allowed_owners, or a check in review.wait_for failed or did not finish within an hour.

For cawdit serve, the webhook’s delivery history in Forgejo shows the answer to each delivery, and a 200 answer carries the reason it was ignored. For the Crow plugin, the step’s log shows it.

Every review fails with 403 Forbidden

The token lacks a scope. Cawdit needs write:repository, write:issue and read:user, as Set up the forge explains. Cawdit checks the scopes at start and names the missing ones, unless the forge could not be reached then. Forgejo cannot add scopes to a token, so create a new one.

Approving pull requests

No. Its review verdict is “comment”, or “request changes” when review.request_changes is on and a new finding has severity error. The commit status cawdit is the signal to require in branch protection.

The cost of a review

It depends on the model and the size of the pull request. The footer of every review shows its tokens and its cost, and cawdit serve adds both up in the metrics cawdit_llm_tokens_total and cawdit_llm_cost_usd_total. A smaller context.budget and a cheaper verify model lower the cost. A re-review after a push looks only at what changed, so it costs less than the first review.

Running the model yourself

Yes. Configure a provider of kind openai-compatible that points to Ollama, vLLM or another server that speaks the OpenAI API, as Configure a model provider shows. The server must support JSON schema output.

Trying Cawdit without posting anything

Yes. Set dry_run: true in the Crow plugin, or run cawdit review --dry-run, and Cawdit prints the review instead of posting it. A dry run still calls the model.

Reviewing again

Push a new commit, and Cawdit reviews what changed. With cawdit serve, comment @cawdit-bot review on the pull request, with the bot’s login, to review the whole pull request again, also a draft. In the Crow pipeline, restarting the pipeline runs the step again, and Cawdit skips a commit it reviewed already.

Disagreeing with a finding

With cawdit serve, reply @cawdit-bot ignore in the finding’s thread, and Cawdit marks the finding dismissed and never posts it on that pull request again. Any other reply that mentions the bot is a question, and Cawdit answers it in the thread from the finding and the code. Only accounts with write access to the repository can steer Cawdit this way. In an organisation’s repository, ignore may need admin access for the bot account, as setting up the forge explains.

Changes to .cawdit.toml in a pull request

No. Cawdit reads .cawdit.toml from the base branch, so a change to it applies to the pull requests after it is merged. The security model explains why.

GitHub and GitLab

No. Cawdit is built for Forgejo and Arbinel. It also accepts the webhooks of Gitea, whose API is close to Forgejo’s, but Gitea is not tested.

Where Cawdit keeps its data

cawdit review keeps none of its own. It reads its earlier findings back from the hidden markers in its comments. cawdit serve also keeps a history of its review runs and its job queue in Postgres.

Finding out why a finding was dropped

Run Cawdit with --log-level debug, or set log.level = "debug", to log each stage and how many findings the passes dropped. cawdit eval replays pull requests through the same engine without a forge, so you can compare prompts and models on the same cases.