Results and operations
Read the result
The connector publishes confirmed findings in MR discussions and a completion note. Review the full session history to inspect the agent's checks, tool output and hook errors. The agent's final response is part of that history, not a comment automatically copied to GitLab.
On later reviews, the connector can update its own discussions, resolve addressed findings and report recurring problems. It leaves human discussions and other reviewers unchanged.
If the MR diff changes, the MR closes or the bot is removed from reviewers during analysis, the connector cancels or discards obsolete work. When the diff changes, it also removes itself from reviewers without posting a note. Assign the bot again to review the new revision. It does not publish the old review against the changed diff. For another review after completion or failure, assign the bot again. An unsuccessful analysis is not retried automatically.
Failure notes show the public Orpheus error message when available, or a safe application failure description. Public error text can include the original harness message. To omit it, use a template without the Reason placeholder. The default note has no session link. To include one, use a custom failure template.
Discussion status distinguishes a finding explicitly resolved by a participant from one made outdated by a push. A recorded resolver is enough to recognize explicit resolution when GitLab omits its cause.
Run one instance
Run exactly one connector for each GitLab installation and bot identity. Stop the old process before starting its replacement. Increasing replicas does not increase supported review capacity. Use MAX_CONCURRENT_REVIEWS instead, within the capacity of Orpheus and your providers.
The connector exposes no HTTP API or health endpoint and needs no separate database. Monitor process status and JSON logs on stdout. Configuration, startup and runtime failures exit with code 1. LOG_LEVEL controls verbosity.
On shutdown, accepted Orpheus runs continue. After restart, the connector recovers them and resumes publication. Transient publication errors are retried without starting another analysis.
To update the Compose example, change the image version in .env, then:
docker compose pull
docker compose stop review
docker compose up -d review
docker compose logs --tail=100 reviewDiagnose a review
| Symptom | What to check |
|---|---|
| The connector exits at startup | JSON error log, required connection settings, readable workflow directory and file validity |
| An assigned MR does not start | Bot access, reviewer assignment, open MR state, poll interval and concurrency limit |
| Skip note in GitLab | Add the project to a workflow, restart and assign the bot again |
| Session creation fails | Existing profile, template and service codes, API key and both request-size limits: the connector limit and Orpheus MAX_REQUEST_BYTES |
before_run fails | HTTPS clone/fetch access, required sandbox tools and whether the MR changed during preparation |
| Agent completed, publication failed | Inspect after_run, connector logs and GitLab API permissions. A final answer alone is not a publishable review |
| No session link in a failure note | The default failure note omits the link. To include one, configure the public web URL and a custom notes.fail using the SessionURL placeholder. Before session creation the URL is empty |