Dry-run review

The aicr review command runs a single review without the long-running server. With --dry-run it prepares and runs the full review pipeline but skips every output channel, so you can validate the LLM, agent, and sandbox without spamming a PR or IM channel. This page expands on the dry-run section of Quick start.

For the complete CLI command reference, see CLI commands.

When to use dry-run

• Validating that LLM credentials, the agent CLI, and the sandbox are wired up before opening the server to webhooks.
• Iterating on prompts, skills, or model choice without publishing.
• Reproducing a review locally from a checkout.

Dry-run does not publish to any channel, regardless of outputs.routes. A non-dry-run review without a matching output route is recorded as skipped with skipReason="no_output_publisher" — that is routing, not a dry run.

Running a dry-run

export AICR_LLM_API_KEY=sk-xxx

node packages/cli/dist/index.js review \
  --config example/config.yaml \
  --repo "my-org/my-repo" \
  --provider gitea \
  --source-root . \
  --dry-run

Flags

Flag	Description
--config <path>	Path to the config YAML file
--repo <ref>	Repository reference (e.g. owner/repo)
--provider <name>	Trigger provider kind from the config schema (gitea, github, gitlab, p4, svn)
--trigger <name>	Trigger name
--reason <text>	Review reason
--source-root <path>	Source root directory to review
--base-prompt <path>	Path to a base system prompt template (overrides the workspace prompt file)
--changed-file <path>	Changed file (repeatable)
--base-sha <sha>	Base revision SHA
--head-sha <sha>	Head revision SHA
--url <url>	PR / MR / commit URL
--author-username <u>	Author username
--author-email <e>	Author email
--dry-run	Run without publishing to output channels
--max-prompt-tokens <n>	Maximum prompt token budget

--provider selects which trigger profile’s VCS adapter and filters apply. --source-root points at the local checkout; combine it with --changed-file (or --base-sha / --head-sha) to scope the diff.

Reading the result

A dry-run prints the resolved review summary, the list of reported problems (if any), and the skip reason when the review was skipped. Common skip reasons:

Skip reason	Meaning
lgtm	No actionable problems found
no_reviewable_code	Nothing reviewable in the change
no_output_publisher	No output route matched (non-dry-run only)
no_problems_suppressed	All selected summary channels suppress zero-problem results
output_dispatch_failed	Every dispatch attempt failed

If the dry-run reports problems, the LLM and pipeline are healthy. If it errors with AgentContextOverflowError, enable llm.model_catalog or set a context_window override — see Troubleshooting.

Next steps

Once dry-run is clean, wire up your first webhook to let real VCS events drive reviews.