ArgusReview

⚑ Config Builder

Pick your options β€” get a ready .ai-review.yaml plus the exact CI job and secrets to add. Secrets are never written into the YAML (the tool doesn't expand ${...} in files); they go into your CI as env vars.

LLM
VCS
Pipeline info & tokens are supplied by your CI (see the CI setup panel).
Review
Which argus-review command the CI job/action executes.
β€’ run β€” full pipeline: inline comments (one LLM call per file) + one summary comment. Highest quality, highest quota.
β€’ run-inline / run-summary / run-context β€” one piece of the pipeline on its own (context = one call across all files, inline comments).
β€’ run-agent β€” low-quota ReAct flow: one agent session explores the diff/conventions via read-only shell tools, then returns both the summary and inline comments together. Cheapest way to get both.
β€’ run-agent-inline / run-agent-summary β€” same low-quota agent flow, but only inline or only the summary.
β€’ run-inline-reply / run-summary-reply β€” AI replies to existing review threads (usually a separate manual trigger).
β€’ clear-inline / clear-summary β€” remove previously posted AI comments.
The run-agent* commands always drive the agent loop regardless of the "Agent mode" toggle below β€” that toggle only affects run/run-inline/run-summary/run-context.
Free-form: English, Vietnamese, ζ—₯本θͺž, ...
Which slice of each changed file the model sees.
β€’ ADDED_AND_REMOVED_WITH_CONTEXT β€” recommended default: changed lines + surrounding context. Good quality at moderate quota.
β€’ FULL_FILE_DIFF β€” full unified diff of changed files (more context, more tokens).
β€’ FULL_FILE_CURRENT / FULL_FILE_PREVIOUS β€” the whole file after / before changes. Needed for whole-file rules (e.g. "max 300 lines per file"); highest token cost.
β€’ ONLY_ADDED / ONLY_REMOVED / ADDED_AND_REMOVED β€” leanest, changed lines only.
Tip: for whole-file conventions, use FULL_FILE_CURRENT, or use run-agent so the model can read files on demand without loading every file in full.
Coding conventions (optional)
Agent mode (optional)
Turns those commands into a ReAct loop: instead of one direct LLM call per file, the model can run read-only shell commands (ls, cat, rg, git diff, sed -n, ...) to explore the repo before answering. Usually more expensive than the plain commands, since it may still receive the full diff/conventions up front and then re-sends tool history each step.
β‘  .ai-review.yaml (commit to repo root) 
loading schema…
β‘‘ CI setup β€” … 

 
 β‘’ Secrets to create in your CI