| MINIMAX_API_KEY | MiniMax API key | yes | — |
| MINIMAX_API_URL | MiniMax chat-completions endpoint. Override for a different region (e.g. https://api.minimax.chat/... for the China endpoint). | no | https://api.minimaxi.chat/v1/chat/completions |
| MINIMAX_MODEL | MiniMax AI model | no | MiniMax-M3 |
| MINIMAX_SYSTEM_PROMPT | Custom system prompt for the AI reviewer | no | You are an expert code reviewer reviewing a GitHub pull request diff. Output ONLY the final review in GitHub Markdown — do not narrate your reasoning, think out loud, or describe your process. Start directly with a one-paragraph "## Summary", then list concrete findings under "## Critical", "## Warning", and "## Nitpick" headings (omit any heading that has no findings). For each finding, reference the file and line and keep it to a sentence or two. If the diff looks good, say so in one line under Summary and add no other sections. |
| MINIMAX_REVIEWER_NAME | Name shown in the review comment header | no | BoolMax3 Code Review |
| MINIMAX_TRIGGER_PHRASE | Phrase that, when included in a PR comment by a repo owner/member/collaborator, triggers a fresh review posted as a NEW comment. Requires the workflow to also listen for issue_comment events. | no | @BoolMax3 |
| COMMENT_MODE | Where the review is placed on new commits: 'smart' (edit in place while it's the last comment, else post a new one), 'update' (always edit in place), or 'new' (always a new comment). | no | smart |
| MINIMAX_MAX_TOKENS | Maximum tokens for the model RESPONSE (output). Default 131072 is M3's recommended max output; 900K input + this still fits M3's 1M context. Set to 0 to let the API decide. | no | 131072 |
| MINIMAX_TEMPERATURE | Sampling temperature (empty = model default). | no | "" |
| EXCLUDE_PATTERNS | Comma-separated file patterns to exclude from review (e.g. '*.lock,dist/**,*.min.js') | no | *.lock,package-lock.json,yarn.lock,pnpm-lock.yaml,dist/**,*.min.js,*.map |
| MAX_INPUT_TOKENS | Approx. max INPUT tokens of diff to send, so large PRs stay within the model context and don't fail. Default 900000 fits M3's 1M window. Note: inputs over 512K tokens bill at MiniMax's higher long-context rate. | no | 900000 |
| MAX_DIFF_CHARS | Optional hard character cap on the diff. 0 = derive from MAX_INPUT_TOKENS. Set a value to override the token-based budget. | no | 0 |
| MAX_RETRIES | Number of retries on rate-limit (429), server (5xx), or network errors | no | 2 |
| GITHUB_TOKEN | GitHub token for API access | no | ${{ github.token }} |