karandash ai
Ad-hoc AI helpers that call a language model directly: suggest a fix for a finding, draft a reply to a review comment, or draft an MR description. They do not fetch an MR or run validation; for AI review during validation see AI Rules and the AI Features Reference.
The command group is contributed by the built-in AI plugin (AiCliCommandModule). When a command runs it loads the
discovered config and reads its ai: block:
| Key | Effect |
|---|---|
ai.provider |
openai, anthropic, gemini, ollama, or custom (needs ai.baseUrl); other providers are not supported here |
ai.baseUrl |
The provider's endpoint (an ollama URL must be a loopback or private-network host) |
ai.model, ai.temperature |
Model and temperature, unless --model / --temperature is given |
Without ai.provider, the first of these variables that is set picks the provider:
| Variable | Provider |
|---|---|
OPENAI_API_KEY |
OpenAI |
ANTHROPIC_API_KEY |
Anthropic |
GEMINI_API_KEY |
Google Gemini |
OLLAMA_HOST |
Ollama at that URL |
The API key always comes from the provider's variable (CUSTOM_AI_API_KEY for custom). With neither ai.provider
nor a variable set, the command fails to initialize. Without a model, each command uses the provider's cheap default
model at temperature 0.3; the token limit (1000, 1500 for describe) is fixed.
Synopsis
karandash ai fix --rule <rule-id> --message <finding-message>
karandash ai reply --comment <review-comment> [--context <additional-context>]
karandash ai describe --title <mr-title> [--diff <diff-summary>]
Each flag takes the next word as its value; quote values that contain spaces. Unknown words are ignored.
Subcommands
fix
Suggests concrete steps to fix one finding. Both --rule and --message are required.
karandash ai fix --rule title-format --message "Title must start with a conventional commit type"
reply
Drafts a reply to a code review comment. --comment is required; --context adds background for the model.
karandash ai reply --comment "Why not reuse the existing cache?" --context "The cache is per-run and this value is global"
describe
The mr-summary analyzer gives the same title and description suggestion in-pipeline during validate.
Drafts an MR description (title, What/Why/How to test/Risks sections) as JSON. At least one of --title and
--diff is required.
karandash ai describe --title "feat: add retry to webhook reporter" --diff "WebhookReporter: wrap send in HttpRequestRetry"
Output and exit codes
The model's answer goes to stdout as plain text. Usage and errors:
| Exit code | When |
|---|---|
0 |
The model answered |
1 |
No usable provider (No AI provider found. Set ..., on stderr) |
2 |
A required argument or flag is missing (the usage line is printed), or --temperature / ai.temperature is not a number |
3 |
The provider call failed (AI provider error: ..., on stderr) |
6 |
The ai plugin module failed to initialize |
See also
- explain: explains your configuration against module schemas, without a model (
--aiadds AI examples) - AI Features Reference