llm

Uses LLM inference to generate or modify a file in the target directory.

Usage

- name: generate-deployment
  llm:
    provider: anthropic
    model: claude-sonnet-4-20250514
    systemPrompt: "Output only valid YAML. No explanation."
    prompt: |
      Generate a Kubernetes Deployment manifest for {{ .serviceName }}
      in namespace {{ .namespace }} using image {{ .image }}.
    target: "deploy/{{ .serviceName }}.yaml"
    mode: generate
    maxTokens: 2048
    retries: 2
    retryDelay: "3s"
    providerConfig:
      tokenEnv: ANTHROPIC_API_KEY

Field	Required	Description
provider	yes	LLM provider. See Providers.
model	yes	Model identifier (provider-specific).
prompt	yes	User prompt. Templated. In modify mode, existing file content is prepended automatically.
systemPrompt	no	System prompt. Omitted if empty. Templated.
target	yes	Output file path relative to the target directory. Templated.
mode	no	"generate" (default) or "modify". See Modes.
maxTokens	no	Maximum output tokens. Provider default used when omitted.
retries	no	Maximum retry attempts on failure. Default: 0 (no retry).
retryDelay	no	Initial delay between retries (e.g. "2s", "500ms"). Doubles each attempt. Default: "2s".
providerConfig.tokenEnv	no	Env var name holding the API key. Overrides provider default.
providerConfig.project	no	GCP project ID. Required for vertex. Templated.
providerConfig.location	no	GCP region. Required for vertex. Default: "us-central1". Templated.

All string fields are templatable.

Providers

Provider	Value	Auth	Default env var
OpenAI	openai	API key	OPENAI_API_KEY
Anthropic	anthropic	API key	ANTHROPIC_API_KEY
Google Gemini	gemini	API key	GEMINI_API_KEY
Google Vertex AI	vertex	Application Default Credentials	—
OpenRouter	openrouter	API key	OPENROUTER_API_KEY
AWS Bedrock	bedrock	AWS credential chain	—

Set providerConfig.tokenEnv to override the default env var for key-based providers. vertex and bedrock do not use API keys.

Modes

generate (default)

Sends the rendered prompt to the model and writes the response to target. Fails if the file already exists. Use this to create new files.

modify

Reads the existing file at target, prepends it to the prompt, then calls the model and overwrites target with the response. Fails if the file does not exist.

The prompt sent to the model is composed as:

Here is the existing file content:

```
<existing content>
```

<your prompt>

Retry

When retries > 0, failed attempts are retried with exponential backoff. The delay doubles on each attempt:

Attempt	Delay
1st retry	retryDelay
2nd retry	retryDelay × 2
3rd retry	retryDelay × 4

Retries trigger on inference errors and on empty responses. If the context is cancelled during a wait, the operation fails immediately.

Dry Run

In dry-run mode, the model is not called and no file is written. A log entry records the provider, model, target, and prompt length.

Examples

Generate a new file

- name: create-service-config
  llm:
    provider: openai
    model: gpt-4o
    prompt: "Generate a Kubernetes Service for {{ .serviceName }} on port 8080."
    target: "k8s/{{ .serviceName }}-svc.yaml"
    providerConfig:
      tokenEnv: OPENAI_API_KEY

Modify an existing file

- name: update-readme
  llm:
    provider: anthropic
    model: claude-sonnet-4-20250514
    prompt: "Add a ## {{ .serviceName }} section describing this service."
    target: README.md
    mode: modify
    providerConfig:
      tokenEnv: ANTHROPIC_API_KEY

Vertex AI (GCP)

- name: generate-policy
  llm:
    provider: vertex
    model: gemini-2.5-flash
    prompt: "Generate a network policy for {{ .serviceName }}."
    target: "policies/{{ .serviceName }}.yaml"
    providerConfig:
      project: "{{ .gcpProject }}"
      location: us-central1