LLM

The LLM module provides AI-powered TODO validation using Claude (Anthropic API) or any OpenAI-compatible endpoint.

LlmProvider

#![allow(unused)]
fn main() {
pub enum LlmProvider {
    Claude(ClaudeProvider),
    OpenAi(OpenAiProvider),
}
}

Dispatches LLM calls to the configured provider. Follows towl's existing enum dispatch pattern (FormatterImpl, WriterImpl).

call_raw

#![allow(unused)]
fn main() {
pub async fn call_raw(
    &self,
    user_content: &str,
    system_prompt: &str,
    api_key: &SecretString,
) -> Result<(String, LlmUsage), TowlLlmError>
}

Sends a prompt to the LLM and returns the response text and token usage.

build_provider

#![allow(unused)]
fn main() {
pub fn build_provider(config: &LlmConfig) -> Result<LlmProvider, TowlLlmError>
}

Factory function that creates the appropriate provider from configuration.

analyse_todos

#![allow(unused)]
fn main() {
pub async fn analyse_todos(
    todos: &mut [TodoComment],
    config: &LlmConfig,
) -> Result<AnalysisSummary, TowlLlmError>
}

Main entry point for TODO analysis. For each TODO (up to max_analyse_count):

1. Reads expanded context (~30 lines around the TODO + full function body)
2. Constructs a prompt with the TODO description, file path, and code context
3. Calls the LLM to determine validity
4. Parses the structured JSON response into an AnalysisResult
5. Attaches the result to TodoComment.analysis

Concurrency is bounded by max_concurrent_analyses via a tokio Semaphore.

Errors:

• NotConfigured -- TOWL_LLM_API_KEY not set
• UnsupportedProvider -- Provider is not "claude" or "openai"
• ApiError, AuthError, RateLimited -- From the LLM API

Validity

#![allow(unused)]
fn main() {
pub enum Validity {
    Valid,
    Invalid,
    Uncertain,
}
}

Whether a TODO is still valid:

AnalysisResult

#![allow(unused)]
fn main() {
pub struct AnalysisResult {
    pub validity: Validity,
    pub reasoning: String,
    pub is_resolved: bool,
    pub is_relevant: bool,
    pub is_actionable: bool,
    pub confidence: f64,
    pub enrichment: String,
}
}

AnalysisSummary

#![allow(unused)]
fn main() {
pub struct AnalysisSummary {
    pub valid_count: usize,
    pub invalid_count: usize,
    pub uncertain_count: usize,
    pub error_count: usize,
}
}

Summary counts returned by analyse_todos().

Providers

ClaudeProvider

POST to https://api.anthropic.com/v1/messages with headers:

• x-api-key: API key
• anthropic-version: 2023-06-01

System prompt is a top-level system field (not in the messages array).

OpenAiProvider

POST to {base_url}/chat/completions with Authorization: Bearer {key}. Default base URL: https://api.openai.com/v1. Configurable for Ollama, vLLM, etc.

System prompt is the first message in the messages array with role: "system".

ClaudeCodeProvider

Invokes the claude CLI as a subprocess with -p --output-format json. The combined system prompt and user content are passed as the final argument. No API key required.

Default command: claude. Configurable via llm.command and llm.args.

Auto-falls back to ClaudeProvider (API) if the CLI binary is not found on PATH.

CodexProvider

Invokes the codex CLI as a subprocess with -q. The combined prompt is passed as the final argument. No API key required.

Default command: codex. Configurable via llm.command and llm.args.

Auto-falls back to OpenAiProvider (API) with gpt-4o if the CLI binary is not found on PATH.

is_cli_provider

#![allow(unused)]
fn main() {
pub const fn is_cli_provider(&self) -> bool
}

Returns true for ClaudeCode and Codex variants. Used to skip the API key requirement for CLI-based providers.

Configuration

See Configuration for the [llm] config section.

Errors

#![allow(unused)]
fn main() {
pub enum TowlLlmError {
    ApiError { message, status },
    AuthError,
    RateLimited { retry_after_secs },
    ParseError { message },
    NotConfigured,
    UnsupportedProvider { provider },
    AnalysisLimitExceeded { count, max },
    HttpError(String),
}
}