Struct Magi 

pub struct Magi { /* private fields */ }

Main entry point for the MAGI multi-perspective analysis system.

Composes agents, validation, consensus, and reporting into a single orchestration flow. The analyze method launches three agents in parallel, parses and validates their responses, computes consensus, and generates a formatted report.

Examples

// let magi = Magi::new(provider);
// let report = magi.analyze(&Mode::CodeReview, content).await?;

Implementations

impl Magi

pub fn new(provider: Arc<dyn LlmProvider>) -> Self

Creates a MAGI orchestrator with a single provider and all defaults.

Equivalent to MagiBuilder::new(provider).build().unwrap(). This cannot fail because all defaults are valid.

Parameters

• provider: The LLM provider shared by all three agents.

pub fn builder(provider: Arc<dyn LlmProvider>) -> MagiBuilder

Returns a builder for configuring a MAGI orchestrator.

Parameters

• provider: The default LLM provider.

pub async fn analyze( &self, mode: &Mode, content: &str, ) -> Result<MagiReport, MagiError>

Runs a full multi-perspective analysis.

Launches three agents in parallel, parses their JSON responses, validates outputs, computes consensus, and generates a formatted report.

Parameters

• mode: The analysis mode (CodeReview, Design, Analysis).
• content: The content to analyze.

Errors

• MagiError::InputTooLarge if content.len() exceeds max_input_len.
• MagiError::InsufficientAgents if fewer than 2 agents succeed.
• MagiError::InvalidInput if nonce collision detected (probability ~2^-64 per call; fastrand effective state ~64 bits; see ADR 001 §Decision: Nonce RNG choice).

Concurrency

The internal rng_source is guarded by a std::sync::Mutex, so concurrent calls to analyze() from multiple tasks serialize on nonce generation. In practice nonce generation is a single u128 read (~nanoseconds), so contention is negligible under typical workloads. If profiling shows this becomes a bottleneck in a multi-tenant deployment, consider wrapping Magi in a pool of instances (one per tenant), or await v0.4 which may expose with_rng_source publicly to allow a thread-local RNG strategy.