Project Scope

Project Scope

Migaki focuses on one responsibility:

Given a logical agentic workflow and explicit constraints, produce the best measurable execution plan and evidence for why it is acceptable.

The project should stay disciplined around that responsibility.

Core Problem

AI execution infrastructure already exists, but it is fragmented. Routing lives in gateways. State lives in frameworks. Retries live in workflow engines. Context management lives in provider APIs. Verification lives in eval platforms. Tool interop lives in MCP-style protocols. Tracing lives in observability systems.

Migaki exists to represent these execution concerns together as an optimizable system-level artifact.

In Scope

• Provider-neutral logical execution plans.
• Capability-aware provider lowering.
• Model, tool, retrieval, context, validation, branch, retry, fallback, and approval nodes.
• Explicit constraints for cost, latency, quality, governance, retention, provider allow/deny lists, replay, and audit.
• Context block ledgers with provenance, mutability, cache policy, privacy class, and retention policy.
• Optimization passes that emit plan diffs and evidence.
• Evidence bundles that make optimization observable, benchmarkable, and auditable.
• SDKs, adapters, wrappers, and trace consumers for existing frameworks and runtimes.

Out of Scope

• Replacing agent frameworks.
• Replacing provider gateways.
• Replacing durable workflow engines.
• Replacing MCP or tool protocols.
• Replacing observability or eval platforms.
• Claiming identical answers after lossy optimization.
• Hiding provider differences behind a lowest-common-denominator abstraction.
• Global autonomous routing without task-specific measurement.

v0 Scope

The first implementation should validate the abstraction rather than chase completeness.

v0 should include:

• TypeScript runtime.
• mIR schema v0.
• Fetch-compatible provider wrapper.
• Provider adapters for OpenAI-style and Anthropic-style APIs.
• Optional LiteLLM-compatible backend.
• Context block ledger.
• Token and cost estimation.
• Exact duplicate context elimination.
• Stable prefix detection.
• Prompt-cache layout reporting.
• Simple retry and fallback policy.
• Basic static routing policy.
• Execution evidence bundle.
• Replayable trace artifact.
• CLI report.

v0 should avoid leading with:

• General semantic compression.
• General semantic caching.
• Autonomous global model routing.
• Cross-provider magical equivalence.
• Claims of identical answers.

Design Principle

Never perform lossy optimization without a declared acceptance test.