-
Notifications
You must be signed in to change notification settings - Fork 0
Risks and Open Questions
This page tracks places where Migaki needs discipline, evidence, or design decisions.
Migaki sits near agent frameworks, gateways, workflow engines, evals, observability, and provider SDKs. It can accidentally grow into all of them.
Open question:
- What is the smallest mIR contract that supports useful execution optimization without modeling the whole application?
Optimization can change behavior. Token reduction alone is not success.
Open questions:
- Which validators are required for each example?
- How should allowed regression be expressed?
- How should model-based judges be calibrated or bounded?
Semantic compression and summarization are high leverage but risky.
Open questions:
- What acceptance tests make lossy context transforms acceptable?
- Which context roles must never be compressed?
- How should provenance survive summarization?
Provider capabilities change quickly.
Open questions:
- How should capability registries be versioned?
- How often should provider adapters be tested?
- Should provider capability declarations include source links and retrieval dates?
Evidence bundles can contain sensitive prompts, documents, tool outputs, and user data.
Open questions:
- What is the default redaction policy?
- When is full trace replay allowed?
- How should evidence bundles encode omitted data?
Tool calls can mutate external systems.
Open questions:
- How does mIR represent idempotency?
- Which tool calls require human approval?
- How does partial replay avoid repeating side effects?
Gateways already perform routing, fallback, budgets, retries, caching, and monitoring.
Open questions:
- Which decisions belong in Migaki versus a gateway?
- How should Migaki express optimization plans to a gateway without duplicating gateway policy?
Workflow engines already preserve state and replay history.
Open questions:
- What does Migaki replay mean when Temporal, LangGraph, or another runtime owns durable replay?
- How should evidence handles attach to workflow history?
The cleanest integration requires frameworks to compile into mIR, but early adopters may need wrappers.
Open questions:
- Is the first useful integration a fetch wrapper, framework adapter, or CLI optimizer?
- What can Migaki prove from traces when it cannot see the full workflow?