v0 Roadmap

v0 Roadmap

v0 should prove that Migaki's representation and evidence model are useful before it tries to become broad.

Goal

Demonstrate that a multi-step agentic workflow can be represented as mIR, optimized by a small number of passes, lowered into concrete execution, and reported with evidence.

Primary Demo

A TypeScript RAG workflow:

1. Receive a user question.
2. Retrieve candidate document chunks.
3. Deduplicate overlapping chunks.
4. Preserve stable instructions as a cacheable prefix.
5. Estimate token and cost deltas.
6. Route chunk ranking to a cheaper model or mock backend.
7. Use a stronger model or mock backend for final synthesis.
8. Validate answer against cited chunks.
9. Retry only failed synthesis.
10. Export an evidence bundle and CLI report.

Packages

• packages/mir: schemas, validators, examples.
• packages/runtime: planner, pass runner, evidence, tracing.
• packages/providers: OpenAI-style, Anthropic-style, LiteLLM-compatible, and mock backends.
• packages/adapters: fetch wrapper first; framework adapters later.
• packages/cli: report and replay commands.

Milestones

M0: Repository Skeleton

• Package manager and TypeScript config.
• Workspace layout.
• Formatting and test tooling.
• Minimal README update.
• Example fixtures for baseline and optimized plans.

M1: mIR Schema v0

• Plan, node, edge, context block, constraint, capability, and evidence types.
• JSON Schema or Zod validators.
• Example plans for RAG and code review workflows.

M2: Pass Runtime

• Pass interface.
• Pass runner.
• Plan diff model.
• Warnings and evidence emission.
• Deterministic test fixtures.

M3: Context and Cache Passes

• Exact duplicate context elimination.
• Dead context flagging.
• Stable prefix detection.
• Prompt-cache layout report.
• Token and cost estimation.

M4: Execution and Backends

• Mock backend for repeatable tests.
• Fetch-compatible provider wrapper.
• OpenAI-style adapter.
• Anthropic-style adapter.
• Optional LiteLLM-compatible backend.

M5: Evidence and CLI

• Evidence bundle serialization.
• Trace artifact.
• CLI report for plan diff, token estimate, cost estimate, routing decisions, validator results, and warnings.
• Replay command for deterministic or mock-backed runs.

M6: Demo Benchmark

• Baseline workflow.
• Migaki-optimized workflow.
• Reported token, cost, latency, validator, and trace deltas.
• Clear statement of what changed and what did not.

Explicit Non-Goals for v0

• Learned routing.
• General semantic compression.
• Semantic caching.
• Complex durable workflow semantics.
• Multi-agent orchestration.
• Universal provider abstraction.
• Claims about identical output quality.