StatePlane¶
StatePlane is a typed context compiler plus provider request preparation layer.
Current supported product surfaces:
stateplanePython SDK for embedding in customer codestateplaneinstallable CLI for local compile, render, debug, and demo workflows- lower-level packages such as
context_coreandprovider_openai
Start Here If You Are New¶
If you are new to the project, start with Getting Started.
That page shows one complete flow:
- create
StatePlaneSession(...) - call
transform(...)withmessages = [...] - inspect selected and excluded context
- inspect the final provider payload
- optionally execute the provider call
What It Does Today¶
StatePlane currently helps you decide what context should go into one provider request, explain why that context was chosen, and render the exact payload that would be sent.
What goes in:
- a task summary
- a transcript or bucketed context inputs
- provider and model identifiers
- a token budget and response reserve
What changes:
- inputs are normalized into typed context items
- required items are preserved
- optional items are selected or excluded deterministically
- a typed report explains the decision
- the selected items become one provider-specific payload
What comes out:
prepared.requestprepared.compiled_contextprepared.compilation_reportprepared.rendered_request
Recommended learning paths¶
Choose one path based on your goal:
- First-day onboarding with commands + checkpoints → First-Day Tutorial
- Quick embed into an app → Getting Started
- Deep parameter understanding → OpenAI Parameter Tutorial
- Stateful coding-agent workflows → Structured Coding-Agent State with OpenAI
- Replay and diff run history → Replay and Diff a Coding-Agent Run
- Evaluate continuation-context quality → Coding-Continuation Evaluation
- Evaluate task outcomes on external benchmarks → External Coding Benchmarks
- Measure which state groups help or waste tokens → State Value Lab
- Train and validate a learned selector for optional state → State Selector v0
- Prepare the selector-v1 dataset contract → Selector v1 Dataset
- Inspect selector-v1 target schema v2 → Selector v1 Targets
- Train the schema-v2 selector artifact → Selector v1
- Review the current selector-v1 proof-step results → Selector v1 Results
- Inspect the latest harness-evaluated SWE-bench snapshot → External Coding Benchmark Results
SDK Versus CLI¶
The SDK is the main product surface. The CLI wraps the same SDK operations for:
- local inspection
- reproducible demos
- CI checks
- support and debugging
Start here:
- Getting Started
- Structured Coding-Agent State with OpenAI
- Replay and Diff a Coding-Agent Run
- Codex Hooks Integration
- Coding-Continuation Evaluation
- External Coding Benchmarks
- State Value Lab
- State Selector v0
- Selector v1 Dataset
- Selector v1 Targets
- Selector v1
- Selector v1 Results
- External Coding Benchmark Results
- OpenAI Context Demo
- StatePlane SDK
- Provider OpenAI