graff — a fast agentic coding harness in Zig: multi-provider, MCP, workflows, DGM evolution loop, TS/Python SDKs
Python
Kernel: transport
Source of truth: lean-proofs/Graff/Transport.lean.
Process kernel, not a Turing machine: finite Event / step, no tape.
The 96-cell cube is the snapshot. A sub never takes WS; only one cell is
WS (a live root Responses turn). PromptCache isolates the child key; this
kernel is the pipe that child is forbidden from opening. Shape stays a
cube of one observation — fleet topology is not a Shape cell.
The diagram is the projection of the live Python step. Emit it with
python3 spec/conformance.py --diagram transport.
stateDiagram-v2
[*] --> Sse
Blocked --> Sub: markSub
Blocked --> Ws: setCodexWs 1
Blocked --> Ws: setHasOut 1
Blocked --> Sse: setKind anthropic
Blocked --> Sse: setKind openai
Blocked --> Ws: setQuiet 0
Blocked --> Ws: setWsOff 0
Sse --> Sub: markSub
Sse --> Blocked: setKind responses
Sse --> Ws: setKind responses
Sub --> Blocked: joinRoot
Sub --> Sse: joinRoot
Sub --> Ws: joinRoot
Ws --> Sub: markSub
Ws --> Blocked: setCodexWs 0
Ws --> Blocked: setHasOut 0
Ws --> Sse: setKind anthropic
Ws --> Sse: setKind openai
Ws --> Blocked: setQuiet 1
Ws --> Blocked: setWsOff 1