burn-workflow

burn-workflow is the V6 executable-program counterpart to burn-graph. An author writes the whole workflow as code, the compiler validates its structured graph, and the runtime—not a main AI—owns serial, parallel, and bounded-loop scheduling after launch.

The current repository contains P01, an executable architecture proof. It demonstrates:

• code-defined sequence, parallel, and bounded loop control flow;
• AI nodes with explicit session: { lane, mode: "start" | "continue" };
• public-safe session handles that cannot expose provider transcript paths;
• compile-time rejection of parallel access to the same session lane;
• immutable compiled programs and immutable JSON output handoff;
• registered hard-coded Action nodes;
• a read-only, output-capped Git status node invoked with direct argv;
• structured events and JSON results.

It intentionally uses a mock AI adapter. A real provider adapter and durable run recovery belong to later locked vertical slices.

Run the proof

bun install
bun run verify

Inspect the normalized graph:

bun run compile:example

Execute it:

bun run run:example

The example starts a builder session, runs an independent reviewer and a hard-coded policy Action in parallel, continues the builder through a repair loop until a hard-coded verifier passes, and then continues the same builder session for synthesis.

Repository boundary

This code is a new V6 implementation. The historical ChaosRealmsAI/open-dynamic-workflow repository at commit e22ee1ba075e4658bcf4df082413400ab07d7df7 informed the comparison, but its code and documents were not copied. That repository resumes workflow calls but documents its real Codex nodes as single-shot; named cross-node session lanes are the new product contract here.