Skip to main content

0.2.0-wave-1 — a task becomes a plan

· 3 min read
0.2.0-wave-1 — a task becomes a plan | AEPThe first release you can run against something. A task plus a document tree produces a plan, evidence decides what may be done and whether the work is finished, and a capability decision names the rule that decided it.AEPfield-noteaepfield-noteadopterdevelopertry-spec-driven-developmentunderstand-safe-agentic-codingresearch

Source-owned field note · website/blog/2026-08-20-0028-the-execution-core.md · revision 8f106fad0926

0.1.0 was a vocabulary. This is the first release where a question goes in and an answer comes out: a task plus a document tree produces a plan, and evidence decides both what may be done and whether the work is finished.

What landed

  • Resolution — a task resolves against the document tree into the principles, capabilities and requirements actually in force for it, rather than a global rulebook you read yourself.
  • Live execution with derived facts, so the state of a run is computed from what has been recorded rather than tracked in parallel.
  • Capability decisions that name the rule that decided. A refusal that says "denied" is a refusal you argue with; one that names the document, the rule and the line is one you act on.
  • Completion as a checklist, not a declaration — the beginning of the idea that finishing is a predicate over recorded facts.
  • 49 protocol documents, the protocol CLI, and a worked example that walks a task to completion.

163 tests, roughly 62% of the v0.2 scope.

The design decision inside it

The engine evaluates; it never manufactures. Nothing in this release can produce the evidence that would satisfy a rule it is evaluating, and that separation is why a later driver could be given a shell without also being given the ability to approve its own work.


Written retrospectively from this release's CHANGELOG.md section and its annotated tag; no re-run command output.