Skip to content

0004. Approximate, and always report it

  • Status: Accepted
  • Date: 2026-10-14

Context

Real trees contain nodes Phase 1 cannot model exactly: reactive sequences and fallbacks (preemption), Retry and Repeat (loops), Parallel, and custom controls and decorators with unknown semantics. Refusing such trees would make the tool useless on real missions. Silently approximating them would make it dangerous: a reader would trust a graph that is subtly wrong.

Decision

Phase 1 walks these nodes with a simple, documented approximation (for example, a reactive sequence is walked like a plain sequence), and always emits a diagnostic with the node type, the approximation used and the source location.

  • The CLI prints diagnostics on stderr (--quiet hides them).
  • The JSON output carries them.
  • beehave stats counts them per node type: this is the plan's "breakage" metric, the list of what Phase 2 has to fix.

Consequences

  • Beehave runs on any tree from day one.
  • The graph is honest about where it is approximate.
  • Each Phase 2 rule removes a diagnostic. Progress is measurable.