mirror of
https://github.com/EveryInc/compound-engineering-plugin.git
synced 2026-10-07 23:49:12 +02:00
Previous trigger criterion ("include when behavior is hard to pin
down without a concrete scenario") was judgment-based and produced
real variance: same `--quiet` flag synthesis, two test runs, one
agent generated 4 AEs and another generated 0. Both were defensible
reads of the spec.
The asymmetric failure mode favors more-inclusion: under-inclusion
makes downstream planners invent missing context; over-inclusion is
just ceremony. Per AGENTS.md "match prescription to failure mode,"
tightening here is justified.
New rule: AEs are REQUIRED for behavioral-conditional requirements
(any "When X, Y" or "If X, Y" framing) regardless of tier — even
Lightweight. Conditional framing signals state-dependent behavior
where prose alone leaves implicit ambiguity (e.g., "When --quiet is
set, errors continue to surface" — does that cover warnings? binary
errors? AE pins it down).
Non-conditional requirements remain triggered (Standard/Deep) or
omit-unless-triggered (Lightweight) per existing rules. The section
is still not exhaustive — AEs cover ambiguity, not every R-ID.
Updated:
- requirements-capture.md section matrix: Acceptance Examples row
- requirements-capture.md trigger criterion paragraph
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>