plan now lists every external request the Build will make, one entry per
request, and runs the plan's Producers that need nothing from outside so
each request's exact constraints exist before submission. Parameters are
read off the request itself (scalars by their own names, prose as a word
count, referenced media counted by kind); no model-specific field table
lives in the CLI. A request whose references are still to be generated
reads the author's request draft upstream and says the references are
pending. Requests with the same shape are grouped and counted, word counts
shown as a range.
Core scheduling becomes incremental: a ready Producer no longer waits for
an unrelated external request that shared its scheduling turn, which is
what let plan-time evaluation reach every request and also stops Builds
serializing cheap steps behind long paid ones.
Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
kashorin