Repository navigation
session-flow: reconcile orchestrate's token-multiplier figures and annotate the condensed-return norm #3564
Description
Context
Signed-off candidate C8 of the context-engineering integration contract (2026-09-01): the
orchestrate skill already states a compressed-verdict return norm; the corpus adds the
1,000-2,000-token condensed-return figure (Anthropic context-engineering post, byte-verified
slice) and the multi-agent research post's economics (~4x chat tokens per agent, ~15x
multi-agent, token usage explaining most eval variance). Devils-advocate flagged a numeric
conflict to resolve first: orchestrate's existing line says "Multi-agent costs 3-10x the
tokens".
Proposed work
- Reconcile the 3-10x line with the ~15x source figure (cite both, or bound the range with
provenance) inside plugins/session-flow/skills/orchestrate/SKILL.md. - Annotate the compressed-verdict line with the 1-2k condensed-return norm and its citations
(the engineering post's sub-agent section; the multi-agent post's economics). - Keep the edit small; this is an annotation, not a rewrite.
Acceptance criteria
- No two token-multiplier figures in the skill without provenance distinguishing them.
- The condensed-return norm carries its two citations.
References
- docs/topics/context-engineering-integration/PLAN.md (C8 constraints)
- plugins/session-flow/skills/orchestrate/SKILL.md:44,149-154
Generated by Claude Code
Activity
feat(conventions): absorb the context-engineering corpus and add the …
kyle-sexton commentedon Sep 2, 2026
Done in #3588, merged to
mainasae3aa802. Both acceptance criteria are satisfied.The 3-10x line is no longer a bare figure: it is labelled as this plugin's own operating figure
and as a floor rather than a ceiling, with the upstream measurement cited beside it (~4x a chat
interaction per agent, ~15x multi-agent, token usage explaining most of the performance variance
regressed), plus the instruction to size research-shaped fan-outs against the higher number. So
the two figures now carry the provenance that distinguishes them.
The compressed-verdict line carries the 1,000-2,000 condensed-return shape with its citation, and
the caveat that it is a target shape rather than a budget to spend up to — the return norm was
otherwise unfalsifiable. The multi-agent economics citation sits on the multiplier line above it,
so both sources are present with attribution.
The edit stayed an annotation, as asked; the only other change to the file was rewrapping a
paragraph the inserted citation had turned into a run-on line.
Generated by Claude Code