mirror of
https://github.com/EveryInc/compound-engineering-plugin.git
synced 2026-09-19 01:09:55 +08:00
Trevin Chow
c828af5dcd
Default ce-doc-review at ce-plan Phase 5.3.8 to mode:headless so safe_auto fixes apply silently and remaining findings surface as a one-line summary above the post-generation menu. Add Run deeper doc review as a first-class menu fixture so interactive engagement is opt-in but obvious; hide it when no actionable findings remain (FYI-only state included, since the walkthrough is gated to anchor 75/100 gated_auto/manual findings and would otherwise dead-end). Tighten the adversarial activation trigger so routine plans no longer trip it just for having structural rationale; activation now keys on real high-stakes signals (high-stakes domains, new abstractions, missing or extended origin, requirements-shape premise content, explicit alternatives). Move document-type classification into the orchestrator using content-shape signals (frontmatter, R-IDs vs U-IDs, section structure) with file path as a tie-breaker only. Add an Origin: slot to the subagent template so personas that adapt on origin (product-lens, adversarial, scope-guardian) read it authoritatively instead of re-parsing frontmatter themselves -- one classification, all personas trust the slots. Per-persona doc-type adaptation: - product-lens, adversarial, scope-guardian suppress premise/prioritization/ simplification techniques on plan-shape docs with Origin: set, since the brainstorm validated those upstream - feasibility scopes shadow-path tracing, implementability, and migration mechanics to plan-shape docs only; requirements-shape gets a tight check for fundamental-rework signals - coherence, design-lens, security-lens carry light doc-type framing (different identifiers and granularity expected per type) since they are net-positive at both levels Net effect: routine plans dispatch the cheap minimum (coherence + feasibility), high-stakes plans dispatch the full team, and the user is never asked to adjudicate findings from auto-runs unless they explicitly opt into the deeper review. Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>