mirror of
https://github.com/EveryInc/compound-engineering-plugin.git
synced 2026-09-19 01:09:55 +08:00
Trevin Chow
86b9748ad7
Reduce always-loaded SKILL.md sprawl (347 -> 285 lines) and remove single-source drift, while fixing two latent contract issues found in an independent dual-model (Claude + Codex) skill evaluation. Fixes: - Path A/B contradiction: the already-clear fast path labeled the synthesis "announce-mode" unconditionally (old SKILL.md:146), conflicting with the Phase 2.5 / synthesis-summary.md tier guard (Path A only for Lightweight + no blocking questions). A richly pre-loaded Standard/Deep prompt could skip its confirmation gate. Fast path now defers to the reference's Path A/B gate. - Stale path: visual-probes.md pointed the durable artifact at docs/brainstorms/ instead of the Phase 3 unified plan under docs/plans/. Consolidation / placement: - Visual-probe gate: Phase 1.3 is now the single in-body owner; the gate's full contract (state-based timing, blocking-question precedence, ASCII-mockup-never- satisfies) lives in visual-probes.md; the four satellite restatements (Interaction Rule 4, Phase 0.3 tripwire, redundant 1.3 bullet, Phase 2) are reduced to pointers. - Extracted branch-specific detail to phase-loaded references: verdict-routing.md (0.1c), product-pressure-test.md (1.2 lenses), model-tiers.md (dispatch tiers). Output-mode (0.0) intentionally kept in body: its config read depends on the Claude Code `!`-backtick pre-resolution that only fires in a loaded SKILL.md. - De-duplicated the repo-relative-path rule (kept owner in brainstorm-sections.md; clarified artifact paths repo-relative vs chat report absolute). Hardening: - Phase 1.3 exit condition made observable (checklist vs "until the idea is clear"). - Visual helper now invoked via the SKILL_DIR Tier-3 anchor, matching the repo-profile-cache convention. Updated the guard test that asserted the superseded <resolved-helper-path> placeholder. - Frontmatter description gains a quiet-condition for already-specified work; spawn_agent reference phrased as a capability example. Verified: bun test 1686/0, release:validate green, and a skill-creator regression eval (4 changed branches x edited-vs-pre-refactor) showing identical correct routing on both versions (no behavior regression).