framework_evolution — 2026-04-14
Summary
Analyzed Yeachan-Heo/oh-my-claudecode against our setup (functionally equivalent/superior on 13/16 dimensions) and executed full Tier 1 + Tier 2 cherry-pick: commit trailers protocol, ast-grep CLI v0.42.1 + skill, 4 user-scope agents (critic/architect/tracer/scientist at ~/.claude/agents/), ai-slop-cleaner skill, Serena MCP v1.1.1 as Permanent #11. Caught + fixed unexpected drift during Serena install: dual ~/.claude.json divergence (researchclaw/xint only in root config, not home — reconciled to 11 identical MCPs in both), npm —force overwrote system /usr/bin/sg symlink (restored to newgrp), brainstorm.service Restart=on-failure missed clean idle-exits (changed to Restart=always). Deep post-upgrade audit (5-cluster parallel pattern) surfaced 11 findings: 7 stale “Permanent 10” references, 4 orphan capabilities (new agents/skills not wired into protocols), 1 internal architecture.md contradiction, vague/terminology issues. All 11 resolved with cascade edits + protocol wiring. Commit trailers protocol validated end-to-end (first dogfood: commit 3bc1c2d with Constraint/Rejected/Directive/Confidence/Scope-risk). System state: 11 Permanent MCPs (sync verified), 63 skills (3 new: ast-grep/ai-slop-cleaner/serena), 4 user-scope agents, brainstorm.service active HTTP 200 Restart=always, 7/7 Universal Completion Gate PASS.