harness-checklisted

# Harness Check — Agent Framework Structure Audit You audit whether **the framework runs reliably as wired**. Core stance: **does "the system the docs describe" equal "the system the code can actually run"?** One broken glue point (a trigger, a registry, a symlink, a script seam) silently kills a chain while docs claim full functionality. Generic audit — assume no specific layout, role names, or scripts. Establish the target root first (user path or cwd). ## Flow (each step MUST print one `[harness-check] …` line) > No visible output = silently skipped; a missing print = step not done. > **Self-review flag**: this session already edited the audited harness → prefix every print with `[self-review]`, and a fresh sub-agent blind re-audit (target root, NO expected findings) MUST run before declaring done; print `[harness-check] self-review re-check: <agent> → <verdict>`. Cannot run → mark **unverified**, never silently. ### Step 0 · Discovery reality check (the lifeline) **Run this skill's `scripts/check_discovery.py <root>` first** — dead symlinks, glob coverage, name inventory. Cite its lines — script lines = leads, YOU verify (can't run → by hand, say so; exit 2 = bad root, fix the root, don't fall back). Then judge: - **Registry completeness** — hardcoded member/role lists ≟ the set on disk (vs its inventory). - **Naming consistency** — roster ≟ dispatcher/launcher name (mismatch → lookup fails). - **Forced-trigger targets** — every injected "must use Skill/tool X" reso