ai-config Worldclass Upgrade Audit — Completion Log

Task span: 2026-07-16 → 2026-07-17 (7 phases + follow-ups). Spun off from Fable-usage (Fable 5 deep-review candidate #3). Pattern: Fable 5 plans → cheaper model executes → Fable 5 verifies.

What was decided

Plan approved 2026-07-16: 6 phases ordered SSOT-completeness → docs truth → SMTM skills → dev-project rule files → global AGENTS.md → peripherals; per-phase model recommendation (Sonnet for mechanical, Fable/Opus for judgment). Phase 7 (fresh-session Fable verification) added mid-flight.
project-* skill trio: updated to STATUS.md-driven SMTM v5.2+ (not deprecated) — Talbot will use discrete projects going forward.
Skills under ai-config deploy with per-dir rsync --delete (kills the cp-not-mirror trap).
Local FOSS model routing: parked in UPGRADES — not for vault-writing tasks (agentic reliability too low); niche = read-only overnight work with morning review gate.
Model-check gate: task frontmatter model: + skills check own identity, stop on mismatch (agent can detect, can’t switch). Landed in task-start/task-continue.
Shared skills serve ALL agents (CEO directive 2026-07-17): skills promoted from claude/skills/ to tool-agnostic top-level skills/ SSOT; deploy.sh fans out to Claude Code (WSL + Windows) + AntiGravity (global + monorepo workspace). Future agent surfaces get added to the fan-out — content is never forked per tool.
Windows Claude hooks gap accepted: hooks (SessionStart inject, SessionEnd digest) are bash+qmd+WSL-path bound; no PowerShell port. Run sessions from WSL. Documented in ai-config UPGRADES.md.

What changed (by commit)

Commit	Repo	Content
`8da02ef`	ai-config	Phase 1: 10 untracked commands + 12 user skills migrated into repo; plugins-manifest; deploy.sh skills step
`d6e5671`	ai-config	Phase 2: README truth pass (config-ai→ai-config, hooks/skills/plugins reality); Component-SSOT table; STATUS/UPGRADES portal paths
`ab55367`	ai-config	Phase 3 mechanical: labels, qmd wiring (task-prep/task-from-chat/task-complete), model-check gate
`6275fe7`	ai-config	Phase 3: task-from-chat dept routing; project-* trio aligned to SMTM v5.4
`ec47ceb` / `67a8c1d` / `f5c7021`	monorepo / web-deploy / KB	Phase 4: AGENTS.md=fat SSOT + CLAUDE.md=4-line pointer inversion; web-deploy lessons → LESSONS.md; sd-app dead-path fix; MBR CLAUDE.md v5.4 alignment
`6c7f302`	ai-config	Phase 5+6: AGENTS.md dept-routing rules refresh, 2 new hard rules (sudo hand-off, move-aside), cursor/antigravity frozen, Obsidian filter cleanup, Inbox hand-offs
`efba0a7`	ai-config	Phase 7 verification: doc-truth fixes (STATUS/ROADMAP/UPGRADES bookkeeping lag)
`6076e97`	ai-config	Skills → tool-agnostic `skills/` SSOT + all-agents deploy fan-out; graphify block rewritten tool-agnostic in AGENTS.md; Skill-tool mechanics → claude/CLAUDE.md wrapper

Phase 7 verification verdict

Fresh Fable 5 session re-ran the ORIGINAL audit scope (not intervening conclusions): zero substantive gaps — stale-path greps clean, labels clean, source↔deployed byte-identical, wrapper pattern correct across 4 dev repos, KB-OS split holds, Skills Dashboard current. Only bookkeeping lag found (fixed in efba0a7). World-class confirmed.

Gotchas / warnings

ai-config post-commit deploys the working tree, not the commit — one phase = one commit = clean tree, always. (Pre-existing rule; held throughout.)
Doc-truth lag is the recurring failure mode: STATUS/UPGRADES/ROADMAP checkboxes lagged work already shipped — tick trackers in the same commit as the work.
Deployed-path references (~/.claude/skills/...) survived the skills SSOT move unchanged — only source-path references needed the sweep. Deploy targets are the stable interface.
AntiGravity-specific skills (aic/ptr/ss, antigravity/skills/, lowercase skill.md) remain a separate frozen set alongside the shared fan-out — no collisions.

Open threads handed off

KB-OS ROADMAP checkbox lag + M4 task-from-chat half-landed → Processes Inbox (Phase 6.4).
Future agent surfaces (Gemini CLI, Cursor commands) → extend deploy.sh fan-out per UPGRADES.md note.