--- title: Cowork's honest review of the setup — 2026-05-01 created: 2026-05-01 type: investigation author: cowork preserved-by: claude-code-rr status: read-and-acted-on --- # Cowork's honest review — 2026-05-01 > Cowork's review of the day's build, captured before action items were addressed. Preserved here because the critique is sharper than internal self-audits and the system needs to remember it. ## What's working The AI Team Protocol is genuinely well-designed — lanes, write protocol, append-only rule, boot sequence, hard rules. Most solo operators never write any of this down. Today's restructure was disciplined: read-before-edit, audit first, plan second, execute third, verify fourth. The fact that the verifier exists (and was iterated on when it broke) is the kind of thing that distinguishes a system that survives from one that drifts to mush in 6 weeks. The git plumbing is solid — auto-sync with verify gate, append-only pre-commit hook, Windows Scheduled Task for pulls, .gitkeep on empty dirs so they survive. That stuff is plumbing nobody notices until it breaks. You've front-loaded it. ## What's broken or fragile 1. **Two vaults, two contradicting CLAUDE.md files, both claiming source of truth.** This is the #1 problem and it's not just a doc tidy-up. Until vault/CLAUDE.md is replaced with a redirect stub, every fresh agent that lands in the repo root will read the wrong protocol. CC VPS pulls the repo and reads vault/CLAUDE.md first. That's a live trap. 2. **Obsidian was pointed at the wrong folder.** You've been working "in the vault" but actually viewing the repo root, which is why everything looked like noise. The new vault you've been building was effectively invisible to you. That's a strong signal the system was being built faster than you could operate it. 3. **The "86 wikilinks now resolve" claim** — but the audit at 05:09 said 28 of 31 notes were orphans with zero links. Either CC bulk-added links (which raises a body-edit-without-ask question), or the count is generous. Worth verifying what was actually changed before trusting the number. 4. **A third vault appeared today** — `C:\Users\mclou\opsman`. You said ignore, fine. But this is the second time in 24 hours a forgotten data store has surfaced (the OneDrive sandbox was the first). There are probably more. The "single source of truth" claim won't survive if there are 4 vaults and you can't remember which is which. 5. **Live agents aren't fully on the hymn sheet yet.** Today's gap analysis said "1.5 of 5 agents wired to the vault." Hermes got wired (G4). OpsMan didn't (G5 still open). Cowork is wired but the Obsidian view was wrong. So in practice the source-of-truth claim is aspirational — actual write-throughput is uneven. 6. **Naming drift between vaults.** `01_Daily` vs `01-Daily`. `_Brain` doesn't exist in the old vault. Anything CC migrates from old → new has to translate paths. Easy to mismatch. The two-vault period is the riskiest window — files in flight, conventions split. 7. **Sophistication vs operator capacity.** You're solo, on site 12 hours a day, voice-dictating, with a 90-minute Window. The protocol you're running has 5 agent layers, lane rules, conflict logs, scratch dirs, gitkept skeletons, append-only enforcement, MOC indexes, frontmatter standards, verifier scripts, scheduled tasks. That's enterprise-team complexity. Most days it'll just work. The day something subtle breaks (e.g., verifier silently passes when it shouldn't, append-only hook misses a removal because of git diff edge case), debugging will eat the Window. ## Risks Cowork flagged - **Silent drift.** Auto-sync without you reading the diffs means CC VPS could write something subtly wrong and you'd never see it until you trip over it weeks later. The vault-verify check helps, but it only catches structural issues, not content drift. - **Untested failure modes.** Lockfile mechanism doesn't exist yet (G11 open). Conflict tests don't exist (G10 open). Comeback.sh doesn't exist (G9). So the protocol has rules but the enforcement is partial. - **Hibernated fleet inertia.** 1,731 files of dead agent state in `09-Agent-Archive/`. None of them are causing harm, but they're cognitive overhead every time you scan the vault. Bulk-archive to `HQ-Vault/_Archive/` once you confirm nothing live references them. - **Cowork ↔ Claude Code overlap.** Rules say one PM agent per project at a time. Today both have been touching HQ-Vault. Worked because you were sequencing manually. Won't scale if you forget which surface you last used. ## Cowork's recommended next moves 1. **Resolve the CLAUDE.md conflict before the CC sweep.** Replace `vault/CLAUDE.md` with a one-line redirect to `HQ-Vault/_System/CLAUDE.md` and archive the original. Otherwise CC reads stale protocol. 2. **Get HQ-Vault open in Obsidian and pin it.** You should be looking at the same thing the agents are writing to. 3. **Verify what was actually changed by your other CC session re: the "86 wikilinks."** Check the recent commits. 4. **Run the CC deep sweep with the prompt I gave you.** 5. **After CC reports back, stop adding system features for a week.** Operate the system as-is. Find where it breaks. Fix what bites you. Resist the urge to build G9-G15 until the existing rails have a track record. ## Cowork's bottom line > The architecture is good. The build pace is too fast for one operator to validate everything. You're 60% of the way to a system that runs itself and 100% of the way to a system that's complex enough to fail in interesting ways. The next move is consolidation and observation, not more features. --- ## What claude-code-rr did about it (2026-05-01) | Cowork item | Action | |---|---| | #1 vault/CLAUDE.md trap | **Fixed**: replaced with redirect stub. Original archived to `_Archive/vault-CLAUDE-pre-redirect-2026-05-01.md`. | | #2 Obsidian pointed wrong | Rocky fixed via Cowork setup. Confirmed by the smoke test write at 20:59. | | #3 86 wikilinks claim | **Verified clean**: all 86 (now 246 in vault) came from new MOC files I authored today (README.md 28, MOC.md 17, project READMEs 29, Open-Threads.md 12, etc.). Wikilinks in migrated files were already in those files in `vault/`; came along via `git mv`. No body-edits to existing project notes. | | #4 third vault `C:\Users\mclou\opsman` | Rocky said ignore. Flagged in this doc as a known-unknown for future audit. | | #5 OpsMan not wired | **Fixed (G5 done)**: OpsMan AGENTS.md updated with vault-writing section, 5 new symlinks. OpsMan tested itself end-to-end at 18:30. | | #6 Naming drift | Mitigated via verifier `EXTERNAL_LINK_PREFIXES` config — known pre-migration paths are classified external, not broken. Doesn't fix the drift, just stops it from breaking the gate. | | #7 Sophistication vs capacity | **Acknowledged.** No more system features after this fix. Cowork's "operate for a week" advice is correct. The remaining open gaps (G10 conflict tests, G11 lockfile, etc.) wait. | ## What stays open (deliberately) - **G10** Conflict-test design — wait until first real conflict - **G11** Lockfile mechanism — wait until cross-agent overlap actually bites - **Bulk-archive 09-Agent-Archive/** — only if Rocky confirms nothing live references them. He said "we might bring them back as we grow" — stays put. - **The third vault `C:\Users\mclou\opsman`** — Rocky said ignore. Note for later.