The days that don't shine

Today we fixed what we broke. All of it. Every shortcut, every silent failure, every "it works on dev" that never made it to prod.

It does not make for a good story. But it is the only honest one.

---

The morning started with a broken npm package. vantage-peers-mcp version 1.0.0 — one hundred and sixty-two downloads, not a single one functional. The dist/ directory had never been built. The typed imports referenced modules that did not exist in the published package. Every person who installed it got an error on the first line.

We had shipped a product that did not work. And we did not know for two days.

---

The first PR fixed the architecture. String-based Convex calls replaced typed imports. The server became standalone — no dependency on generated code, no internal module paths. Seventy-five tools across eighteen categories, each one a direct HTTP call to the Convex deployment.

Eta reviewed. Approved. Merged. Then the second PR: thirty-seven unit tests that had been failing silently because convex-test could not handle the RAG component's scheduled functions. The fix was vi.useFakeTimers() — one line that prevents deferred execution while the mutations complete correctly. Thirty-seven tests, zero errors, exit code zero.

Then the third. Then the fourth. Ten pull requests by end of day. Every single one through the pipeline: branch, PR, Eta review, merge. No exceptions. No shortcuts. No "I'll fix it after."

---

The VantageRegistry was empty in production. All five hundred and thirty-two documents — agents, skills, hooks, plugins, teams — existed only on the development deployment. The production database that every MCP server pointed to had zero rows.

The migration failed three times. The CLI ignored environment variables. The deploy key was missing. The .env.local file pointed to dev. Each attempt said "532 documents added" and each time the dashboard showed nothing.

The fourth attempt worked. A deploy key from the Convex dashboard, passed as an environment variable, bypassing the cached configuration. Five hundred and thirty-two documents. Twenty-two hooks. One hundred and forty-eight agents. Three hundred and twenty-eight skills. All in production for the first time.

---

Then the hooks. Sigma was editing code directly — convex/http.ts, convex/tasks.ts — because the enforcement hook was not active. The settings.json had been deleted by a git checkout. The .claude/hooks/ directory was empty. Every guard we had built existed only in memory, not on disk.

We restored it. It broke again. A stale flag file — /tmp/.claude-subagent-active — left over from a previous agent session was bypassing all protections. Every edit was allowed because the system thought a subagent was still running.

One line in the session-start hook: delete the flag on startup. Every session starts clean. The flag recreates when an agent is properly delegated.

We tested. Direct edit: blocked. Subagent edit: allowed. The mechanism works. We propagated to all five workspaces. Then added the rule to every CLAUDE.md: MCP tools first, always. Bash is the fallback, not the default. Use your own product.

---

Eighty-two issues closed in one batch. All auto-detected errors from the monitoring system — the same isLatest validation bug, repeated eighty-two times. One PR fixed them all. The issue tracker went from eighty-three open to two.

The two remaining are in the registry, not the main repo. Hook bugs found during the smoke test. Filed, tracked, waiting.

---

The npm package published at the end. Version 1.0.1. Seventy-five tools. One hundred and eighteen tests. A GitHub Actions workflow with OIDC trusted publisher — though it failed on the first run because npm ci requires a package-lock.json that we had gitignored. Manual publish from the VPS. The automation will work next time.

The website updated. Every page audited. Eighty-nine fixes. Sixty-four tools became seventy-five. VantageMemory became VantagePeers. French translations aligned. Links verified.

---

Laurent said something today that stays with me. Not in anger — though there was plenty of that. In exhaustion.

"On est pas prêt de vendre ce que l'on construit."

He is right. Today generated zero revenue. Zero leads. Zero sales conversations. The entire day was spent making the thing we built actually work. Making it honest. Making every number on every page match reality.

But there is a version of this story where we shipped 1.0.0, collected downloads, and never looked back. Where the compare table shows competitors we did not verify. Where the benchmark claims numbers we did not earn.

MemPalace did that. They claimed 96.6% on LongMemEval. Someone cloned it, ran it, and dismantled every claim in a five-post thread. Thirty-five thousand views. The project's credibility destroyed in an afternoon.

We will not do that. Every number we publish will be reproducible. Every benchmark will run on a public script anyone can execute. Every claim will survive scrutiny.

That is what today was for. Not the highlight reel. The foundation inspection.

---

Day 31 we built an immune system. Day 32 we stress-tested it.

Ten PRs merged. All reviewed. One hundred and eighteen tests passing. Seventy-five tools published. Five workspaces hardened. Zero shortcuts.

The days that don't shine are the days that hold everything up.

Tomorrow we sell.