The next-action field in a project's frontmatter is supposed to be an executable pointer, two or three lines. Mine grew to 5,600 characters — added to a little at a time by my own agent, every single addition reasonable — until it went stale a full day behind the log sitting right below it. Co-maintained systems fail in ways neither party produces alone.
My daily review ran clean against a version of my day that was hours out of date. The tool had handed back a path to the note instead of the note, and a path is an invitation to skip the read. v0.5 of para-vault-mcp is about freshness: make the tool defeat the cache instead of asking the caller to remember.
A find_project query returned an empty array, and the LLM confidently told the user they had nothing — while four projects sat invisible behind one line of unquoted YAML. v0.4 of para-vault-mcp is about silences: wrong-but-confident answers, and the review discipline that fixes them even below the public bar.
Twenty years ago, a manager told me the fix I'd written to save a deal was a mistake. He wasn't wrong about the system — but my comp plan paid on closes, and the comp plan is the contract that tells you what your job actually is.
Two of para-vault-mcp's tools just absorbed three Bash ops I was running every evening to start the daily review. The shape they took came from the review, not the vault.
I asked my own MCP a routine question and it caught me lying to myself. The thing worth writing about isn't that the MCP worked — it's which thing it got right.
Forward Deployed Engineer is having a moment. The discourse is missing a cohort that has been doing audit-integrate-deploy-stay work for twenty years. AI did not create the role. It unblocked us.
I shipped one tool with a parameter. Eighteen hours later I split it into three. A day after that, I deleted one of the three. The lesson is that tool surface follows real use, not design.