Shipping 1.0, then the memory an agent can read

Two releases in one day. The morning was the road to 1.0 — proving the whole loop end to end and fixing what a pre-release audit turned up. The afternoon was v1.1, which added the feature I think matters most strategically: a way for coding agents to read postmortem’s memory. In between was a Windows bug that had no business being as annoying as it was.

The loop, proven — then a shim that wouldn’t spawn

1.0 is the smallest thing that delivers the promise: watch → detect → explain → predict, proven end to end rather than demoed in pieces. I cut it, tagged it, and immediately had to ship 1.0.1.

The bug was pure Windows. postmortem’s default brain shells out to the claude CLI, which on Windows isn’t an executable — it’s an npm/nvm shim (claude.cmd). Node’s spawn can’t resolve or exec a .cmd directly; it just ENOENTs, so the marquee “free AI via your subscription” path silently failed on the platform I actually develop on. The fix is to spawn with shell: true and pass a single validated command string (not an args array, which trips a deprecation warning). It’s now a documented rule: anything that shells out to a PATH command has to do the same. The tool that explains other people’s incidents having a silent one of its own on launch day was a fittingly humbling way to start.

Incident memory as an MCP server

The headline of v1.1 is mort mcp: a read-only Model Context Protocol server over the SQLite incident history. Point Claude Code or Cursor at the command mort mcp and the agent can call list_incidents, get_incident, query_events, and predict — risk-scoring the diff it’s about to write against every outage you’ve ever had.

This is the real thesis of the project, stated as a protocol. postmortem isn’t just a watcher; it’s the local ops-memory layer that agents plug into. The agent writing your code should be able to ask “have we broken this file before?” and get an answer grounded in your actual history. Read-only is a hard boundary — no database writes, no actuator triggers over MCP. Agents read the memory; they don’t pull the levers. That line is what makes it safe to hand an autonomous tool a connection to your incident record.

The contract pays off

Everything else in v1.1 was the NormalizedEvent bet cashing in. The Netlify sensor was built with the Vercel poller as a literal template — a new source of truth, zero changes to the brain, the database, or any output, because none of them care where an event was born. Auto-start units landed for launchd, systemd, and Windows Task Scheduler, all invoking the headless daemon. And Telegram alerts went out through a BotFather bot, the first output that leaves your terminal — formatted incident, ☠ and all.

Released as 1.1.0. The roadmap from here is v2: actuators that don’t just tell you what broke but can act — rollbacks, GitHub issues, PagerDuty — plus community sensors. The community builds sensors and actuators; the harness stays the product.