Constantine Makoid

Mine starts with a number too: 220 seconds. That's how long it took, on a real
dependency advisory, from pressing a button to an open pull request the alert
landed, an embedded coding session opened already holding the repository and
the package, edited an isolated checkout, and the runner pushed the branch. I build Celmis: one tree-sitter symbol graph across a set of repositories, read
by four surfaces cross-repo Q&A with citations re-checked against the files
on disk, a pull-request reviewer, generated documentation, and an MCP server.
Self-hosted, AGPL-3.0, one machine under docker compose. The number I'd rather not hide: 17th of 50 on a public PR-review benchmark. I
audited all 79 findings the benchmark scored as false positives against the
source and published the verdicts 33 turned out to be real defects the gold
set missed. A wrong number said honestly is worth more than a good one said
quietly. Launching Thursday, 00:01 PT.

Scheduled for Thursday, 00:01 PT. Before it goes live, the decision I'm least sure about: "Fix from here" opens a full agent session rather than filing an issue or opening a draft PR directly. The session edits an isolated checkout, runs, verifies, and only then pushes a branch and opens the pull request a person still merges it. Slower than a one-click auto-PR, on purpose: I wanted a point where a human reads what actually changed before it enters a checkout, not after. Genuinely unsure this is the right tradeoff for everyone. If you've built something in this space where would you make the agent commit faster, and where would you make it slower?