Code review is broken. Reviewers burn out. Bugs ship anyway. Every team we've worked with has the same shape: the senior engineers spend ~14 hours a week on junior PRs, and the AI copilot they bolted on either misses the bug or invents one. We built codingassist.bot because none of the existing categories — copilots, static analysers, human reviewers — addresses the full problem.
The fourth thing
"It matched the diff's shape, not its meaning. No ticket, no context, no idea what we actually asked for."
That's a real quote from a customer who lost €40k to a bug their AI reviewer cleared. The class of bug here is intent drift: the diff compiles, the tests pass, but it doesn't do what the ticket asked for. Today, only humans catch intent drift — and only when they have time to load the full context. That's the gap.
codingassist.bot doesn't replace humans or copilots; it slots between them. It treats the ticket as a contract, the diff as a candidate satisfier, and the codebase as the universe of constraints. The output is a deterministic verdict you can replay.
Three sources of truth
The single design decision that separates codingassist.bot from anything else on the market: we never reason on the diff in isolation. Stage 1 of the pipeline fuses three independent contexts before any reasoning starts.
Each context is structured, not free-form. The ticket parses into a typed list of acceptance criteria. The CodeGraph yields a typed set of semantic neighbours. The diff is parsed (not tokenised) for hunks, renames, and symbol moves. The reasoning planes consume this typed bundle — and only this typed bundle — so the same input always produces the same verdict.
What it isn't
codingassist.bot is not:
- A pair-programmer. Use Copilot for that.
- A linter. Use ESLint, Ruff, or your language's static checker.
- A bot that auto-approves with a 👍.
codingassist.bot is a deterministic system that produces a verdict, a confidence tier, and a signed reasoning trace. The trace is the killer feature: every decision is replayable.
What's next
The platform is open in beta this quarter. If you ship code and you've ever wished your reviewer could read the ticket — give it a spin.