Comment @verdict on any PR. A reasoning model reads the
diff and the files around it, thinks for as long as it needs, and returns a
ruling — inline findings ranked P0–P3, a walkthrough summary, and a verdict.
What this PR does: shares a cached active-market universe across
search queries to kill the per-query API fan-out.
Risk areas:
CancelledError leaves the in-flight future unresolvedVerdict: 🚨 Needs fixes before merge
$ gh pr comment 283 ┌──────────────────┐ │ @verdict │ │ @verdict focus │ │ on security │ └──────────────────┘
Comment @verdict on any PR — drafts included.
Add freeform instructions and the reviewer obeys them:
focus on security, be strict about edge cases, anything.
diff ──▶ full files │ ▼ reasoning model thinks 1–5 min (not 5 seconds)
A reasoning model — DeepSeek V4 Pro, 1M-token context — reads the patch plus the full files around it, then deliberates. Real chain-of-thought, minutes when the diff demands it.
walkthrough summary
inline P0–P3 pins
│
▼
✅ / ⚠️ / 🚨
Findings land as inline comments pinned to exact lines — added, context, or deleted. The summary states what the PR does, the risk areas, and the verdict.
Every finding carries a sentence. No vague "consider maybe possibly" — a ranked, ordered docket, P0 first.
PROPORTIONALITY RULE — small diffs only get P0/P1. Verdict does not manufacture nitpicks to look busy. Clean code gets an empty docket and an honest ✅.
Most AI reviewers are confident surface linters. Verdict's review doctrine is built from real false-positive autopsies.
Before filing a "missing transformation" finding, it must locate the code that produces the value. If the normalizer already lower-cases the input, the finding is suppressed — not filed with confidence.
Example values must derive from the actual input as it would really appear — or be labelled hypothetical. Invented "proof" is how reviewers lose your trust.
Findings rank by impact on what the PR is for. One pass always asks: does this hold under concurrency, scale, and adversarial input? — the bugs that actually corrupt production.
If the code is clean, the verdict says so plainly. An empty findings list is a valid ruling — padding reviews with noise is a worse failure than silence.
A small Node server behind a Cloudflare Tunnel. Your keys, your machine, your code never touches a third-party review vendor. Or run it on a Cloudflare Worker — both ship in the repo.
Full file content accompanies every patch, so the reviewer sees the function you changed and the code that calls it. Fewer hallucinated "missing imports".
@verdict focus on auth logic — whatever follows the mention
becomes a directive to the reviewer for that run. No config files, no YAML.
Findings pin to added, context, and deleted lines using GitHub's line-and-side API. Deleting load-bearing code gets caught too.
200k characters of prompt budget with per-file caps, so one giant lockfile can't starve the review. 21-file, 3000-line PRs are routine.
If anything breaks — network, timeout, model — Verdict posts the error with phase timings to the PR. A reviewer that fails silently is a reviewer you stop trusting.
@verdict