Batch security reviews into one process: 61 invocations paid 61 cold prefixes #174

New issue

Closed

opened 2026-08-15 17:15:57 +00:00 by john · 0 comments

john commented

2026-08-15 17:15:57 +00:00

Owner

A 25-turn session never amortizes its cold start. Measured, 3.x era:

opus-5 main reads 128K per call against 3K written
security-review reads 37.8K against 8.2K

One process reviewing N diffs pays one cold prefix instead of N. 61 invocations paid 61 cold starts.

This is the second exit on #173 and the one reachable without an upstream change: batching also lands consecutive invocations inside the 5-minute window, so it fixes the TTL mismatch as a side effect rather than needing the TTL raised.

Open question: whether a batched review over N diffs finds what N single-diff reviews find. A per-commit review is scoped to one change on purpose, so this is a real trade rather than free.

A 25-turn session never amortizes its cold start. Measured, 3.x era: - opus-5 main reads 128K per call against 3K written - security-review reads 37.8K against 8.2K One process reviewing N diffs pays one cold prefix instead of N. 61 invocations paid 61 cold starts. This is the second exit on #173 and the one reachable without an upstream change: batching also lands consecutive invocations inside the 5-minute window, so it fixes the TTL mismatch as a side effect rather than needing the TTL raised. Open question: whether a batched review over N diffs finds what N single-diff reviews find. A per-commit review is scoped to one change on purpose, so this is a real trade rather than free.