[cabibbz]

A real event stream settled this. Two lines from it:

item_2 — "I'm checking the decision's external premises against primary/official sources, especially Windows capture semantics and the cited graphics-kernel proof of concept."

item_3, 115 seconds later — "The web-search attempt returned no usable source, so no external source was relied upon."

So: the tool existed, the policy permitted it, the reviewer used it, and it came back empty. Every previous diagnosis in this chain was looking at the wrong end of the pipe.

Cause

The mode was live, which fetches pages from the open web. Reviews run --sandbox read-only inside a Vercel Sandbox with no outbound egress, so those fetches find nothing.

Mode	Served from	Works here
cached (Codex default)	OpenAI-maintained index, pre-indexed results	✅
indexed	External access gated through the search index	✅
live	Fetches the most recent data from the open web	❌ needs egress

I picked live back when I reasoned "live is the only one that reaches the web." It reaches the web from a machine that can't. cached reaches the web through OpenAI, over the connection the model already holds.

Default is now cached. SOL_RESEARCH_MODE accepts indexed or live for a deployment whose sandbox has real egress.

This also narrows the security surface. docs/SECURITY.md warned that retrieved pages are untrusted input; with cached the reviewer reads pre-indexed results rather than arbitrary live pages, so the prompt-injection channel is smaller than what was documented. Updated accordingly.

The trace was also lying

That same stream contains no web_search item at all — despite a search having been attempted and reported by the model itself. A search that returns nothing need not be emitted as an event.

So "No search ran / the reviewer chose not to search" was an unfounded claim about the model's intent. It now reads:

No search observed — Research was allowed and no search was recorded. That can mean the reviewer did not need one, or that a search returned nothing and was never reported as an event. The review's own text usually says which.

Search mode was "cached".

Same correction applied to the cited-sources-without-a-search case, which was accusing the reviewer of fabrication on the same faulty inference.

Tests

• The worker validates the mode against the documented variants and defaults to one that works without egress
• applyWebSearchMode writes cached for research, disabled otherwise, and passes live through intact when a deployment asks for it
• The trace never claims a choice was made

107 unit tests (106 pass / 1 Windows-only skip), mocked e2e, lint, build clean.

---

Generated by Claude Code

[vercel]

The latest updates on your projects. Learn more about Vercel for GitHub.

Project	Deployment	Actions	Updated (UTC)
sol-review-gate	Building	Preview, Comment	Jul 27, 2026 2:23pm