[mikerowe23]

I'm on the Codex team. This moves two of the bench's Codex subscriptions to our new edge node, iad.edge.codex.io: Base, and Robinhood Chain at the Processed commitment level. Solana stays on fra.edge.codex.io and BNB on graph.codex.io. The change is in two commits, so you can take Base without Robinhood.

Changes

• Base (first commit): subscribes on wss://iad.edge.codex.io/graphql from every region, at Processed as before. CODEX_BASE_WS_URL overrides the endpoint; setting it to wss://graph.codex.io/graphql rolls Base back without a code change.
• Robinhood (second commit): subscribes on wss://iad.edge.codex.io/graphql at Processed instead of the default level. At the default level iad is no faster than graph.codex.io (about 50 ms slower from Singapore), so the gain comes from the Processed stage. graph.codex.io doesn't serve that stage on Robinhood: it accepts the subscription and delivers nothing. So the level follows the endpoint, and CODEX_ROBINHOOD_WS_URL=wss://graph.codex.io/graphql rolls back both.
• Sockets: each edge chain gets its own socket, so each probe holds four Codex sockets on the same API key (Solana, Base, Robinhood, and BNB on graph.codex.io). Base and Robinhood share the node but not a socket. Since head-lag: a full race book no longer eats pool trades; Codex edge errors end the subscription #3256 an edge error ends the whole socket, and a silent pool trips the flow watchdog for everything on it, so a Robinhood problem can't disconnect Base. All four feed the existing ws_connected{aggregator="codex"} series, which reads 1 only while every socket is up. The spec queries are unchanged.
• Code: codexSocketPlan splits the chains across sockets and codexProcessed picks each chain's level; both are tested. The socket log tags are [solana], [base], [robinhood] and [graph] (previously [edge] and [graph]).
• Spec: the "Codex endpoints" methodology line names both edge nodes. A new "Codex commitment levels" line says Robinhood reads Processed and that earlier Robinhood figures read the default level on graph.codex.io. It's two lines to stay under the 500-character limit.

Verified

Base (2026-10-22). I ran a probe in each of the bench's probe regions on Railway (us-east4, europe-west4, asia-southeast1) for about 15 minutes. Each one subscribed the bench's Base pool at Processed on both iad.edge.codex.io and graph.codex.io and compared arrival times trade by trade.

Region	Trades on both	iad.edge first	iad.edge lead p10 / p50 / p90
us-east	139	138	192 / 223 / 309 ms
eu-west	120	119	186 / 221 / 321 ms
sgp	120	119	138 / 174 / 278 ms

Both sockets stayed connected for the whole run in every region. In each region one trade reached graph.codex.io but not iad.edge.codex.io within the 15 s matching window.

Robinhood (2026-10-27). I ran probes in the same three regions for about 10 minutes, each subscribing the bench's Robinhood pool on three setups at once: today's (graph.codex.io, default level), this PR's (iad.edge.codex.io, Processed) and iad.edge.codex.io at the default level.

How far this PR's setup was ahead of today's, trade by trade:

Region	Trades	Ahead	Lead p10 / p50 / p90
us-east	139	139	57 / 79 / 139 ms
eu-west	139	139	55 / 74 / 136 ms
sgp	138	136	8 / 30 / 89 ms

The bench's own figure (arrival minus block timestamp), mean over the same trades:

Region	Today	This PR	Change
us-east	766 ms	674 ms	−92 ms
eu-west	804 ms	716 ms	−88 ms
sgp	827 ms	785 ms	−42 ms

iad.edge.codex.io at the default level ran 1–4 ms behind graph.codex.io in us-east and eu-west, and a steady 51 ms behind in sgp. Every subscription stayed connected in every region, with no errors.

Harness and tests. The branch ran locally with the Codex key: all four Codex sockets connected on the first attempt, every chain delivered, and ws_connected{aggregator="codex"} read 1. go test ./cmd/script (with -race on the Codex tests), pnpm validate and bun test src/ pass.

Deploy notes

• No new secrets: all four sockets use the existing CODEX_API_KEY. Redeploy the three head-lag services (us-east, eu-west, sgp).
• Expect Codex's Base figures to drop by roughly 0.2 s, and its Robinhood figures by roughly 40–90 ms depending on region. The methodology lines say why.
• The two methodology lines are undated. You may want to add the deploy date, as the other dated notes have.
• I rebased the branch onto current dev (a705fe05) since opening the PR.

🤖 Generated with Claude Code