[opencode-agent]

Problem

The stats sync counted every Go/Free generation.completed event, including requests rejected before routing (attempts = 0). Those requests have no route, no endpoint and no usage, but their model_requested string still became a model dimension and got a page on /data.

One rejected request on 2026-09-19 for the nonexistent glm-5.5-flash (HTTP 400, 0 attempts, no workspace) produced /data/zhipuai/glm-5.5-flash. That page is now being shared as evidence of an upcoming model.

Change

• Add attempts > 0 to the source filter of the usage/geo and retention queries, so only requests that reached an upstream route are counted.

R2 check: across all generation.completed rows since 2026-05-28, every attempts = 0 row has no route, failed, and has null usage, and attempts is never null. The filter therefore drops no tokens or cost.

Effects

• The lake contains no attempts = 0 events after 2026-09-20 13:44 UTC. For current traffic, this is a guard rather than a behavior change.
• Inside the display window (2026-09-02 to 2026-09-20), the daily history pass will stop counting pre-routing rejections, mostly Free-tier 429/403s and Go 400s. Request, session, unique-user and error counts for those days drop accordingly. Token and cost totals stay the same.
• The sync only upserts, so rows that already exist for a dimension no fresh query returns are not deleted. The existing glm-5.5-flash row stays until it leaves the 56-day window (about 2026-11-14) unless it is added to the StatsHiddenModels secret.

Verification

• bun test src/domain in packages/stats/core: 34 pass, including the new query test
• bun run typecheck in packages/stats/core
• oxlint on the changed files: clean

Requested by: @adamdotdevin (Adam via Slack)

[opencodejr]

The change looks sound. Both lake queries, usage/geo and retention, now share the same attempts > 0 guard, placed next to the existing product filter. A request rejected before routing can no longer create a model dimension or a /data page.

I couldn't run this against the R2 lake: there's no access from the review environment, and the lake schema isn't in the repo. So I couldn't confirm that attempts is a column there, or that rows with attempts = 0 never carry usage. If that holds, token and cost totals are unchanged.

bun test src/domain in packages/stats/core passes (34 tests), and tsc --noEmit is clean. The package's typecheck script (bun check --no-script) fails with "Script not found" on the base branch too, so that's unrelated to this change.

Public API: unchanged.