mirror of
https://github.com/AgentSeal/codeburn.git
synced 2026-10-31 19:40:54 +00:00
fe560f7f4d
Roo Code was discontinued in May 2026 (its repository is archived). Cached Roo sessions and daily history slices are dropped on load.
83 lines
3.1 KiB
Markdown
83 lines
3.1 KiB
Markdown
# Vercel AI Gateway
|
|
|
|
Cloud usage for [Vercel AI Gateway](https://vercel.com/docs/ai-gateway) via the reporting API.
|
|
|
|
- **Source:** `src/providers/vercel-gateway.ts`
|
|
- **Loading:** lazy (`src/providers/index.ts`)
|
|
- **Test:** `tests/providers/vercel-gateway.test.ts`
|
|
|
|
## Where it reads from
|
|
|
|
Not local disk. CodeBurn calls:
|
|
|
|
```
|
|
GET https://ai-gateway.vercel.sh/v1/report?start_date=...&end_date=...&date_part=day&group_by=model
|
|
```
|
|
|
|
See [Custom Reporting](https://vercel.com/docs/ai-gateway/capabilities/custom-reporting).
|
|
|
|
## Authentication
|
|
|
|
Set one of:
|
|
|
|
- `AI_GATEWAY_API_KEY`
|
|
- `VERCEL_OIDC_TOKEN` (from `vercel env pull` when using `vercel dev`)
|
|
|
|
## Caching
|
|
|
|
None. Each parse issues one API request for the requested date range.
|
|
|
|
## Deduplication
|
|
|
|
Per `vercel-gateway:<day>:<model>`.
|
|
|
|
## Not counted in totals by default
|
|
|
|
A report row is a **daily aggregate per model**: one cost, one token count and
|
|
one `request_count` for a whole day, with no request ids, timestamps or
|
|
attribution. Nothing in it can be matched against the local tools you pointed at
|
|
the gateway — Claude Code (`ANTHROPIC_BASE_URL`), Codex, OpenCode,
|
|
Cline/Kilo, Cursor — which already record those same requests from their own
|
|
session files. Counting both double counts the same spend.
|
|
|
|
So gateway spend is:
|
|
|
|
- **always shown** as its own provider row (marked "not in total" in the app),
|
|
and as `overview.excludedGatewayCost` in `--format json`;
|
|
- **excluded** from every headline total, per-model row, daily row and history
|
|
point by default;
|
|
- **unaffected** under `--provider vercel-gateway`, which reports the full
|
|
amount so you can inspect it.
|
|
|
|
The exclusion is one rule in one place (`parseAllSessions`), so every surface
|
|
follows it: the report, the interactive dashboard, the menubar payload, `models`,
|
|
`sessions`, `export`, `compare`, `compare-periods`, `spend`, `yield`, `audit`,
|
|
`budget`, and the Teams sync push (which would otherwise hand the backend the
|
|
same double count). The session cache and the daily cache are the deliberate
|
|
exception — they keep storing the gateway slice, because a past day's aggregate
|
|
row can never be fetched again.
|
|
|
|
To include it in totals instead:
|
|
|
|
```
|
|
codeburn gateway-totals include # codeburn gateway-totals exclude to undo
|
|
codeburn gateway-totals # show the current setting
|
|
```
|
|
|
|
The setting is read-side only. The daily cache always stores the gateway slice,
|
|
so flipping it applies retroactively to sealed days without re-fetching.
|
|
|
|
## Quirks
|
|
|
|
- Requires Pro/Enterprise Custom Reporting on your Vercel account.
|
|
- Data can lag by a few minutes after requests complete.
|
|
- Rows are daily aggregates per model, not per chat session (see above).
|
|
- `request_count` is used as the row's call count, so one row can stand for many
|
|
requests while carrying a single cost and token figure.
|
|
- `total_cost` is used as `costUSD`; token fields map directly when present.
|
|
|
|
## When fixing a bug here
|
|
|
|
1. Confirm env vars are set in the same shell running `codeburn`.
|
|
2. Reproduce with `codeburn report --provider vercel-gateway -p week --format json`.
|
|
3. Compare totals to the Vercel dashboard AI Gateway usage view.
|