Knowledge transfer · Nov 2026

Handover: keeping @ai running

I'm away from 12 to 24 November. This page is for whoever covers: where @ai lives, how to ship a fix, and what to do when someone says it's broken. For how it all works, see the KT.

Where it lives

Production: ai.bhaskarapp.com. Staging: dev.agents.bhaskarapp.com, a Vercel preview alias.
Vercel project: agents. Functions and sandboxes run in sin1.
Repo: dbdigital-vercel-programmatic-sandbox/agents. Changes reach main through PRs.
Data: Neon Postgres (DATABASE_URL) and Upstash Redis (memory, spend reports, caches).
Slack connects through Vercel Connect: slack/ai in production, slack/appbuilder on preview. There are no Slack env vars.

One gotcha: schedules and triggers only run in production. Vercel Cron doesn't fire on preview deployments, so a schedule made on staging never runs.

Shipping a fix

Open a PR to main. Before merging, run pnpm typecheck, pnpm build and pnpm build:eve.
The build doesn't run migrations. If the PR adds a file in drizzle/, apply it with pnpm db:migrate (or from /cms/migrations) before the new code serves traffic. /cms/migrations only opens for the emails in appconfig.json → migrationAdminEmails, so add yours.
Rolling back: promote the previous production deployment in Vercel. Migrations don't roll back with it (the revert route only deletes the bookkeeping row), so keep schema changes additive.
Don't upgrade eve. It's pinned at 0.87.0 with our patch in patches/, and going back below 0.83 breaks sessions created after it.

When someone says it's broken

They see	Where to look	What to do
"Your rolling AI budget is exhausted"	Their row in `/cms/users`	Raise their AI budget.
"Budget verification is temporarily unavailable"	Vercel logs, Upstash and Neon	The spend lookup failed (Redis, the database or the Gateway report). It clears once they're back.
They can't use @ai at all	`/cms/users`	Tick their AI access.
@ai answers messages meant for people, or ignores ones meant for it	Observability → Jev → `slack-follow-up`	Relax the mention screen first, not the 0.8 threshold.
A new chat can't get a sandbox	Logs: "Vercel Sandbox create failed"	The log has the Sandbox API's error code. `snapshot_region_mismatch` means the template snapshot is in another region: `scripts/snapshots/README.md`.
A schedule didn't run	The schedule's page shows its last error	Production only. A failed start is retried after 5 minutes.
A trigger didn't run	The trigger's page shows its last error, including Jev's skip reason	No error means the message never matched: check its channel, author and filter.
A CloudWatch query is blocked	Their CloudWatch budget in `/cms/users`	Raise it. `CLOUDWATCH_GUARDRAIL_ENABLED=false` turns blocking off; spend is still recorded.
Errors or slow replies from the model	Logs: "Model call failed", "AI Gateway used a fallback route"	Every call has a fallback model, so a provider outage shows up here first.
A bug report	Observability → Bugs	Filed from the web form or the Gluon MCP's `report_bug` tool.

Where to look

Vercel logs. Each line is a fixed message plus JSON, so search for the message, or for a session ID (wrun_…).
Observability (/observability, admins only): overview, users, reliability, events, Jev and bugs.
Tables, read-only: analytics_events (every turn, model call and tool call), jev_evaluations, bug_reports, trigger_fires, scheduled_task_runs, background_commands.
No error tracker. The app has no Sentry SDK, so server errors only reach Vercel logs.
EVE_LOG_LEVEL=debug logs how long each part of a turn's setup takes.

Emergency levers

Cut off a user: untick their AI access in /cms/users, or set their budget to $0.
Stop an automation: the owner can pause it on its Schedules or Triggers page. For someone else's, set its status to paused in the database.
Turn off an integration: blank its token env var and redeploy.
Turn off the Gluon MCP: block /mcp.

Accounts it depends on

Vercel: hosting, Sandbox, Cron, Workflow, AI Gateway, and Connect for Slack and the MCP connectors.
Neon and Upstash.
GitHub: read, push and PR tokens for each org. If one expires, clones, pushes or PRs fail. The push token for dbdigital-vercel-programmatic-sandbox is Naman's, and it also merges PRs there.
AWS: S3 for artifacts, and the role the AWS MCP assumes.