Production checklist

Use this checklist before the first production release.

Runtime

• Pin Flary, Flue, Pi, and Cloudflare adapter versions.
• Run typecheck, unit tests, Worker integration tests, and the clean npm installation test.
• Apply the D1 run migration and verify the Flue Durable Object migration.
• Set an explicit default mode, model, cache policy, and run limits.
• Confirm the production R2 binding exists when large files or history need it.

Security

• Authenticate every route before resolving TrustedRunContext.
• Derive tenant, application, project, agent, and identity from trusted server data.
• Keep provider credentials and MCP credentials out of prompts, events, logs, and API responses.
• Use encrypted secret storage and a single refresh lock per connection.
• Authorize every R2 operation. Do not treat a tenant prefix as a security boundary.
• Block private-network MCP destinations and unapproved outbound access.
• Require approval for writes, deletes, secret access, external messages, commits, merges, deploys, and Sandbox imports.

Durability

• Submit an idempotency key for every run and state-changing tool call.
• Persist the run before starting model or tool work.
• Reconnect with the last event sequence after browser or network loss.
• Test Durable Object eviction during a model stream and during a tool write.
• Test waiting for user input and approval across a restart.
• Test credential rotation and revoked credentials during recovery.
• Reconcile outcome_unknown instead of repeating an external write.

Operations

• Record provider, model, billing mode, cache policy, token usage, tool calls, latency, retries, and trace IDs.
• Keep subscription cost unknown when the provider reports no price.
• Project the Flue stream to immutable history without creating a second transcript authority.
• Index Recall records with organization and project filters.
• Define retention and deletion for source history, R2 blobs, and search index records.
• Set quotas for runs, subagents, tools, tokens, duration, and storage.

Failure drills

Run these scenarios in a staging environment:

1. Close the browser during a long run. Reconnect and verify no duplicate events or tool calls.
2. Rotate a BYOK credential after admission. Verify recovery uses the admitted generation or returns a typed recovery failure.
3. Approve a write, stop the Worker, and verify the journal prevents a second write.
4. Use the same public run ID under two tenants. Verify both bindings remain isolated.
5. Delete a project and verify history, blobs, and derived Recall records are removed or tombstoned.