CHECKLIST.md

Pre-Deployment Checklist

Model configuration

• Model IDs verified against provider -- exact strings, not approximations
• Fast tier and think tier both tested with a real message
• Model names in system prompt match actual TIERS config
• Reasoning mode disabled on fast tier if using a hybrid model

Formatting enforcement

• Formatting directive present in system prompt
• Formatting directive injected into every user message before dispatch
• Tested with a prompt that would naturally produce markdown -- confirmed plain text output

Truncation detection

• finish_reason checked on every response
• Truncation notice appended when finish_reason is "length"
• max_tokens set appropriately for each tier -- not too low, not unbounded

Token management

• Token logging active -- input and output tracked per call
• Daily spend limit configured with warning threshold
• Cost rates accurate for current models

Process management

• systemd service file in place with Restart=on-failure
• WatchdogSec configured -- detects hung processes, not just crashes
• Port conflict procedure documented -- stop service, kill orphans, confirm port clear, start clean
• Service tested: stop, start, restart all behave correctly

Secrets

• No credentials in source code or environment variables
• Secrets encrypted at rest -- SOPS+age or equivalent
• Secrets decrypted at runtime only, not stored as permanent env vars
• Secrets file excluded from version control

Comms

• Inbound message routing tested end to end
• Outbound message delivery confirmed
• Error states surface to operator -- agent does not fail silently