AI agent usage logs

One row per LLM call a Chatwoot/iframe/specialist agent turn made — README_AIAGENTUSAGE_TOKENS.md. `bots/engine.py::run_bot_turn` already computes `usage` (token buckets) and `cost_usd` (real cost under the provider's pricing) for every turn, but until this model existed that number only ever reached a `logger.info` line ("bots.cost: ...") and was discarded — there was no way to answer "how much did agent X spend last week" without grepping raw server logs with no `agent_key` in the line. Written best-effort right after each `run_bot_turn` call (see `bots/usage_tracking.py::record`) — a failure saving this row must never break the reply the customer is waiting for.

Field	Description
AI agent	Null only for a turn where bot_config carried no agent_id — should not happen on the DB-backed path, kept nullable defensively so a bad turn never loses the whole row.
Company
Conversation	Channel-qualified conversation id, e.g. chatwoot:536 — empty for synthetic test turns with no real conversation behind them.
Channel	Options: Chatwoot Iframe (test channel) Synthetic test (handle_test_message) Internal specialist call (orchestrator)
LLM model
LLM calls in this turn	A turn may loop through several API calls (tool-use round-trips) — this is the count actually accumulated, not always 1.
Input tokens
Output tokens
Cache read tokens
Cache creation tokens
Cost (USD)	Real cost of this turn under the provider's per-token pricing (bots/engine.py::usage_cost_usd) — not a raw token count, an output token costs several times an input one.

AI agent usage logs It has the following related modules and may be of interest to you: