Codex
Send OpenAI Codex's call events to Provantix and its model calls and tool calls appear as records, tied to the person using it, with no code to write. This page says exactly what that carries and what it does not.
Status: built from OpenAI's published configuration documentation for Codex's [otel] settings (read 2026-11-15) and from third-party descriptions of its event attributes. OpenAI's pages describe the events in prose and do not list every attribute name. It has not been run from a live Codex session, so check the first delivery against Connectors, and expect to adjust the attribute names if a live session differs.
What Codex exports
Codex exports log events and metrics (and, on a separate setting, traces). Telemetry is off by default. Provantix takes the log events only: a metric is a total with no call behind it, and Codex's trace spans are not read.
Set it up
- On Setup, under Connect a gateway or OTLP source, create a token.
- In your user-level
~/.codex/config.toml(a project-level.codex/config.tomlignoresotel):
[otel]
environment = "dev"
log_user_prompt = false # leave it false: Provantix would drop prompts anyway
[otel.exporter.otlp-http]
endpoint = "https://app.provantix.tech/v1/logs"
protocol = "json" # Provantix takes the JSON encoding, not binary
headers = { Authorization = "Bearer pvx_otlp_..." }
Leave the trace and metrics exporters unset or none.
What becomes a record
| Codex event | Record |
|---|---|
codex.sse_event with token counts (the response completing) |
A model call: input, output, cached and reasoning token counts, how long it took. |
codex.tool_result |
A tool call: the tool name, whether it succeeded, how long it took. |
The model. Codex reports the token counts on one event and the model on others (codex.conversation_starts, codex.api_request, codex.websocket_request). Provantix remembers the model each conversation named and uses it for the token events that follow, in the same delivery or a later one. The memory is in the running service and is lost on a restart. A token event whose conversation has named no model is recorded with no model and is not counted in the cost views. Say so if you see that.
Tokens. OpenAI's counts include the cached part, so the cached and reasoning counts are parts of input and output. They are priced at their own rates when you have set them. See Token kinds.
Cost. Codex reports no cost, so none is stored. The estimate is from your own price table: add a price row for each model (Settings, Cost tab, Model prices).
The person. The event's user.email, else user.account_id, is the person. Only a handle (a one-way code) is kept, never the address, so the people directory can place the call in a department.
Tags. Set provantix.cost.department and the other documented names where your Codex version puts attributes on its events. OpenAI's documentation describes [otel.span_attributes] for spans; whether those reach events is not stated, so check the first delivery. Until then Codex calls are unallocated for a tag unless an agent tag or a rule says otherwise. See Cost tags.
What is not captured
- No content. Prompts, agent responses, tool output and code are not read, whether or not
log_user_promptor the response settings are on. Only the numbers and names above are taken. - Metrics, traces, session starts, prompts and permission decisions are not taken.
- No destination check. A tool call is the assistant's own account of what it did.
- Who the agent is is the account on the event, or the default agent you set. It is not a verified identity.
If nothing appears
- Connectors shows nothing received from this source yet for a token that has never delivered.
- A
wrong_signalerror means metrics are pointed at this endpoint. Apayload is not valid JSONerror usually means the protocol isbinary: set it tojson. - Records with no model: the conversation named none before its token events. See The model, above.