Export Codex Activity

The Logfire Exporter plugin sends Codex’s own activity telemetry to Logfire. It is separate from the main Logfire plugin, which gives Codex skills and MCP access for instrumenting applications and querying existing telemetry.

Plugin	Purpose
Logfire	Helps Codex instrument your code, query Logfire data, and open Logfire UI views.
Logfire Exporter	Exports Codex turns and tool calls to Logfire.

The two plugins solve different problems and can be installed together.

Install

Add the Pydantic marketplace and enable the exporter plugin:

Terminal

codex plugin marketplace add pydantic/skills --ref main
codex plugin add logfire-exporter@pydantic-skills

You can also enable Logfire Exporter from the Pydantic marketplace in the Codex plugin UI.

After enabling the plugin:

Restart Codex so hook configuration is loaded.
Run /hooks if Codex asks you to review or trust the plugin hooks.
Complete a new Codex turn — the exporter sends telemetry when the turn finishes.

Configure

Create a config file at:

${XDG_CONFIG_HOME:-~/.config}/logfire-exporter/config.env

Example for Logfire Cloud:

LOGFIRE_TOKEN=<your Logfire write token>
LOGFIRE_BASE_URL=https://logfire-api.pydantic.dev

For a local Logfire instance:

LOGFIRE_TOKEN=test-e2e-write-token
LOGFIRE_BASE_URL=http://localhost:3000
CODEX_LOGFIRE_DEBUG=true

LOGFIRE_URL is still accepted as a compatibility alias, but LOGFIRE_BASE_URL takes precedence.

The exporter sends Authorization: <LOGFIRE_TOKEN>, matching Logfire’s direct OTLP client configuration. If you need a scheme-prefixed header, set CODEX_LOGFIRE_AUTH_SCHEME=Bearer.

How It Works

The exporter registers Codex lifecycle hooks:

Hook	Purpose
`SessionStart`	Initialize local telemetry state
`UserPromptSubmit`	Store the submitted prompt and turn metadata
`PostToolUse`	Store tool-call results for export as child spans
`Stop`	Export the completed turn and its tool spans to Logfire

Export boundary

Only completed turns are exported. The exporter sends telemetry on the Stop hook — it does not live-stream every Codex action. Interrupted turns that never reach Stop are not exported.

Trace model

There is no long-lived session root span. Codex hooks do not provide a reliable session-end event, so the exporter deliberately avoids a session-level trace.

Instead, each completed turn becomes a root-level span in a deterministic conversation trace. Tool calls appear as child spans of that turn span. Spans for the same Codex conversation share a stable trace ID derived from Codex conversation metadata.

Content Capture

Content capture controls whether prompt text, assistant text, tool inputs, tool outputs, and tool errors are included in exported span attributes. By default, the exporter uses CODEX_LOGFIRE_CONTENT_CAPTURE_MODE=full, which includes redacted content for all of those fields.

For sensitive projects, set a narrower mode in config.env:

CODEX_LOGFIRE_CONTENT_CAPTURE_MODE=metadata_only

Available modes:

Mode	Captures
`full`	Redacted user prompt, final assistant message, tool input/output, and tool errors (default).
`no_tool_content`	Redacted user prompt and final assistant message, but not tool input/output.
`metadata_only`	No prompt, assistant, tool input, or tool output content.

When content capture is enabled, spans include pydantic_ai.all_messages, the Pydantic AI message-schema attribute that lets Logfire show Codex turns in the generic LLM conversation details panel.

Troubleshooting

Debug logs are written to:

${XDG_STATE_HOME:-~/.local/state}/logfire-exporter/logs/

If no spans appear after a completed Codex turn, check the Codex TUI log: