Skip to content
Confabulous

Codex

Confabulous has first-class support for OpenAI Codex sessions, including subagent spawns and skill invocations.

What gets parsed

Section titled “What gets parsed”
  • Full conversation history.
  • Per-message token counts (input, output, cached input, reasoning).
  • Model identifier (gpt-5, gpt-5.5, o3, etc.).
  • Tool calls.
  • Subagent spawns (spawn_agent / wait_agent) — bucketed by agent role.
  • Skill invocations (<skill> user-message wrappers) — bucketed by skill name.
  • Parent-child thread relationships (recursive tree of spawned subagents).

Analytics cards

Section titled “Analytics cards”
  • Tokens — including reasoning tokens (preserved for display; billed at output rate).
  • Cost — using the OpenAI pricing table.
  • Tools, Agents & Skills — Codex-specific breakdown.
  • Conversation — Codex synthesizes reasoning time into active time.
  • Code activity — files modified and lines added/removed, from whichever edit format your Codex CLI version emits.
  • Repo activity.

Codex CLI version differences

Section titled “Codex CLI version differences”

Codex reshaped its rollout format in version 0.149.1. Confabulous reads both formats, so old and new sessions analyze correctly, but two things legitimately differ between them.

Tool names are reported as Codex recorded them. Sessions from 0.130.0 and earlier show exec_command, apply_patch and write_stdin; sessions from 0.149.1 onward show exec, wait and web.search. Codex renamed its tools, so a session from before the change and one from after will list different tool names for the same kind of work. Confabulous does not remap them onto a common vocabulary — that would mean displaying names that never appeared in your transcript.

Files read and searches are only available from 0.149.1 onward. Newer Codex versions record what each shell command was doing, so Confabulous can count file reads and searches. Older versions recorded only the command line itself, with no indication of its purpose, so those two figures stay at zero for older sessions rather than being guessed at from command text.

Subagent aggregation

Section titled “Subagent aggregation”

When a Codex session spawns subagents, Confabulous aggregates the main thread plus every subagent thread for most analytics cards. The Conversation card stays main-only by design.

Pricing nuances

Section titled “Pricing nuances”
  • cached_input_tokens is a subset of input_tokens (not a separate count).
  • reasoning_output_tokens is a subset of output_tokens (billed at output rate).
  • OpenAI does not charge for cache writes.

Other supported providers

Section titled “Other supported providers”

Confabulous treats every provider as a first-class citizen. Claude Code, Cursor, and OpenCode are also supported today. New providers slot into the same sync, storage, and analytics pipeline.