Codex Source #

Data source:

${CODEX_HOME:-~/.codex}/sessions/
${CODEX_HOME:-~/.codex}/archived_sessions/

When both directories contain the same relative JSONL path for one Codex home, the active sessions/ copy wins.

Relevant JSONL event:

• type === "event_msg"
• payload.type === "token_count"
• payload.info.total_token_usage is cumulative.
• payload.info.last_token_usage is the current turn delta.
• If only cumulative totals exist, subtract prior totals to recover deltas.

Relevant speed-setting event in Codex CLI 0.144.0 and later:

• type === "event_msg"
• payload.type === "thread_settings_applied"
• payload.thread_settings.service_tier === "priority" (or legacy "fast") selects Fast.
• payload.thread_settings.service_tier === "default" selects Standard. Codex Desktop spells the same tier "standard"; both appear in the same CLI version, so this is a value mapping and not a version split.
• Token usage inherits the latest recognized setting in the rollout. A settings event without a service_tier key leaves the previous tier in place (auto-review threads emit these); a tier that is present but unrecognized clears it so a stale value is not inherited.
• thread_settings_applied is not emitted per turn, so short rollouts carry no tier at all and stay unclassified for report policy to resolve.

Relevant MultiAgent V2 subagent replay markers:

• A subagent rollout replays its parent's history before the child's own turn, so the prefix must not be counted again.
• type === "event_msg" with payload.type === "task_started" ends the replayed prefix.
• type === "inter_agent_communication_metadata" with payload.trigger_turn === true starts the child turn. The event was named inter_agent_communication before Codex CLI 0.143.0-alpha.15; both spellings are matched. trigger_turn is also emitted as false, so the value must be checked.
• Usage between those two markers belongs to the child; usage before them is inherited history.
• Which replay boundary a rollout uses is decided by scanning it for the marker, never from session_meta. Resuming a session re-records the original session's cli_version, so that field cannot be used to gate behavior.

Token mapping:

• input_tokens - total input tokens.
• cached_input_tokens - cached prompt tokens.
• output_tokens - completion tokens, including reasoning cost.
• reasoning_output_tokens - informational breakdown; already included in output billing.
• total_tokens - provided directly or recomputed as input plus output for legacy entries.

Pricing uses model metadata from turn_context. Early sessions without metadata fall back to gpt-5, mark isFallbackModel === true, and expose fallback rows as approximate in aggregate JSON.