Repository navigation
codex: re-emitted token_count events are double-counted (usage overstated ~30-67% on real data) #1288
Description
Summary
The codex adapter sums last_token_usage per token_count event, deduping on a key that includes the millisecond timestamp. But Codex CLI re-emits token_count events with an unchanged total_token_usage (UI/rate-limit refreshes, not new API calls), and each re-emit has a fresh timestamp — so every re-emit is counted as new usage. Forked/subagent sessions amplify this: at spawn they replay the parent thread's history as a sub-second burst of token_count events that are pure duplicates of values already counted in the parent file.
On my real ~/.codex/sessions (229 files, ~141k token_count events, 5,357 of 9,157 events in the largest file are pure re-emits): ccusage reports 8.58B codex tokens; counting deltas of the cumulative total_token_usage per session (replay baselines removed) gives 5.15B — a ~67% overstatement. Verified in v20.0.9 and v20.0.11.
Repro fixture
One session, two real turns (cumulative ends at 3,300 tokens), with two re-emits of the first turn's counter — the kind Codex writes constantly:
{"timestamp":"2026-05-01T10:00:00.000Z","type":"session_meta","payload":{"id":"01970000-0000-7000-8000-000000000001","timestamp":"2026-05-01T10:00:00.000Z","cwd":"/tmp","originator":"codex_cli_rs","cli_version":"0.73.0","source":"cli","model_provider":"openai"}}
{"timestamp":"2026-05-01T10:00:01.000Z","type":"turn_context","payload":{"model":"gpt-5"}}
{"timestamp":"2026-05-01T10:00:02.000Z","type":"event_msg","payload":{"type":"token_count","info":{"total_token_usage":{"input_tokens":1000,"cached_input_tokens":800,"output_tokens":100,"reasoning_output_tokens":50,"total_tokens":1100},"last_token_usage":{"input_tokens":1000,"cached_input_tokens":800,"output_tokens":100,"reasoning_output_tokens":50,"total_tokens":1100},"model_context_window":258400},"rate_limits":{}}}
{"timestamp":"2026-05-01T10:00:02.500Z","type":"event_msg","payload":{"type":"token_count","info":{"total_token_usage":{"input_tokens":1000,"cached_input_tokens":800,"output_tokens":100,"reasoning_output_tokens":50,"total_tokens":1100},"last_token_usage":{"input_tokens":1000,"cached_input_tokens":800,"output_tokens":100,"reasoning_output_tokens":50,"total_tokens":1100},"model_context_window":258400},"rate_limits":{}}}
{"timestamp":"2026-05-01T10:00:03.000Z","type":"event_msg","payload":{"type":"token_count","info":{"total_token_usage":{"input_tokens":1000,"cached_input_tokens":800,"output_tokens":100,"reasoning_output_tokens":50,"total_tokens":1100},"last_token_usage":{"input_tokens":1000,"cached_input_tokens":800,"output_tokens":100,"reasoning_output_tokens":50,"total_tokens":1100},"model_context_window":258400},"rate_limits":{}}}
{"timestamp":"2026-05-01T10:00:10.000Z","type":"event_msg","payload":{"type":"token_count","info":{"total_token_usage":{"input_tokens":3000,"cached_input_tokens":2300,"output_tokens":300,"reasoning_output_tokens":130,"total_tokens":3300},"last_token_usage":{"input_tokens":2000,"cached_input_tokens":1500,"output_tokens":200,"reasoning_output_tokens":80,"total_tokens":2200},"model_context_window":258400},"rate_limits":{}}}Save as $CODEX_HOME/sessions/2026/05/01/rollout-2026-05-01T10-00-00-01970000-0000-7000-8000-000000000001.jsonl, then:
CODEX_HOME=... ccusage codex monthly --json --offline
Expected: totalTokens = 3,300 (the final cumulative).
Actual (v20.0.9–v20.0.11): totalTokens = 5,500 (the two re-emits of the 1,100 counter are added again).
Suggested fix
Count per-session deltas of total_token_usage (only when the cumulative moves), or dedupe last_token_usage events on the usage payload alone (excluding timestamp) and skip events whose cumulative is unchanged. For forked sessions, the replayed burst at spawn re-states cumulative values already present in the parent file and needs a baseline subtraction.
Thanks for ccusage — happy to test a fix against my real data.
Activity
pullfrog commentedon Jun 11, 2026
github-actions commentedon Jun 11, 2026
This issue was auto-closed. Issues from new contributors are auto-closed by default.
Maintainers review auto-closed issues and reopen worthwhile ones. Issues that do not meet the quality bar in CONTRIBUTING.md may not be reopened or receive a reply.
Keep the issue short, concrete, and written in your own voice.
If a maintainer replies
lgtmi, your future issues will stay open. If a maintainer replieslgtm, your future issues and PRs will stay open.See CONTRIBUTING.md.
fix(codex): dedupe replayed usage (#1435)
ryoppippi commentedon Aug 31, 2026
Historical audit: this discussion was auto-closed by the legacy contributor gate. That closure did not assess technical importance.
Audit result: resolved. A later merged change or the current
mainimplementation covers this request. This item is kept for history and does not need to be reopened.