Codex usage telemetry gotchas: keep explicit zeros, as OpenBench adapter nulls cache_write_input_tokens
Due_Emu_8229 · reddit · 2026-09-05
The author built Astra Sweetspot (MIT-licensed, publishing coding experiments with patches and reproducible runs) with Codex. Auditing the first eight run receipts exposed a telemetry mistake in his own collector: cachewriteinputtokens: 0 and reasoningoutputtokens were present in all eight original logs but discarded; v0.1.1 restores them and publishes usage-only events with hashes — no model reruns, timing and grades unchanged.
Replaying the event also revealed OpenBench's legacy Codex adapter doesn't recognize cachewriteinputtokens, turning an explicit zero into null and downgrading the usage basis to "estimated". A fix with an offline regression is pending (PR #50).
Practical lesson: collectors should preserve an explicit zero separately from an absent field, and check the emitted event before declaring a metric unavailable. Reasoning is included in output tokens, cached input in input, and these counters don't attribute subscription quota. All eight corrected receipts are inspectable via npx [email protected] report --.
More from coding & agent
- Hugging Face details Moon Bot, its Slack-native coding agent with codebase and DB access — victormustar · 2026-09-05
- Computer use is now where coding agents were in Nov 2025, says agentic tooling author — austinvhuang · 2026-09-05
- Dev boosts GLM 5.2 TPS on a B300 and swaps it into Claude Code in place of Anthropic models — abhijithneil · 2026-09-05
- What forces LLM teams to optimize inference when going from MVP to production? — Ok_Philosophy_4031 · 2026-09-05
- Astra's trick: map the GUI once, then drive computer use with programmatic tool calls — anpaure · 2026-09-05
- Local Qwen 27B vibecodes a playable Godot dungeon game in just 4 prompts — jacek2023 · 2026-09-05