codex adapter: total_tokens undercounts by omitting cached_input_tokens and reasoning_output_tokens

[fjgbue]

⚠️ This was automatically generated by Javi OpenClaw BugFinder Bot. Not reviewed by a human. Verify independently. Author: 12122J. Co-author: fjbgu OpenClaw Assistant.

Severity: Medium

File & Line

src/adapters/codex.mjs — observeCodexLine() function, turn.completed event handler (approx. line 68)

Root Cause

When processing turn.completed events from Codex CLI, the total_tokens field is computed as:

total_tokens: numberOrZero(usage.input_tokens) + numberOrZero(usage.output_tokens)

This omits two token categories that Codex may report:

1. cached_input_tokens — tokens retrieved from the context cache, which still count toward the session total
2. reasoning_output_tokens — tokens consumed by extended thinking / reasoning models (e.g., models with reasoning/chain-of-thought capabilities)

Compare with the Claude Code adapter in claude-code.mjs, which includes both:

const total = input + cachedRead + cachedCreate + output;

Impact

Token usage is systematically under-reported for Codex sessions. For reasoning-heavy models, the undercount can be substantial (reasoning tokens often exceed output tokens). This affects:

• Cost estimates (pricing calculations use total_tokens)
• Session summaries (tt summarize)
• Dashboard visualizations

Suggested Fix

if (payload.type === "turn.completed" && usage) {
  observations.push({
    type: "usage.tokens",
    input_tokens: numberOrZero(usage.input_tokens),
    cached_input_tokens: numberOrZero(usage.cached_input_tokens),
    output_tokens: numberOrZero(usage.output_tokens),
    reasoning_output_tokens: numberOrZero(usage.reasoning_output_tokens),
    total_tokens: numberOrZero(usage.input_tokens)
      + numberOrZero(usage.cached_input_tokens)
      + numberOrZero(usage.output_tokens)
      + numberOrZero(usage.reasoning_output_tokens),
  });
}

Additional Context

The usage object destructured from Codex events may use different field names than the DeepSeek/Anthropic convention. Verify that Codex uses cached_input_tokens and reasoning_output_tokens as field names in the actual event payload. If Codex uses different names, add those variants to the destructuring.

Also consider making the total_tokens computation consistent with the Claude Code adapter (claude-code.mjs), which already accounts for all token categories.