transcript adapter: reasoning_output_tokens always zero — extended thinking models invisible

[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/transcript.mjs — extractFromTranscript() function, return statement (approx. line 120)

Root Cause

The usage object returned by extractFromTranscript() hardcodes reasoning_output_tokens to 0:

usage: hasUsage ? {
  input_tokens: rawUsage.input_tokens,
  cache_creation_tokens: rawUsage.cache_creation_input_tokens,
  cache_read_tokens: rawUsage.cache_read_input_tokens,
  output_tokens: rawUsage.output_tokens,
  reasoning_output_tokens: 0,  // <-- ALWAYS ZERO, never extracted
  total_tokens: total
} : null,

The function parses Claude Code transcript JSON lines and aggregates input_tokens, cache_creation_input_tokens, cache_read_input_tokens, and output_tokens from assistant messages — but never reads reasoning_output_tokens (also known as thinking_tokens or reasoning_tokens in the API response).

Claude Code with extended thinking models (Opus 4 Thinking, etc.) DOES include this field in the transcript. For example, an assistant message from a thinking model contains:

{
  "type": "assistant",
  "message": {
    "usage": {
      "input_tokens": 1234,
      "cache_creation_input_tokens": 0,
      "cache_read_input_tokens": 500,
      "output_tokens": 800,
      "reasoning_output_tokens": 12000
    }
  }
}

The 12,000 reasoning tokens are silently dropped.

Impact

• Hook-recorded sessions (automatic recording via Claude Code Stop hook) show zero reasoning tokens even when the session used an extended thinking model.
• Cost estimates are wrong — reasoning tokens are billed (though often at a lower rate), making the recorded cost lower than actual.
• Dashboard and summaries show incomplete token breakdowns for thinking-enabled sessions.

Suggested Fix

In the assistant message processing loop, extract reasoning_output_tokens from the usage block:

if (entry.type === "assistant") {
  const msg = entry.message ?? {};
  const u = msg.usage ?? {};

  if (msg.model && !model && !msg.model.startsWith("<")) model = msg.model;

  if (u.input_tokens != null || u.output_tokens != null) {
    hasUsage = true;
    rawUsage.input_tokens += u.input_tokens ?? 0;
    rawUsage.cache_creation_input_tokens += u.cache_creation_input_tokens ?? 0;
    rawUsage.cache_read_input_tokens += u.cache_read_input_tokens ?? 0;
    rawUsage.output_tokens += u.output_tokens ?? 0;
    // FIX: capture reasoning tokens
    rawUsage.reasoning_output_tokens += u.reasoning_output_tokens ?? 0;
    turns.push({
      input: u.input_tokens ?? 0,
      cacheWrite: u.cache_creation_input_tokens ?? 0,
      cacheRead: u.cache_read_input_tokens ?? 0,
      output: u.output_tokens ?? 0,
      reasoning: u.reasoning_output_tokens ?? 0,  // add to turns too
    });
  }
  // ...
}

Then in the return statement, use the tracked value instead of the hardcoded 0:

usage: hasUsage ? {
  input_tokens: rawUsage.input_tokens,
  cache_creation_tokens: rawUsage.cache_creation_input_tokens,
  cache_read_tokens: rawUsage.cache_read_input_tokens,
  output_tokens: rawUsage.output_tokens,
  reasoning_output_tokens: rawUsage.reasoning_output_tokens,
  total_tokens: total
} : null,

The total_tokens computation on the line above may also need updating to include reasoning tokens (currently it is input_tokens + cache_creation_input_tokens + output_tokens), depending on whether reasoning tokens are meant to be counted as new work tokens.