Skip to content
Trazum

2.7.7

Released 2026-11-05

Fixed#

  • A Claude Code call with no requestId was billed once per line. Claude Code writes a response as one line per content block, and the reader collapsed those lines by requestId. A line without one was kept as a call of its own. On this repository's own transcripts every Claude Fable 5.1 line came without a requestId: 289 lines for 124 responses, billed at $47.82 where the calls cost $18.46. The lines are now keyed by the response's message.id when there is no requestId; where both are present they pair one to one, measured on 5,774 calls. The $29.36 difference is also what the "same request, sent again" finding had flagged on that model: it was this. noRequestId now counts only the lines that carry neither id.
  • A session's subagents were read as turns of the main conversation. Claude Code gives a subagent its session's sessionId, and the subagent keeps a context and a cache prefix of its own. The findings that follow a conversation turn by turn grouped by session alone, so the main thread and its subagents were one conversation to them. The growth ceiling took its floor from whichever thread had the smallest prompt: on this repository's own transcripts, measured on 2026-11-05, the main thread's smallest Opus 5.5 prompt is 110,262 tokens and a subagent's 58,190, and the ceiling came out at $412.48 where the threads support $366.74. A subagent's call seconds after a truncated main-thread answer read as a retry of it, and repeats, the cache TTL fit and the cache writes never read back paired calls across threads the same way. A call now carries a thread, which bill and from-claude-code take from a sidechain's agentId, and those findings group by session and thread. What a session costs, and the per-session budget, still count the whole session, subagents included.

Changed#

  • The growth sentence says when its figures are averages. Across several conversations the smallest and largest turn are means over conversations, and the sentence read as if they were one conversation's extremes. It now says "across N conversations, the smallest turn averages X tokens and the largest Y", and keeps the old wording for a single conversation.
  • What a session costs is said per session. The median, p95 and maximum per session, and the count beside --max-session-usd, said "conversations" for what was always a whole session. With a subagent now a conversation of its own, the same log could read "1 conversation" in one line and "across 31 conversations" in the next. Those sentences now say "sessions", and that subagents are counted in them, on the CLI, the web Bill tab and the MCP report.