[Claude Code] Token usage always 0 — "transcript flush sentinel not found within timeout" on every TurnEnd

[heychad]

Summary

On Entire CLI 0.5.2 (5affe173) with Claude Code 2.1.100 (claude-opus-4-6[1m]), every TurnEnd logs a transcript flush sentinel not found within timeout, proceeding warning, the model field on the turn-end event is empty, and entire status reports tokens 0 even across many completed turns.

Environment

• Entire CLI: 0.5.2 (5affe173), Go 1.26.1, darwin/arm64
• Claude Code: 2.1.100
• Model: claude-opus-4-6[1m] (Opus 4.6, 1M context)
• macOS 15.x (Darwin 24.6.0)
• Mode: manual-commit on main

Repro

1. Start a Claude Code session in a repo with Entire enabled.
2. Submit several prompts, let each turn complete.
3. entire status → session shows tokens 0.
4. tail .entire/logs/entire.log → every turn logged:

"transcript flush sentinel not found within timeout, proceeding"
  component: agent.claudecode
  timeout:   3000000000   (3s)

And every corresponding turn-end event has model:"" (empty), even though the session was started with model:"claude-opus-4-6[1m]".

Expected

entire status should reflect actual token usage after each stop, and model should populate on turn-end.

Actual

tokens 0 forever. Empty model on every turn-end. Token parsing silently fails.

Guess at root cause

Either (a) the 3 s sentinel timeout is too tight for Opus on long-context turns (Claude Code is still flushing the JSONL transcript when Entire reads it), or (b) the claude-code adapter's sentinel detection / transcript parser doesn't match the current Claude Code 2.1.x transcript format — the empty model field on turn-end suggests the parser is running but failing to find expected fields, not just timing out.

Willing to help

Happy to grab a redacted transcript sample and/or run a patched build with more verbose sentinel logging if it would help.