[Bug] Subagents inherit session model without cost confirmation, causing unintended quota drain on limited-quota models
Bug Description
Fully attributed now — zero ambiguity. The 2,448 Fable subagent messages are
in 16 files under this session's subagents/ directory, and those 16 agent IDs
are exactly the deep-audit auditors I launched. The entire audit ran on Fable,
then this session switched to Opus-4.8 for the doctor/this investigation. So
your instinct was correct: a brief Fable window, but it contained a 16-way
fan-out I spawned.
Here's the complete, paste-ready package.
---
Bug report evidence — Fable 5 usage spike
Environment
- Claude Code 2.1.223 (native install, macOS, EDT/-0400)
- Limit reset: Tue Aug 4 2026, 4:00 PM EDT (2026-08-04T20:00:00Z)
- Report generated from local transcripts in ~/.claude/projects/
The event
- A 16-agent audit workflow launched from one session (fb9157a6-…) while the
session model was claude-fable-5.
- Subagents inherited the session model, so all 16 ran as Fable, in parallel.
- Burst window: 2026-08-06T19:41:55Z → 20:00:19Z = 18 min 24 sec (3:41–4:00 PM
EDT)
- Attribution: 16 transcript files under fb9157a6-…/subagents/agent-*.jsonl,
2,448 assistant messages total.
Post-reset usage by model (since 2026-08-04T20:00:00Z)
┌──────────┬──────────────────┬───────┬────────────┬──────────────┬────────┐
│ Model │ Kind │ Msgs │ cache_read │ cache_create │ output │
├──────────┼──────────────────┼───────┼────────────┼──────────────┼────────┤
│ fable-5 │ subagent (the │ 2,448 │ 724.0M │ 29.7M │ 0.59M │
│ │ burst) │ │ │ │ │
├──────────┼──────────────────┼───────┼────────────┼──────────────┼────────┤
│ fable-5 │ main (your │ 65 │ 4.1M │ 0.5M │ 0.49M │
│ │ typing) │ │ │ │ │
├──────────┼──────────────────┼───────┼────────────┼──────────────┼────────┤
│ opus-5 │ main (your real │ 6,566 │ 2,430.9M │ 26.6M │ 6.93M │
│ │ 2-day workload) │ │ │ │ │
├──────────┼──────────────────┼───────┼────────────┼──────────────┼────────┤
│ opus-5 │ subagent …
Note: Content was truncated.
This issue has 1 comment on GitHub. Read the full discussion on GitHub ↗