Summary

Fixes #2571 — a subagent infinite tool-call loop burned $350+ in one afternoon due to two missing safeguards.

Changes

Bug 1: Subagent infinite loop — no circuit breaker

File: src/tools/delegate-task/sync-session-poller.ts

Added a maxAssistantTurns parameter (default: 300) to pollSyncSession. Each time a new assistant message is detected, a counter increments. When the limit is reached, the session is aborted and a clear error message is returned to the parent.

Task aborted: subagent exceeded 300 assistant turns without completing.
This usually indicates an infinite tool-call loop. Session ID: ...

The default of 300 is well above any normal task's requirements while still preventing runaway loops. The parameter is optional and backward-compatible.

Bug 2: Silent model routing — no user visibility

File: src/tools/delegate-task/sync-task.ts

When a task completes, the return string now includes the actual model used. If the subagent was routed to a different model than the parent session (via category resolution), a ⚠️ warning is shown:

⚠️  Model routing: parent used openai/gpt-5.4, this subagent used google/gemini-3.1-pro-preview (via category: visual-engineering)

Root Cause (from issue)

A Sisyphus-Junior subagent was silently routed to google/gemini-3.1-pro-preview via the visual-engineering category, despite the parent session using GPT-5.4. The subagent then entered an infinite tool-calls loop (809 turns, 3.5 hours) with no safeguard to stop it. Gemini 3.1 Pro has zero prompt caching, so each of the 809 turns sent 880K input tokens at full price ($1.76/turn displayed, ~$3.52/turn actual due to >200K context tier).

Testing

bun test src/tools/delegate-task/
# 199 pass, 0 fail

All existing delegate-task tests pass. No breaking changes — maxAssistantTurns is an optional parameter with a safe default.

Related

• Closes [Bug]: Subagent burned $350 in 3.5 hours with ZERO safeguard — silent model routing + no loop limit + cost display showing HALF the real bill #2571
• Related: [Feature]: Improve repeating text loop #2527 (repeating text loop — similar symptom)
• Related: [Bug]: Category variant dropped for categories without CATEGORY_MODEL_REQUIREMENTS entry #2538 (category variant routing bugs)

Summary by cubic

Adds a turn-limit circuit breaker for subagents and makes model routing visible in task results to prevent infinite loops and surprise model costs. Addresses #2571.

• Bug Fixes

  • Added maxAssistantTurns (default 300) to subagent polling; aborts the session and returns a clear error when the limit is hit.
  • Task output now includes the actual model used; shows a warning if it differs from the parent session’s model, with the routing category.

• Migration

  • Backward compatible. No changes required; maxAssistantTurns is optional and defaults to 300.

^{Written for commit f2b26e5. Summary will update on new commits.}