[Bug] Model ignores conditional logic in gated instructions, spawns unnecessary subagents

Status Open
Reported on v2.1.218
Maintainer reply None cached
Activity 0 comments · opened Jul 23, 2026

Bug Description
Feedback: model ignored an explicit conditional instruction

What I asked: Check whether a runbook/manual already exists for a task. My instruction was explicit and conditional: only if no manual exists, create one or give me instructions — and in that case use cheap Sonnet subagents to research online.

What the model did: It correctly found that the manual already existed and was current (5 days old, verified against official docs). But it still launched a Sonnet web-search subagent to "confirm console screens" — spending tokens on the exact branch of work that my instruction had gated behind "if no manual exists." The condition was false, so that subagent should never have run.

Why this is a problem:

  • It burned tokens/cost on work I explicitly conditioned away.
  • It shows the model parsed the topic of the instruction ("research Twilio console with subagents") but dropped the condition attached to it ("only if there's no manual"). The condition is the whole point of the instruction, not a detail.
  • When I pushed back, it acknowledged the error only after I pointed it out twice — it didn't self-catch the contradiction between "the manual already exists and is current" and "let me launch the research agent anyway."

What I expected: Once it confirmed the manual existed, it should have stopped there and reported, without spawning any research subagent. If it thought the existing manual had gaps worth verifying, it should have asked me first ratherthan assuming and spending resources.

Pattern to fix: The model treats conditional instructions ("only if X, do Y") as if they were unconditional ("do Y").It needs to evaluate the condition and let a false condition actually suppress the gated action — especially when thegated action has a real cost (spawning subagents, spending tokens).

Environment Info

  • Platform: linux
  • Terminal: WezTerm
  • Version: 2.1.218
  • Feedback ID: 7278a263-0512-4846-8902-61c075ed9ad7

Errors

[]

View original on GitHub ↗