[FEATURE] Pause-and-steer for extended thinking — and formalizing delegation as steerable reasoning
Preflight Checklist
- [x] I have searched existing requests and this feature hasn't been requested yet
- [x] This is a single feature request (not multiple features)
Problem Statement
I steer Claude primarily by reading its visible extended thinking. When a wrong assumption enters the reasoning chain early, my only tool today is interrupting the turn — which disrupts in-flight state and confuses the flow.
Correcting after the final answer is far more expensive: by then the model has built four or five reasoning steps on the bad premise, so unwinding costs much more time and many more tokens than an early nudge would have.
I hit this daily. I am a pharmacist, not an engineer, and I have shipped production work across eight domains with Claude since January 2026 — an app on both stores, a full construction execution package, trading indicators, invoice automation. In all of them, the single biggest source of wasted work is a wrong premise I could see forming but could not correct in place.
Proposed Solution
Two parts. The second is the one I care about most.
A) Checkpointed thinking
Let the user inject a correction during extended thinking and then resume, instead of killing the turn. A lighter version would already help: the model surfaces its working assumptions at natural checkpoints and briefly waits for optional steering before building on them.
B) Formalize delegation as steerable reasoning — this already works today
Pausable thinking effectively EXISTS through delegation. When heavy reasoning is fanned out to background subagents, the orchestrator stays free and conversational: my input lands instantly, and agents can be redirected, killed, or respawned mid-flight. The reasoning becomes a steerable system instead of one uninterruptible breath.
I found this by accident while running multi-agent workflows, and it changed how I work. It deserves to be a first-class interaction mode — a "think with me" mode where heavy reasoning always externalizes into agents while the main loop stays open to the human.
Why it matters: prevents error compounding, saves tokens and wall-clock time, and turns reasoning from a monologue into a collaboration.
Alternative Solutions
_No response_
Priority
Medium - Would be very helpful
Feature Category
Interactive mode (TUI)
Use Case Example
_No response_
Additional Context
_No response_