Steering ignored: messages typed during long agent tool-call sequence never reach the model

Resolved 💬 2 comments Opened May 21, 2026 by lukesoftware Closed May 25, 2026

Summary

In Claude Code (VS Code extension, Opus 4.7 1M context), while the assistant is running a long sequence of tool calls (e.g. a multi-file batch with successive Edit/Write/Bash calls), any message the user types into the input and submits with Enter never reaches the model. The message visually leaves the input box and shows up in the chat history, but the next assistant turn shows no awareness of it: the assistant just keeps marching through its plan.

This effectively disables steering during the most useful moment to steer (long autonomous executions).

Repro

  1. Ask the assistant to perform a 5+ step refactor across multiple files in one turn.
  2. While the assistant is mid-execution (lots of consecutive Edit/Bash/Write calls), type two or three short messages and submit each with Enter (e.g. "wait, stop", "change of plan, do X instead").
  3. The messages appear in the conversation transcript but the assistant never references them, never pauses, never adjusts.

Workaround currently used

The user has to either:

  • Close the chat / re-open it (force a fresh turn boundary), or
  • Wait until the assistant finishes the entire batch before speaking.

Both are bad: the first loses minor context state, the second defeats steering entirely.

Expected

Either:

  • Hard interrupt — typing during execution cancels the in-flight stream (like API streaming.cancel()), and the user message starts a new turn. (Preferred.)
  • Soft queue — the user message is queued and surfaced as the first thing in the models next turn (with a visible "queued" indicator on the input), so the model sees it before continuing.

Either is fine. Today is neither: the message seems to vanish from the models perspective.

Environment

  • OS: Windows 11 Pro 26200
  • IDE: VS Code with Claude Code extension
  • Model: claude-opus-4-7 (1M context)
  • Permission mode: autorun on a dev sandbox (most tools pre-approved, so no permission prompts that could act as natural pause points)
  • Reproducible across multiple sessions on 2026-05-21

View original on GitHub ↗

This issue has 2 comments on GitHub. Read the full discussion on GitHub ↗