Feature request: Codex-style live steering (inject input into the in-flight turn)

[riskeviteHR]

What I'd like

A Codex-style "live steering" capability: the ability to send a message while Claude is mid-turn and have it injected into the current reasoning/tool loop to nudge or redirect — without a hard interrupt and without losing the agent's place.

How it works today

Claude Code already supports:

• Queue while it works — a message typed during a run is delivered at the next step boundary.
• Esc — interrupt the current action, then type a new direction.
• Esc Esc — rewind and edit an earlier message.

These cover redirection well, but a queued message is only picked up at the next step boundary, so it feels less immediate than OpenAI Codex CLI, where input continuously steers the in-flight turn.

Request

Optionally inject queued user input into the current turn as soon as it arrives (between tool calls / mid-stream) so guidance takes effect immediately — matching the Codex steering UX. Ideally a configurable choice between live-steer and the current queue-at-boundary behavior.

Environment

• Claude Code CLI on Windows 11
• Model: Claude Opus 4.x

[github-actions]

Found 3 possible duplicate issues:

1. [Feature Request] Real-time steering: priority message channel for redirecting Claude mid-execution #30492
2. Feature: Real-time steering — send message mid-generation without queueing #64624
3. [FEATURE] Inject instructions into running task without interrupting (mid-execution context injection) #59684

This issue will be automatically closed as a duplicate in 3 days.

• If your issue is a duplicate, please close it and 👍 the existing issue instead
• To prevent auto-closure, add a comment or 👎 this comment

🤖 Generated with Claude Code

[cloverink]

+1. A concrete use-case the original captures well but worth making explicit: adding context mid-turn without redirecting.

On a long agentic run (editing, testing, committing for several minutes) I often want to drop in a fact, not steer — e.g. "FYI, the panel has no unit tests, verify by screenshot" or "FYI, also update the ADR while you're in there."

Both current options are wrong for that: queue lands too late (after the agent already made the decision the FYI would've informed), and Esc aborts and loses its place.

UX idea: a /fyi <message> prefix that's explicitly additive context, never an interrupt — "consider this now, keep going," vs a queued message's "do this after you finish." Under the hood it's just the live-steer path from this request, with intent made explicit.

[github-actions]

Closing for now — inactive for too long. Please open a new issue if this is still relevant.