Follow-up behavior: add a "steer at next tool call" mode

[FilipeSilvens]

Current modes

• Steer: sends the message to the running turn immediately.
• Queue: (since the V2 orchestrator, feat(orchestrator): introduce new orchestrator #2829) holds the message until the whole turn finishes.

Why I'd like it kept as its own mode

Before V2, Queue sent a message after the next completed tool call (#11673, reported as a bug in #12100). I found that timing useful, and it matches how Claude Code handles mid-turn input.

Steer sends right away regardless of what the agent is doing, and Queue can't redirect the agent until the whole turn ends. Delivering at the next tool call sits between the two.

Proposal

Add a "Steer at next tool call" option to Follow-up behavior. It brings back the pre-V2 Queue behavior (V1 builds, 0.0.42–0.0.45). The Queue and Steer defaults stay unchanged.

Related

#6684 · #12100 · #14065

[ianrowe12]

Moving this over from #24399, which was closed as a duplicate of this discussion.

A data point for ACP providers, where this gap is wider than for T3's native Claude or Codex providers.

T3's current ACP integration declares no active steering, so for ACP Registry agents Steer cancels the prompt and starts a new one. (Some ACP adapters offer their own extension for this, such as _session/steering in the Claude and Codex ACP adapters, but T3 doesn't use one.) With GitHub Copilot, that cancel kills any background agents still running (#26695, reproduced at the protocol level on Copilot CLI 1.0.105). Copilot keeps session/prompt open while background agents run, so a message sent during that window is either a destructive Steer, or it sits in Queue until every background agent finishes. Neither gives "read it at the next safe boundary".

For providers without a supported non-interrupting route, T3 does control one safe boundary: the result of a call to its own t3-code MCP tool. One provider-neutral shape:

• A message the user designates for non-interrupting mid-turn delivery can go out with the next T3 MCP tool result in that turn, then be recorded in the timeline as delivered mid-turn. Messages queued for after the turn stay queued, so Queue and Steer keep their meaning, in line with the separate mode Follow-up behavior: add a "steer at next tool call" mode #15287 asks for.
• Agents get a small tool that waits for such input, so an agent waiting on CI or a delegated task can pick up new instructions without ending its turn.
• Messages that need a new turn always wait: attachments, a model or mode change, /compact and restart continuations.

Limits: delivery only happens when the agent calls a T3 tool, so this makes in-turn delivery possible, not guaranteed, and the UI would have to say "delivered at the next T3 tool call", not "steered". It also depends on the T3 tools being available to the agent and the agent using them.

A fork has an implementation of this, tested against a simulated provider; I haven't measured how often real agents pick up messages through it. I'm not proposing it as a fix for the whole discussion, only as a possible delivery mechanism for providers that can't take input mid-turn. If maintainers want to explore it, I can write it up as a separate proposal.

— Opus 5.5 with Claude Code, reviewed by GPT 6.1 Sol

[guryanov-a]

+1. A data point on why today's Steer isn't a substitute, from T3 Code 0.1.6 with Claude Opus 5.5 (Claude Code 2.1.316). The agent had issued two calls: a long Bash search and a ToolSearch waiting behind it. I steered in a short "ok, post it" while the search ran. Bash finished normally, but the waiting ToolSearch was cancelled and returned Claude Code's "The user doesn't want to take this action right now. STOP what you are doing…". The agent acted on my message, then dropped the rest of its task and told me I had stopped it.

A "steer at the next tool call" mode would avoid this. Every call the agent has already issued runs to completion, and the message is delivered before the next model request, in the same turn, like Enter in the Claude Code CLI. With nothing cancelled, the agent has no refusal to misread and carries on with its task unless the message tells it otherwise.

The SDK experiments in #21895 suggest Claude's priority: "next" gives this behaviour during tool calls, with the caveat noted there for messages sent while no tool is running.

[TechyAditya]

+1. Most of my mid-run messages aren't corrections. They're extra context: a constraint I just remembered, a file the agent should look at, or a thought about the approach. I want the agent to read it at its next tool boundary and keep going.

Today I have to pick between Steer and Queue. Steer can cancel the calls in flight and makes the agent think I stopped it (#21895). Queue holds the note until the turn ends, often after the agent has already gone the wrong way. So I either sit on the thought or stop the run myself.

A "steer at next tool call" mode covers this. I'd also like it as the default for Enter, with the interrupting Steer kept for the modifier key. That matches the Claude Code CLI, where Enter delivers after the running tool calls and Esc interrupts.

[ahmedisam99]

+1. My mid-run messages are a mix of extra context and corrections, with extra context a bit more common.

Claude Code handles this more smoothly. Its docs say Enter delivers my message as soon as the running tool calls finish, in the same turn. When I want it in right away, Ctrl+X Ctrl+S sends it now, and running shell commands and subagents move to the background first. My guess is that Enter uses priority: "next" and send now uses priority: "now", which is what T3 Code sends for every steer. This is me speculating.

I've also had a steer make Claude drop my original request, the same way as guryanov-a's case above. I asked it to prepare a new release. While one Bash command was running with a git diff queued behind it, I asked "do we request any new permissions?". The queued call came back with "The user doesn't want to take this action right now. STOP what you are doing and wait for the user to tell you how to proceed." Claude answered my question and ended the turn without going back to the release. So Claude isn't just misreading my steer. The interrupt hands it a message saying I stopped it.