Skip to content

Revisit queue as the default landing for agent messages (due 2027-01-11) #538

Open

Description

@Jacksondr5

Why

Deciding #519 on 2026-10-11, Jackson kept j5_send_message and gave the sender upstream's landing choice (queue, steer, restart), with queue as the default. Upstream's auto was left out, because it steers whenever it can and that brings back the harm D4 records: a steer cancels a Claude agent's in-flight tool calls and makes it report a refusal the person never gave, and on Cursor it destroys the turn.

Jackson asked for the default to be revisited in three months: agents are getting much better at being steered, and that needs to reach the cheaper models before the default can change.

What to check, on or after 2027-01-11

  1. Does a steer still abort tool calls on Claude, or restart the turn on Cursor? Test the models J5 agents actually run, including the cheaper ones, not only the top model.
  2. Has upstream changed when it steers? activeSteeringInterruptsTools is its flag for a steer that interrupts tools (see the Steering and queues entry in docs/j5/product/upstream-convergence.md).
  3. How often did senders choose steer or restart since the option shipped, and did those sends go well?

Outcomes

  • Still harmful: keep queue as the default and set the next check date.
  • Safe on the models in use: make upstream's auto the default, retire D4 and its Astra exception, and follow upstream's delivery.

Related

Activity

  1. Jacksondr5 commented on Oct 12, 2026

    @Jacksondr5
    OwnerAuthor

    Context for the revisit, from Jackson on 2026-10-12, while the #519 stack was being tested:

    • What "steering" means is changing. The Codex GPT 6 models are built to take a message during work and handle it "without missing a beat", which is why feat(j5): peer messages reach every running Codex GPT 6 turn mid-turn #569 lets a peer's ordinary message into a running GPT 6 turn. On Claude a steer is still disruptive (the agent reports that the user does not want an action), and on Cursor it kills the turn and starts a new one.
    • He expects most models to handle steering gracefully within about six months, and accepts the inconsistency between providers until then.

    Two things the #519 live test saw that bear on this:

    • With feat(j5): peer messages reach every running Codex GPT 6 turn mid-turn #569, the default (queue) already steers on GPT 6 receivers. So the question for this issue is which other providers can join them, not whether to start.
    • Asked in plain words to interrupt another agent, Claude Haiku picked mode correctly in four of six uncoached requests: two of its three sessions got both a steer and a restart right, and one got neither. That is accepted as model variance.
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Metadata

Metadata

Assignees

No one assigned

    Labels

    No labels
    No labels

    Projects

    No projects

      Milestone

      No milestone

      Relationships

      None yet

      Development

      No branches or pull requests

      Issue actions