> ## Documentation Index
> Fetch the complete documentation index at: https://docs.mesh.texturehq.com/llms.txt
> Use this file to discover all available pages before exploring further.

# Steer active work

> Send corrections or stop requests while an agent is working.

Live steering is automatic for running Slack and Telegram tasks. A text follow-up
from the original requester in the same conversation and audience joins the active
run. It does not start a second run or require a setting.
Background commitment attempts accept explicit stop commands through this
mailbox, but leave other follow-ups queued for a conversational turn. An
autonomous objective must not swallow a person's new question or force it into
the objective's restricted scope; the next turn reads the conversation history.
In channel threads, the active requester can follow up without mentioning the
agent again, even before the agent has posted its first answer.

* Corrections reach the next model call with author and message attribution. If
  they arrive during a model call, its obsolete proposal is discarded before its
  tools execute or its answer is delivered. Replacement calls use the run's
  existing budgets.
* Every tool checks for updates before starting. A correction skips remaining
  tools from an obsolete batch; observations from completed tools stay in context.
* A whole-message command such as `stop`, `please stop`, or `cancel this task`
  cancels the run's context, including a blocked model or cooperative tool. The
  watcher polls once per second. Cancellation cannot retract a sent message or
  undo an external operation already accepted by another service.
* Mesh acknowledges updates in the original thread. Bursts receive one
  acknowledgement drafted by the agent's configured fast model from the original
  request and latest incoming message. It recognizes questions and corrections
  without quoting them, inventing answers, or exposing runtime jargon. The draft
  gets four seconds within an eight-second acknowledgement budget. Failed,
  malformed, or unnecessary drafts stay silent without losing the steering input.
  Stop acknowledgement does not wait for a model. Source access is checked again
  after drafting and before delivery. Sends are
  reserved durably before delivery; an unknown delivery outcome is not retried.
  Periodic contextual progress continues separately.
* A stop from the requester also cancels commitments created by that run and its
  active commitment attempt, preventing autonomous reconciliation from restarting
  the stopped work. A stop from another participant ends the run only; the
  requester's commitment stays theirs (see below).

The durable mailbox claims ready intake under the conversation lock and records
each message as a run trigger in the same transaction that retires its queued
turn. A generation fences older watchers during takeover. Source authority and
audience are checked again when consuming and acknowledging updates. Trace views
can resolve all steering messages to the same run.

Finalization seals the mailbox only when the model has seen every admitted
message. Messages prepared after that cutoff stay queued for the next turn.
Other participants cannot redirect the requester's task, but any participant may
**stop** it: a whole-message stop from anyone in the thread cancels the run,
while their other messages stay queued for the next turn.
A bystander's stop ends the current run only; a commitment behind it remains the
requester's and the reconciler may schedule another attempt. Bot messages, changed
audiences, file-bearing follow-ups, and authenticated memory-sharing or commitment
resume commands retain their normal queued handling. A direct stop still works
when accompanied by a file. Broader natural-language cancellation reaches the
model at its next boundary; exact stop commands take the immediate watcher path.

The input limit is 128 admitted messages or 64 KiB of text per run. Exceeding it
stops execution rather than silently omitting instructions. This version does
not incorporate new images into an active model context, and shared channel
participants may stop but not redirect another person's run.

Tests cover stale final answers, partial tool batches, blocked model and tool
cancellation, requester isolation, transactional rollback, finalization races,
takeover acknowledgement deduplication, and commitment cancellation.


This documentation is built and hosted on [Mintlify](https://mintlify.com), a developer documentation platform.