Feature request: support steering and interrupting active chat/actions mid-run

[stefanpieter]

Feature request

Please add better controls for steering and interrupting an active chat / action while the agent is working.

Problem

During longer agent runs, I sometimes want to:

• stop the current direction
• steer the agent toward a different approach
• interrupt an action before it finishes
• inject a correction or constraint without waiting for the whole turn to end

Right now this feels too rigid. Once the run is underway, it is hard to intervene in a smooth way.

What I want

A better mid-run control model, for example:

• interrupt the current run/action cleanly
• send a steering message / correction while the session is still active
• pause / resume if appropriate
• keep the workflow context while changing direction

Why this matters

For real coding workflows, steering is often necessary when:

• the agent misunderstood the task
• a new constraint appears
• I want to redirect effort before more time/tokens are spent
• I can already see the action is going the wrong way

Suggested UX directions

Possible implementations:

• a visible Interrupt / Steer control while a run is active
• allow an inline follow-up instruction that becomes a higher-priority steering message
• support a distinction between:
  • hard cancel
  • soft steer / redirect
• preserve visible context so the conversation does not feel reset after intervention

Scope

This request is specifically about live intervention during active work, not only cancelling after the fact.

---

Originally filed by mistake in formulahendry/vscode-acp: formulahendry/vscode-acp#31

[stefanpieter]

Hermes Agent now has first-class /steer <prompt> support upstream: NousResearch/hermes-agent#12116.

Could the Hermes VS Code extension expose this in the active-run UI?

Suggested UX:

• show a Steer action while a Hermes session is actively running;
• accept a short instruction/correction from the user;
• send it through the Hermes active-session steering path instead of treating it as a normal next turn;
• keep a separate Interrupt/Cancel affordance for hard stops, so soft redirect and hard cancel remain distinct.

Why: /steer injects a mid-run note after the next tool call, preserving the active workflow while letting the user redirect before the full run completes. This maps well to VS Code because the user can see the agent going in the wrong direction and correct it without waiting for completion.