[sak0a]

What Changed

• GitHub Copilot, OpenCode and Cursor can now be steered. They can't add a message to a running turn, so a steer stops the turn and the message is sent next. Until now the composer said "steering unavailable" for them, and the message waited for the turn to end.
• Those adapters declare turnSteering: "interrupt-restart". For them, ProviderService.steerTurn interrupts the expected turn and then defers the steer (ProviderTurnNotSteerableError, reason restarting). The message goes back into the node's follow-up queue and is sent as the next turn once the stopped turn settles.
  • The deferral carries restartedTurn (an optional flag), and the node's queue moves that message to its head, so the stopped turn is followed by the steered message, not an earlier follow-up.
  • A node's own stop holds no queue; only a person's Stop does.
  • If the turn cannot be stopped, the message stays queued and the activity says so.
• Their provider status carries turnSteeringMode: "interrupt-restart". It is a plain string, so a mode an older client does not know reads as no steering instead of failing the provider decode. A shared providerSteering() in client-runtime reads it for web and mobile.
• Wording: the composer says "Stop this turn and send this next · ⌘↵ to queue instead", and a queued message's Steer button (web and phone) says "Stop the running turn and send this message next".
• While approvals wait, a message still redirects on these providers: it declines the approvals and then stops the turn (documented).
• Only explicit steers stop a turn: with the follow-up setting on Queue, Enter still waits for the turn to end. ACP registry agents keep queueing, as in T3.
• Docs: docs/providers/approval-response-policy.md → "Steering providers without native steering".

Why

Port batch 9c, item "capability-driven adapter degradation". The report's recommendation was to give every provider steer semantics through an interrupt-and-restart steering mode, rather than rewrite capabilities into T3's 11 groups. This follows T3's policy (docs/orchestration-v2/provider-capability-system.md: "active steering unsupported → interrupt active turn and restart the run"; CommandPolicy.ts decideSteeringExecution; adapters with supportsSteeringByInterruptRestart: true for Cursor, OpenCode and ACP, false for the ACP registry), adapted to Ryco's server-owned follow-up queue.

T3's ForwardCompatibleArray/Optional decoding was not ported wholesale. The new status field uses a plain string, the same forward-compatible pattern Ryco already uses for lineage relationships.

Validation

• bun run check and bun run check:changed pass. Server boot +931 B.
• New tests, all failing without the change:
  • ProviderService.test.ts: a Cursor-shaped adapter in interrupt-restart mode has the expected turn interrupted, gets a deferred steer with the restart wording and never has steerTurn called. A failed interrupt defers with "could not be stopped. The message stays queued."
  • logic.test.ts (client-runtime): providerSteering reads native, restart, an unknown future mode (no steering) and absent status.
  • composerFollowUp.test.ts: the restart wording for both follow-up settings.
  • decider.followUpQueue.test.ts: a steer deferred with restartedTurn moves to the head of a queue with earlier follow-ups; turnSteerFailure.test.ts: only the restarting reason sets the flag.
• Not verified against real Copilot, OpenCode or Cursor processes that a stopped turn always settles promptly. If one does not, the message waits until the turn settles, as before.

🤖 Generated with Claude Code