fix(server): Claude steers no longer fail when only next-turn options differ

[juliusmarminge]

Problem

Steering a running Claude turn from web failed every time with "Claude cannot redirect an active run. Stop it first, then send the message." (seen on a thread started from mobile).

dispatchSteerIntoRun forced interrupt-and-restart whenever the steer's model selection differed from the running run's, however small the difference. The web composer adds an explicit fastMode: false that mobile-started runs don't carry, so the selections never matched. Claude supports live steering (supportsActiveSteering) but not supportsSteeringByInterruptRestart, so the forced restart was always rejected:

CommandPolicyCapabilityUnsupportedError: claudeAgent cannot satisfy interrupt_restart_steering:
providerInstanceId cannot satisfy a required interrupt-and-restart

Fix

A changed selection now forces a restart only when it has to apply immediately:

• Must apply now (provider switch, or a transition plan other than apply_on_next_turn): restart as before, or reject if the provider can't restart.
• Turn-scoped change, provider can restart (Codex, OpenCode, ACP): restart as before.
• Turn-scoped change, provider can't restart (Claude, Pi): the steer joins the running turn, and the thread records the new selection (thread.model-selection-updated) so it applies from the next turn. Previously these providers rejected the steer.

Verification

• New SteeringCompletion.integration.test.ts case: a provider that can steer but not restart, receiving a steer whose options differ (fastMode: false). It steers once, starts no second attempt, keeps the run's selection and stores the new one on the thread.
  • Before (base Orchestrator.ts): × steers a changed turn-scoped selection into a provider that cannot restart, failing with codex cannot satisfy interrupt_restart_steering … cannot satisfy a required interrupt-and-restart
  • After: passes.
• SteeringCompletion, SelectionRestart, CommandPolicy: 38/38 pass. Server typecheck clean.

This is a server-only change, so there are no screenshots; the before/after test above shows the behaviour.

Part of the fix for the hung "Add Find in Diff View" thread. The related client fixes are separate PRs (paused-queue status, Resume queue action).

🤖 Generated with Claude Code

---

[github-actions]

Thread transfer impact

✅ Thread transfer remains within every enforced ceiling.

ℹ️ No successful main baseline artifact is available yet. This run establishes the initial measurement.

Provider	Metric	Main baseline	This PR	Impact	PR ceiling
Codex	Total thread wire	—	4.9 KiB	—	6.8 KiB	✅
Codex	Thread snapshot wire	—	3.7 KiB	—	4.9 KiB	✅
Codex	Live turn WebSocket wire	—	1.2 KiB	—	2.0 KiB	✅
Codex	Live turn WebSocket decoded	—	20.4 KiB	—	29.3 KiB	✅
Codex	Live turn messages	—	2	—	8	✅
Claude	Total thread wire	—	4.9 KiB	—	6.8 KiB	✅
Claude	Thread snapshot wire	—	3.7 KiB	—	4.9 KiB	✅
Claude	Live turn WebSocket wire	—	1.2 KiB	—	2.0 KiB	✅
Claude	Live turn WebSocket decoded	—	20.8 KiB	—	29.3 KiB	✅
Claude	Live turn messages	—	2	—	8	✅

Baseline: unavailable · PR result: ac085d7 · Source CI: success

Scenario and decoded snapshot size

10 historical turns, 5 command tools per turn, 878.9 KiB retained MCP result per historical turn, and a 1.05 MiB retained result in the measured turn.

• Codex decoded thread snapshot: 106.1 KiB
• Claude decoded thread snapshot: 106.4 KiB

Updated in place by a trusted workflow. PR artifacts are strictly validated and never executed.

[macroscopeapp]

Approvability

Verdict: Approved at ac085d7

Macroscope's review found this PR approvable — This is a localized server bug fix that avoids rejecting active Claude steering when only next-turn options differ, while preserving restart behavior when it is required and recording the selection for the next turn. A focused integration test covers the changed path, with no schema, deployment, security, billing, authentication, or default-setting changes.

^{You can add or adjust custom eligibility rules. Learn more.}