feat(codex): investigate opt-in live steering with instant_interrupt #247

Motivation

Investigate explicit opt-in steering of an already-running Codex turn using its native instant_interrupt capability, without changing default session-peer queued delivery.

The feature was introduced in Codex 0.159.0 and remains under development/default-off in installed 0.159.2. It preempts model responses and yields foreground code-mode observation waits when new current-turn user input arrives. It is not turn/interrupt, a process kill, rollback, or generic task-completion/ACK facility.

Current session-peer invokes codex queue. Codex's queued-item service dispatches only when idle; enabling this feature alone does not convert ordinary sends into steering.

Evidence: official changelog, App Server steering, session-peer session_peer_core/codex.py, and Codex upstream tag rust-v0.159.2 (ff6aec96948b70d94983af2641a6b67c94faeff5). Related live-server discovery experiment: #245 / Draft #246. This proposal does not replace those side-session lifetime requirements.

Alternatives considered

• Enable a feature flag only on a sending codex queue process: does not steer an active destination.
• Cancel the current turn with turn/interrupt and send a new message: changes cancellation semantics and introduces a two-operation failure window.
• Use native turn/steer with an exact nonempty expectedTurnId: preferred candidate for explicitly authorized current-turn input.
• Inject keystrokes, kill an existing writer, alter queue/state DB files, auto-resume, or retry ambiguous outcomes: excluded.

Proposed behaviour

• Retain queue-only default. Evaluate a distinct explicit steering mode (CLI spelling is not finalized).
• Connect to the existing owning app-server selected by the operator; no replacement server, parent fallback, or endpoint supplied by an untrusted peer message.
• Exact home/server/thread/active-turn validation; native expectedTurnId handles turn replacement races.
• Target owner selects instant_interrupt. Do not modify global configuration or active-session permissions. Disabled or unverified instantaneous preemption must not be advertised as enabled.
• New steering permission/capability is separate from ordinary send/wake, including any future MCP/paired integration.
• Acceptance is not consumption or ACK; bounded results distinguish refused, accepted, and unknown. No automatic queue fallback/resend.
• Document that foreground yielding can leave code cells, commands or external work running, and cannot undo prior side effects.

Acceptance criteria

• Add isolated experiment code and source-grounded design notes without modifying canonical runtime.
• Run actual pinned native Codex against an isolated loopback streaming model-API fixture: queue vs steer × feature OFF/ON. Record whether the corrected request occurs before releasing the original response; label synthetic model content clearly.
• Confirm queue remains next-turn in both flag states, and native steer targets the original active turn without creating a separate user turn.
• Cover missing/closed/idle targets, wrong expected turn, unavailable server, disabled/unknown feature evidence, and malformed/disconnected/timeout outcomes.
• Verify exact one-shot submission, no permission overrides, no auto-approval, no fallback/retry, and bounded cleanup of owned processes.
• Record provider/model, Codex version, platform, sample count/timing tolerance, and untested real-model/SSH/Windows boundaries.
• Prepare a Draft PR with findings and unmet general-use gates; do not claim a stable new delivery mode or close this issue prematurely.

Milestone: v1.1.0 (new opt-in functionality, not v1.0.3 security scope). User authorized technical review, issue registration, isolated tests and Draft PR creation; no merge, publication, existing-session interruption or production change is authorized.