[abruption]

Summary

Add a local-only experimental steering adapter and an opt-in native/synthetic-provider comparison harness. Refs #247; the issue stays open and this PR remains Draft. Existing-owner side-session research is tracked separately in #245/#246.

The adapter connects to a private owning app-server and uses turn/steer with an exact expectedTurnId. It preserves the same turn and never starts/interrupts/resumes a target, changes a feature, answers approvals, falls back or resends. Metadata dry-run does not prove that the expected turn is still active. Acceptance is not consumption or ACK.

Malformed/contradictory RPC envelopes, invalid error fields and excessive JSON nesting fail with fixed diagnostics; invalid replies after an attempt remain unknown/no-retry. Ordinary connection-close errors cannot erase independently known submission or the original unknown outcome.

The native harness shares one 45-second active-work budget across startup, RPCs and provider waits. Its watchdog sends TERM then KILL after a 0.2-second grace, keeping the leader unreaped until the watchdog has joined. The version probe has a separate five-second timeout. Cleanup has additional bounded waits; these are application budgets, not guarantees against blocked OS calls.

Only experiments/codex_steering/ changes. Installed runtime, native queue, SSH/Relay, package configuration and operational settings are unchanged.

Validation

• Current macOS arm64/Python 3.13.7 offline suite: 43 tests pass, with development mode and ResourceWarning-as-error. Includes fake RPC/deadline tests, batch-abort behavior and a fixture-owned Python process that ignores TERM; no native Codex or model was started.
• Independent subagent findings were checked, fixed and re-reviewed. Generator and diff checks pass.
• Experimental tests are run explicitly; the default repository suite does not discover them. Repository CI is checked separately.

Historical evidence and remaining gates

experiments/codex_steering/RESULTS.md preserves the Codex 0.159.2 native ON/OFF comparison against source ff6aec96948b70d94983af2641a6b67c94faeff5: four conditions with three repetitions each passed using synthetic SSE responses. ON steering produced a second request before release of the first-response gate; OFF steering and either queue condition waited for release. Steering retained the turn, while queue started a new turn. This historical matrix was not rerun for the current hardening changes and does not establish real-provider ACK or latency.

Remaining gates include real-provider/code-mode behavior, existing-owner lifecycle/authorization, current effective feature policy, framing/Reply-To integration, concurrent turn replacement, SSH/Linux/Windows, and any newer native version. expectedTurnId does not prove a session generation or side-UI lifetime. No general-support promotion, merge, publication or operational rollout is included.

[abruption]

Optional RPC robustness improvement: validate a present error object and exact integer code before accessing fields. Malformed null/array/string errors or missing/null/bool/float/string/container codes enter the existing unknown/no-resend path when a submission was attempted, not traceback or a false success. Existing supported-code refusal semantics remain.

28 in-memory RPC tests pass on Python 3.9.6 and 3.13.7; default discovery and generation checks pass. No actual app-server/model/TUI/ACK or instant-interrupt setting was exercised. Experimental feature gates remain outstanding; Draft retained, no merge/release/deployment.