[abruption]

Summary

Add a local-only, explicitly opt-in experiment for addressing a loaded ephemeral Codex thread through its existing owning app-server. Refs #245; the issue remains open. The review scope is this bounded research artifact, not general /btw support.

The adapter uses a private Unix-domain WebSocket, exact UUID selection, required absolute --expected-home matched against canonical server metadata, and a pinned 0.159.2 server version. It never resumes, forks, wakes, retries, redirects to a parent, changes permissions or answers approvals. Acceptance is separate from consumption and ACK.

RPC responses must contain exactly one result or error, with a matching integer request ID and typed error fields. Malformed or deeply nested JSON becomes a fixed refusal before input and an unknown outcome after an attempt. Malformed initialization/discovery fields cannot escape as tracebacks. Ordinary transport-close errors preserve the primary outcome, including known submission and unknown/no-retry results.

Changes are limited to experiments/codex_side_session/ and explicit offline experimental-suite execution in CI. Existing packaging excludes experiments; installed CLI/runtime, SSH/Relay and services are unchanged.

Validation

• Current offline validation on macOS arm64/Python 3.13.7: 32 tests pass, with development mode and ResourceWarning-as-error. Tests use fake RPC/connection objects and make no native/model calls. Additional home/socket/version, approval-request and idle-to-native-rejection cases preserve one-shot semantics.
• Independent subagent review findings were checked, fixed and re-reviewed. Standalone generator and diff checks pass.
• Linux/macOS Python 3.13 CI now explicitly runs available Side Session/Steering experiment suites. No native fixture is launched by that step. Local verification also runs in a dedicated tmux session.

Historical native evidence and remaining gates

The preserved experiments/codex_side_session/RESULTS.md records the isolated native /btw test on macOS arm64 with Codex 0.159.2 and websockets 17.1: an exact on-screen side ACK and unchanged parent-turn hash, plus negative controls. That historical experiment was not rerun for these parser/cleanup changes. submitted:true still has consumptionConfirmed:false; no generic ACK/return-route workflow is implemented.

Two general-use blockers remain:

1. The idle check and turn/start are not atomic; another producer can cause unintended steering. The exclusive-test-session flag acknowledges a dedicated single-producer restriction but cannot enforce it against native clients.
2. Closing the side UI does not immediately unload its thread. Loaded/idle metadata does not establish current side-UI lifetime or generation.

The prototype classifies ephemeral threads as unclassified, not automatically as side conversations. Linux/Windows, SSH, arbitrary app-hosted endpoints, concurrent senders and general return-route integration remain unverified. No merge, publication or operational rollout is included.

[abruption]

Optional RPC robustness improvement: a present error must be an object with an exact integer code (not bool/float/null). Malformed error shapes use the existing pre-attempt refusal/post-attempt unknown handling instead of traceback; a plausible result beside an error does not mask it. Existing known-code classification and no automatic resend are preserved.

17 in-memory RPC tests pass on Python 3.9.6 and 3.13.7; modern/default-floor discovery and generation checks also pass. No real app-server/model/TUI/ACK was invoked. This does not satisfy the experimental feature's remaining provider/platform/lifetime gates; Draft retained.