[abruption]

Summary

The canonical saved-thread queue does not support native Codex /btw ephemeral Side Sessions. Add an isolated TypeScript experiment that connects to the existing owning app-server over a private Unix WebSocket and submits one exact-target input only under an explicit dedicated single-producer prerequisite.

The prototype requires loaded/ephemeral/idle/direct-input metadata, limits frames/inventory/timeouts, refuses unexpected approval interactions and never resumes, forks, falls back to a parent or retries an uncertain submission. Its private manifest pins ws 8.22.0 outside root dependencies. All six changed files are under experiments/codex_side_session/; npm packaging excludes them.

Refs #81; milestone Future — Agent integrations. Keep this PR Draft and the issue open: the experiment establishes transport feasibility, not general-use safety.

Actual TS validation

macOS arm64, Node 24.16.0, Codex CLI/server 0.159.2:

• A dedicated native TUI created a real /btw; no API-fork substitute.
• Exact-target dry-run passed, followed by one turn/start input. The rendered side TUI independently displayed ACK TS81-SIDE-ROUNDTRIP-20260930.
• Parent remained idle with one turn and the same serialized-turn hash. Adapter refused the persistent parent; native side queue probe returned RPC -32600.
• After Ctrl+C closed the side UI, the retained side still passed a dry-run. No further input was sent.
• Strict type checking and 20 contract/actual Unix-WebSocket fixture tests passed. Audit found zero vulnerabilities. Existing package/repository checks passed; packed contents were unchanged and exclude the experiment.
• Dedicated TUI/server PTYs and their test record/socket were cleaned up; unrelated work preserved.

See detailed results.

Unmet gates

turn/start invokes native start_or_steer_turn; the idle preflight cannot atomically exclude concurrent input. The exclusive-test flag does not lock out other clients. Loaded metadata also does not prove the UI remains open: cooperating-owner active registration/generation is needed.

No concurrent-race or multi-client approval test, live Linux/Windows/SSH/Orca shared-server validation, general ACK wait or automatic Reply-To. No canonical CLI, release, global installation or npm publication changes. This experimental adapter is not exposed by the installed CLI.

Response classification correction — 2026-10-14

Current correction head: 0e8ad2077d5701d1be7837d168998f448c659f5d.

• A matching RPC response must contain exactly one of result and error.
  Native refusal requires an error object with a safe integer code. Contradictory
  or malformed envelopes terminate the experiment connection; after an input
  attempt they produce unknown, submitted:null, retryAllowed:false.
  Valid standalone -32600 / -32601 refusal remains refused.
• Added 12 private fake WebSocket-over-Unix + actual experiment main() output
  cases for contradictory/missing outcomes, malformed error codes, and clean
  refusal/internal-error controls. Every case observed exactly one turn/start,
  no queue/fallback/replay, and no message content in output.
• Restoring only the original adapter made the three contradictory result/error
  regressions fail: 29 passed / 3 failed. Restoring the correction gave 32/32
  tests passed on macOS arm64 / Node 24.16.0. npm run typecheck passed;
  npm ci --ignore-scripts audited six packages with zero vulnerabilities;
  git diff --check passed.
• Only three experiment files changed. The historical 0.159.2 native results
  above were not repeated and do not establish current Codex/platform support.
  The existing idle/admission, active UI lifetime, owner generation, concurrent
  producer, Linux/Windows/SSH and ACK limitations remain. Keep Draft; no merge,
  release, packaging/runtime/operations change or new live/native/model test.
• The root CI does not execute this isolated experiment test suite. The 32-test
  result is local fake-owner evidence, not a claim of CI/live native coverage.
  Exact-head repository CI 38488449887
  completed successfully (macOS/Ubuntu Node 22/24, Windows native 22/24,
  policy and release gate). CodeQL 38488444554
  and GitGuardian passed: 12/12 current-head checks. These root checks do
  not run the 32 scoped experiment tests. No automatic review finding on this
  correction head was observed at this checkpoint; review completion is not
  claimed. PR remains Draft, unmerged.

[abruption]

Validation complete at head a6d1daff62d9fee7556fcebba18510136c457e38.

• CLI/package compatibility CI: all 8 jobs passed, including release gate. CodeQL and security checks passed.
• Manual experimental checks: strict TypeScript check, 20 contract/actual Unix-WebSocket tests, audit (0 vulnerabilities), exact package smoke and repository checks passed. Experimental tests are not part of the root CI suite.
• Actual native /btw on macOS arm64 / Node 24.16.0 / Codex 0.159.2: exact side ACK observed, parent hash unchanged, parent target refused, side queue probe rejected.
• UI dismissal still left the side dry-run validated. No subsequent input. Concurrent steering race is not solved or adversarially tested. Linux/Windows live side-session tests remain unverified.
• Dedicated server/TUI PTYs, created parent record, socket/rendezvous and unheld test startup lock were cleaned up; other resources preserved.

PR remains Draft and #81 remains open until the general-use gates are met. No merge or publication.