[abruption]

Summary

Current harness limitation (P1 correction): the current probe no longer uses account-home SQLite rows as freshness authority. It requires server-visible empty-turn snapshots and fails closed when history is unavailable. Native 0.162.1 can reject this lookup before the first user message, so the revised harness may not reproduce the original unsaved-first-input positive. Both modes require explicit --allow-server-history-db-init because history lookup can initialize/migrate a server DB. This remains a research/report PR, not implemented general first-input support; #306 stays open. Older review sections below describe earlier revisions; the final storage-authority section records the current behavior.

Refs #306; keep the investigation open for product acceptance work.

Record isolated first-input experiments without changing the public runtime:

Already-open target topology	Before submission	Result
Dedicated App Server + explicit remote TUI	One loaded idle thread, zero matching stored rows	Queue accepted; exact reply observed in that TUI
Ordinary embedded TUI with effective daemon auto-start disabled	UUID writer lock exists, zero matching stored rows, no external control socket	One official codex queue refused with no rollout found; TUI remained blank
Ordinary TUI already connected to an exclusively owned shared server	No --remote flag or initial prompt; one loaded idle thread, zero matching stored rows	External Unix-WebSocket queue accepted; same TUI displayed ACK CODEX-SHARED-EMPTY-306-8C2F

The third experiment started a foreground server before the plain TUI, then attached an external client after that TUI was open. It did not retrofit the embedded TUI, create a replacement conversation, or prove arbitrary blank-TUI attachment. Queue acceptance is submission evidence; the separately observed exact native-TUI reply is test receipt evidence, not a new generic ACK API.

Add the reports and an opt-in, exclusively owned-server experiment harness under experiments/codex_first_input/. The harness uses existing loaded IDs, thread/read(includeTurns:false) and thread/queue/add; never thread/start, thread/resume, turn/start, fallback, or uncertain-effect retry. Only selected metadata is reported. includeTurns:false is not a universal promise that the API has no preview fields.

Type

• docs
• chore / refactor / test / perf
• feat
• fix

Verification

• python3 tools/generate_session_peer.py --check: PASS.
• python3 -m unittest discover -s tests/docs -t . -v: 60 PASS.
• Experiment harness AST parsing and git diff --check: PASS.
• Exact-version upstream rust-v0.162.1 static source review and separate read-only report/harness review: no blocking findings; this is not an additional independent live run or external CLI review.
• Actual owned macOS TUI observations: both positive topologies displayed their exact requested replies; the embedded negative did not. No unrelated transcript files were read.
• Existing packaging excludes experiments/; no runtime dependency, package metadata, generated standalone, or release change. Websockets 17.0 was installed only in a temporary test environment, then removed. Packaging CI must confirm the existing exclusion.
• No separate Codex review mention was posted; repository automation handles review on PR creation.

Checklist

• Conventional Commit; no private paths, session UUIDs/PIDs, credentials, tokens or session keys in the diff.
• Reports distinguish shipped runtime from feasibility and document exact-version/platform limits.
• Test-owned TUIs/server exited zero; PIDs absent; created sockets and temporary environment removed; existing startup lock preserved.
• Required current-head CI and automatic review completed.
• Ordinary list, SSH list and public send dry-run: not rerun for this report-only change; the experimental RPC target preflight was performed separately. The template's legacy cc_peer.py checklist is not used to claim public runtime support.

Verified against

Codex CLI/App Server 0.162.1, macOS arm64, ordinary shared-server TUI selected GPT-6.1-Sol medium. Source baseline: published session-peer v1.1.1 / b48144bdf2ad1569031562136f5a60e33780020e. Experimental WebSocket client used websockets 17.0. The harness runs only by explicit operator invocation; CI must not launch authenticated real model sessions.

Notes / remaining gates

• No product adapter or stable support is implemented. A future adapter needs endpoint ownership, native PID/start/generation binding, version/capability checks, strict protocol validation, bounded discovery/cleanup and approval routing. The harness's assertions and exclusive-server preconditions are not production safeguards; do not run with optimized Python or against unrelated work.
• Embedded/in-process TUI has no confirmed external attachment/retrofit path. Refuse unavailable targets rather than start another conversation pretending to be the original receiver.
• The preliminary stdio-proxy inspection timed out before initialization and before submission. The successful control used Unix WebSocket; no uncertain send was retried.
• Server exit was zero but emitted a best-effort MCP session-delete HTTP failure; the report does not call shutdown error-free.
• Linux, Windows, remote machines, arbitrary Orca-hosted TUIs, restart recovery and long-term stability remain unverified.
• No merge, tag, release, publishing, global settings, approval/sandbox policy, or operating service changes are authorized by this PR.

Review corrections for the v1.2.0 preparation

• Canonicalize relative Codex home and require the initialized server's authoritative codexHome to match before thread lookup, database inspection, or submission.
• Return exit 1 for RPC rejection; reject malformed error objects/codes without reflecting native text or retrying.
• Close SQLite connections explicitly. Four offline fake-WebSocket regression tests cover these guards; they do not prove live receipt or a production adapter.
• Updated this branch onto published v1.1.2 main. Historical live observations above remain v1.1.1 observations, not new v1.1.2 live tests.
• Current-head local checks: Codex 136 tests OK (3 skipped), documentation 61 tests OK, standalone generator and diff checks PASS. An initial generator invocation used a nonexistent script name; rerunning the repository's tools/generate_session_peer.py succeeded.
• Current-head CI and automatic review are pending; no merge or publication has been performed.

Independent CLI review before the follow-up update

• Claude Code 2.1.296, requested and response-confirmed claude-opus-5-5; tools disabled, non-interactive. Codex CLI 0.162.1, default model with ignored user configuration, read-only sandbox and ephemeral session; actual model name not exposed in captured response metadata. Both ran in parallel on identical frozen stdin, without seeing each other's results, twice (initial and amended candidate).
• Initial input SHA256: 522c18961a8c8b7a4a2e688b574f84c8feab005532b114dfe8ba0152598b04bf. Amended input SHA256: 95927739e86a1c002bd1607354936dbb245288f71567c96e9c4fb1508df37729. Both CLI executions completed successfully in each round; these were static reviews, not independent live experiments.
• Addressed the agreed finding: after any queue attempt, uncertain malformed responses, transport/JSON failures and interrupts preserve accepted:null/outcome:unknown/retryAllowed:false. Expected exceptions are sanitized; no raw native keys/text are reflected. Confirmed server rejection remains distinct from preflight not_attempted, without claiming absence of side effects or permission to retry.
• Amended Codex review found no mandatory changes. Amended Claude review found one documentation wording inconsistency about preflight accepted:false; root corrected it and clarified all review-driven guards lack a fresh native live run. These final explanatory sentences were checked by root, not re-run through the external CLIs.
• Nonblocking follow-up limits: queuedSubmission payload schema and connection-exit fixtures need upstream evidence/additional synthetic cases before production use. The harness is still not a product adapter; raw live evidence and authenticated delivery were not independently retested.
• Final head dc80705: focused 7 tests PASS; documentation 61 PASS; standalone generator/diff PASS. Worker also ran Codex 139 tests OK (3 skipped) before the final wording-only correction. Current-head CI/automatic review remain a separate gate. No merge, tag, release, publishing or operations changes.

Follow-up: new Codex findings on dc80705 (2026-10-15)

Final head: 5c80632.

• Both RPC error paths use rpc_error_code(): code must be an exact integer (not bool), and message must be a string. Missing/null/non-string messages after the queue attempt retain accepted:null/outcome:unknown/retryAllowed:false; a complete rejection returns exit 1 without reflecting message text.
• Queue mode requires ephemeral is False before database access or queue submission. True, null and missing metadata refuse with not_attempted. Inspect mode may report boolean/null without making any queue attempt or proving freshness. Normal positive fixtures now use a persistent thread.
• Ten synthetic tests cover malformed errors, complete rejection, ephemeral/null/missing refusal, non-submitting inspect metadata, prior uncertainty/interrupt/no-retry guards and home binding. Focused 10 tests, docs 61 tests, generator --check and diff checks PASS on final head. No new live native test.
• Pinned native 0.162.1 Thread schema declares ephemeral as an ordinary bool without an omission attribute; the report links the exact source. This is source evidence, not new native serialization evidence. Historical non-ephemeral observations are unchanged.

Independent CLI static review before pushing:

• Same immutable input SHA256: 8777f729ecb68d882651cd1a771deead8b99d4a74f763aae2533879092691880; reviewed implementation candidate 08c26ee.
• Claude Code 2.1.300, requested and response-confirmed claude-opus-5-5, safe mode/tools disabled/non-interactive; exit 0 and valid final response, 04:39:39–04:40:23 UTC.
• Codex CLI 0.162.1, ignored user config/read-only/ephemeral; actual default model not exposed; exit 0 and turn completed, 04:39:40–04:40:02 UTC.
• Both found the two findings addressed and no mandatory new defect. Root adopted optional subTest/inspect/rejection assertions and added pinned-schema documentation. The final probe.py is byte-identical to the reviewed candidate; the last test/doc-only refinements were root-validated, not independently re-reviewed. No reviewer executed tests or native sessions.
• Existing nonproduction limitations remain: response ID strictness, queuedSubmission payload validation and connection-exit fixtures are not newly certified. No product adapter, general empty-TUI support, Linux/Windows/live acceptance, automatic retry, merge or publication is implied.

The two new threads are resolved with implementation/test evidence. Latest-head CI and Codex re-review are requested separately, not claimed complete by this update.

Follow-up: submission payload and queue preflight (2026-10-15)

Final head ec924de addresses the two findings on 5c80632:

• Before the only add attempt, call thread/queue/list for the selected thread, first page, limit1. Require explicit data:[] and nextCursor:null. Missing/malformed/nonempty/paginated/error snapshots refuse pre-effect with not_attempted/accepted:false/retryAllowed:false.
• Validate queuedSubmission as an object with a nonempty string ID, exact submitted plain-text input and matching clientUserMessageId before accepted:true. Missing/null/malformed/mismatched fields remain unknown/null/no-retry after an attempt. The request explicitly includes native text_elements:[].
• RPC IDs must be exact integers matching the request; bool/float/absent IDs cannot silently qualify a response.
• Pinned 0.162.1 schema/service source confirms required string fields and plain-text echo preservation. The previous known queuedSubmission-validation gap is addressed here, not merely relabeled as unsupported.
• No atomic expected-empty/revision API exists in the reviewed add path. This is a point-in-time guard under the existing exclusively owned server/no-other-producer assumption, not an atomic first-input guarantee.
• queue/list can return queued bodies into memory. The harness only tests empty-list/cursor conditions and never prints or inspects items; it is NOT a body-free API and must not be used with unrelated work.
• Fifteen offline tests pass on final head, including fixed independent successful reply, malformed/empty/mismatched submission fields, isolated missing text_elements, nonempty/unproven/error queue snapshots, zero/one submission counts, and response-ID guards on initialization/list/add. Documentation61, generator --check and diff checks PASS. No new native/model/SSH run; historical observations are not retroactively revalidated.

Independent CLI review before push:

• Candidate ad8fe10; identical immutable input SHA256 f46269a528c32b7e6dd6566d1f3c0c6e900c1e19ef89c291fed6b7b1d1ccd1d6 (actual diff, full changed files and pinned schema context).
• Parallel non-interactive Claude Code2.1.300 / requested and confirmed claude-opus-5-5 (04:51:55–04:52:51 UTC), and Codex CLI0.162.1 isolated default/read-only/ephemeral (04:51:56–04:52:19 UTC; actual model not exposed). Both exit0 with valid final responses; no tools, test execution or live delivery by reviewers.
• Both found no mandatory defect in these fixes. Root adopted optional independent wire-fixture, isolated default-field omission and list-stage ID tests. Final probe.py is byte-identical to the reviewed candidate; the last test-only refinements were root-validated, not independently re-reviewed. Exact-final-head CI and automated review are separate gates.
• Unrelated pre-existing loaded-thread metadata/assert hardening and connection-close fault-fixture work are not certified by this change.

The two findings are resolved with implementation/source/test explanations; latest-head re-review is requested. No merge, tag, release, publication, runtime installation or operations change.

Storage-authority correction (P1 on ec924de; 2026-10-15)

Exact final head: 97f7df7.

Read-only design consultation confirmed that probe-local environment/root-config checks cannot establish an already-running server's startup SQLite selection. Root's subsequent pinned-source audit also found an important limitation in the proposed server-history replacement: native history can be unavailable before materialization. The implementation and documentation preserve both facts:

• Remove all direct SQLite imports, home DB globbing/queries and storedRowsBefore freshness claims from the probe. No guessed storage override parser or dependency on unmerged fix(codex): refuse known storage relocation across native config layers #311.
• Bind account home to initialize only as account identity, not startup-storage attestation. Read the same server/thread's metadata with includeTurns:false; require matching ID/cwd, idle/directInput, explicit persistent metadata and empty preview.
• Query thread/turns/list with limit1/sortDirection:desc/itemsView:notLoaded, requiring explicit empty data and null nextCursor/backwardsCursor. Check before inspect output and again after queue/list before the only add.
• Known no-rollout/unmaterialized/unsupported/malformed/error history does not become an empty history. It refuses pre-effect with accepted:false/not_attempted/no retry, no DB fallback, no persist/resume/start workaround. Add-attempt uncertainty still remains unknown/null/no retry.
• Mandatory --allow-server-history-db-init is checked BEFORE connection in both modes. Inspected turns/list paths do not directly request persist_thread, but paginated history can lazily open/create/migrate a history DB. Inspect is non-submitting, NOT globally write-free.
• notLoaded suppresses turn items, but previews/error metadata can still contain text in memory. No raw text is reflected. Exclusively owned test target only; no unrelated history/transcripts.
• Server-visible history is a snapshot, not startup SQLite attestation, atomic freshness or no-input-ever proof. Rollback/compaction, paginated active-overlay limitations and competing producers remain limits. The native rollout-backed handler may replay full history before pagination; limit1 does not bound server work.
• Crucially, pinned native maps a missing rollout to "thread/turns/list is unavailable before first user message". The updated harness safely refuses such a target. This may make the original unsaved-first-input workflow unavailable; synthetic successful empty snapshots do NOT prove native availability. Historical successful sends are not retroactively revalidated. investigate(codex): receive input before the first turn through the owning App Server #306 remains OPEN, and no general first-input product support is claimed.

Pinned primary sources and these availability/privacy/side-effect distinctions are linked in ORDINARY_TUI.md. The fresh-history fallback and explicit unavailable error were independently source-traced, not exercised on a live native server.

Validation on final head:

• Probe suite26 PASS, including rerun under a private task temporary root; documentation61 PASS; generator --check/diff PASS.
• Adversarial stale zero-row DB plus server-visible prior turn refuses; missing/nonempty home DB does not override an empty server snapshot. Fixture root sqlite_home and inherited CODEX_SQLITE_HOME do not cause a local DB read.
• Every invocation mocks/asserts no sqlite3.connect, validates read(false)/notLoaded and forbids thread start/resume or turn start. Both snapshots cover malformed/nonempty/paginated/error history, preview and metadata changes; original queue/error/echo/unknown/no-resend guards remain.
• New source-derived unmaterialized-error regression refuses inspect and queue at both phases without fallback. These are synthetic/offline tests; no new native/model delivery, settings or operations changes.

Independent CLI static review:

• Claude Code2.1.300, requested and metadata-confirmed claude-opus-5-5; safe mode/tools disabled/non-interactive. Codex CLI0.162.1, ignored user config/read-only/ephemeral; actual default model not exposed.
• Initial candidate91db14b10118ae06e0da97dc97239bddeacc0cca input SHA256 b0c2340347e54687229b432a3ed9e2c25a20c561cfc11f78d6d86fb3344d9839. Parallel Claude05:12:00–05:13:06 UTC and Codex05:12:01–05:12:28 UTC; both exit0 and valid completed responses, no mandatory defect.
• Final candidate97f7df702235efa1849aec54f59c68dd9ce3cd46 input SHA256 caefe39b45f131caac61d4f156ce09b85b907676bced68c35f4cbe0ee8bcfb8f includes explicit unmaterialized-history error/source limitations and current files. Parallel Claude05:17:44–05:18:58 UTC and Codex05:17:46–05:18:06 UTC; both exit0, valid final results and no mandatory defect. No reviewers executed tools/tests/native sessions. Source-only review is not native acceptance.
• Recommendations about diagnostic granularity/additional fixtures are not certified as implemented runtime support. Final source was not changed after the exact final-candidate review.

The P1 thread is resolved as a fail-closed safety correction, not by asserting full storage identity or completed empty-TUI support. Current-head CI and Codex automatic re-review are separate gates. No merge/tag/release/publication/native run/runtime install/deployment.

[chatgpt-codex-connector]

Codex Review Summary

This comment shows the latest Codex review activity on this pull request.

Review	Status	Commit	Review trigger
📝 Code Review	✅ Completed 2026-10-16T00:22:43.787150Z	d77a583	Manual request
🔒 Security Review	✅ Completed 2026-10-16T00:22:55.569874Z	d77a583	Manual request

ℹ️ About Codex in GitHub

Your team has set up Codex to review pull requests in this repo. Reviews are triggered when you

• Open a pull request for review
• Mark a draft as ready
• Comment "@codex review" or "@codex security review".

Codex reacts with 👀 while any review is running, comments if it has suggestions, and reacts with 👍 once all reviews finish with no findings.

[chatgpt-codex-connector]

💡 Codex Review

Here are some automated review suggestions for this pull request.

Reviewed commit: 4fdc92139c

ℹ️ About Codex in GitHub

Your team has set up Codex to review pull requests in this repo. Reviews are triggered when you

• Open a pull request for review
• Mark a draft as ready
• Comment "@codex review".

If Codex has suggestions, it will comment; otherwise it will react with 👍.

Codex can also answer questions or update the PR. Try commenting "@codex address that feedback".

[abruption]

@codex review

Please re-review exact current head dc80705. All three original findings have been source-checked, regression-tested and resolved with per-thread explanations. Current-head CI is green; root just reran 7 offline probe tests, 61 documentation tests, generator --check and diff checks successfully. This request does not authorize live native/model tests or operational changes. Historical receipt evidence and remaining production/platform limits are unchanged.

[chatgpt-codex-connector]

💡 Codex Review

Here are some automated review suggestions for this pull request.

Reviewed commit: dc807057db

ℹ️ About Codex in GitHub

Your team has set up Codex to review pull requests in this repo. Reviews are triggered when you

• Open a pull request for review
• Mark a draft as ready
• Comment "@codex review".

If Codex has suggestions, it will comment; otherwise it will react with 👍.

Codex can also answer questions or update the PR. Try commenting "@codex address that feedback".

[abruption]

@codex review

Please re-review latest head 5c80632. Both new findings on dc80705 are addressed and resolved with source/test evidence. Shared RPC error validation requires integer code plus string message; queue mode requires explicit non-ephemeral metadata before database access or submission. Final offline probe tests10/docs61/generator/diff PASS; independent Opus5.5/Codex CLI review provenance and limits are in the PR body. No live native/model test or operations change is requested.

[chatgpt-codex-connector]

💡 Codex Review

Here are some automated review suggestions for this pull request.

Reviewed commit: 5c80632b28

ℹ️ About Codex in GitHub

Your team has set up Codex to review pull requests in this repo. Reviews are triggered when you

• Open a pull request for review
• Mark a draft as ready
• Comment "@codex review".

If Codex has suggestions, it will comment; otherwise it will react with 👍.

Codex can also answer questions or update the PR. Try commenting "@codex address that feedback".

[abruption]

@codex review

Please re-review exact latest head ec924de. Both findings on 5c80632 are addressed: validate the full queued submission and exact request echo before acceptance; require an explicitly complete empty first queue page immediately before the one add. The report preserves TOCTOU/no-other-producer and queued-body privacy limitations. Probe15/docs61/generator/diff pass; independent Opus5.5/Codex CLI review metadata is in the PR. No live native/model input or operations changes are authorized by this review request.

[chatgpt-codex-connector]

💡 Codex Review

Here are some automated review suggestions for this pull request.

Reviewed commit: ec924debf5

ℹ️ About Codex in GitHub

Your team has set up Codex to review pull requests in this repo. Reviews are triggered when you

• Open a pull request for review
• Mark a draft as ready
• Comment "@codex review".

If Codex has suggestions, it will comment; otherwise it will react with 👍.

Codex can also answer questions or update the PR. Try commenting "@codex address that feedback".

[abruption]

@codex review

Please review exact latest head 97f7df7. P1 is addressed by removing local account-home SQL authority, not by guessing server startup env. Server-visible metadata/turn snapshots gate the only add; unavailable history is safely refused. Both modes explicitly opt into possible server history DB initialization/migration. The pinned source can reject turns/list before materialization, so original unsaved-first-input availability is NOT claimed; #306 remains open. Final26 offline tests/docs61/generator/diff pass and exact-head independent Opus5.5/Codex CLI provenance is in the PR. Do not run live native/model sessions or change operations.

[chatgpt-codex-connector]

Codex Review: Didn't find any major issues. Can't wait for the next one!

Reviewed commit: 97f7df7022

ℹ️ About Codex in GitHub

Your team has set up Codex to review pull requests in this repo. Reviews are triggered when you

• Open a pull request for review
• Mark a draft as ready
• Comment "@codex review".

If Codex has suggestions, it will comment; otherwise it will react with 👍.

Codex can also answer questions or update the PR. Try commenting "@codex address that feedback".

[abruption]

@codex security review

Please review exact latest head 97f7df7 before merge readiness. The old security result predates the current experimental server-history guard and opt-in. Static code/security review only; no native/model sessions, operations, settings changes or publication.

[chatgpt-codex-connector]

Codex Review: Didn't find any major issues. Can't wait for the next one!

Reviewed commit: 97f7df7022

ℹ️ About Codex in GitHub

Your team has set up Codex to review pull requests in this repo. Reviews are triggered when you

• Open a pull request for review
• Mark a draft as ready
• Comment "@codex review".

If Codex has suggestions, it will comment; otherwise it will react with 👍.

Codex can also answer questions or update the PR. Try commenting "@codex address that feedback".

[abruption]

@codex security review

[chatgpt-codex-connector]

🛡️ Codex Security Review

Security review completed. No security issues were found in this pull request.

Reviewed commit: 97f7df7022

View security finding report

Only the user who started this review can view the report in Codex.

ℹ️ About Codex security reviews in GitHub

This is an experimental Codex feature. Security reviews are triggered when:

• You comment "@codex security review"
• A regular code review gets triggered (for example, "@codex review" or when a PR is opened), and you’re opted in so security review runs alongside code review

Once complete, Codex will leave suggestions, or a comment if no findings are found.

[abruption]

@codex review

[abruption]

@codex security review

[chatgpt-codex-connector]

Codex Review: Didn't find any major issues. 🎉

Reviewed commit: d77a583b26

ℹ️ About Codex in GitHub

Your team has set up Codex to review pull requests in this repo. Reviews are triggered when you

• Open a pull request for review
• Mark a draft as ready
• Comment "@codex review".

If Codex has suggestions, it will comment; otherwise it will react with 👍.

Codex can also answer questions or update the PR. Try commenting "@codex address that feedback".

[chatgpt-codex-connector]

🛡️ Codex Security Review

Security review completed. No security issues were found in this pull request.

Reviewed commit: d77a583b26

View security finding report

Only the user who started this review can view the report in Codex.

ℹ️ About Codex security reviews in GitHub

This is an experimental Codex feature. Security reviews are triggered when:

• You comment "@codex security review"
• A regular code review gets triggered (for example, "@codex review" or when a PR is opened), and you’re opted in so security review runs alongside code review

Once complete, Codex will leave suggestions, or a comment if no findings are found.