Skip to content

Historical engineering record

This page preserves the design, plan or acceptance evidence for its recorded implementation stage. Earlier behavior, proposed capabilities and old deployment settings do not supersede the current Workspace guide or release history.

Codex-Centric Resume v1: macOS native acceptance ​

Historical acceptance — applies to the 2026-10-20 Resume v1 manual-confirmation implementation (source package 3.0.1, Codex CLI 0.162.1). That flow predates automatic binding at 5fad370 and confirmation-free readiness at f9f0210. /status and explicit confirmation below describe what was tested then, not steps required by the current recovery guide. These results do not certify the later readiness policy.

Date: 2026-10-20. Platform: macOS arm64, Node 24.17.0, native node-pty; Codex CLI 0.162.1; Antigravity through the user's existing agy-proxy wrapper. Real user model, authentication, environment and native approval settings were retained. All implementation files used for the test lived in a disposable Git project. No npm publication, version change or Release was performed.

A — continuing native conversation ​

Task: workspace-82fd905c20a64a1e8ac093f23b40d9a5. Codex native Conversation ID: 01a156e9-1b2d-7eb0-9b36-09dda6d253ef.

A real Default Collaboration Task initialized both CLIs. Codex was asked to remember amber lantern 739182 only in its conversation and replied ACK-739182. A separate followup asked for the first word; it replied amber. The browser closed the Task and clicked Resume Conversation. The resumed TUI's native /status showed the same Conversation UUID and canonical project directory; the operator explicitly confirmed that live result in the Web UI.

RuntimeBefore CloseAfter Resume
Codexsession_a86db2f2587245e5b39ec874370bf07csession_18b91cfa542a4268b0934d4ec85bb63b
Antigravitysession_dbd0bdda650d4a0486f69a4c6d720324session_8fc45902ca6c40039919becc1296d046

After reopening, the question was: "What complete phrase did I ask you to remember earlier? Reply with that phrase only. No tools." The response was amber lantern 739182. The question did not contain the phrase. The application sent no previous prompt or history to the new Codex process. Native launch used the explicit UUID. This establishes real context continuity, beyond process startup.

B — rebuilt executor completes a new implementation ​

After executor replacement and state inspection, Codex explicitly dispatched a new instruction to update resume-proof.txt to executor joined 42018 with one newline. Native tool approval dialogs were handled in the executor terminal. The completed round was round-a8383c2d-2f36-437e-b6dc-a85886a6e31d:

RevisionFormal state
7SENT
8WORKING
9DONE
10REVIEWING
11APPROVED

Codex inspected the actual file and reported approval. A separate filesystem read confirmed the exact 22 bytes executor joined 42018\n. Existing unrelated untracked content was retained; neither agent committed, pushed or released test changes.

One earlier attempt exposed normal tool-approval behavior: Codex tried to answer a native approval using send, creating a different round. The old DONE was refused; no false approval was recorded. That executor was replaced, and subsequent instructions explicitly left native approvals to the operator. Resume neither answers native approval dialogs nor changes the user's permission configuration.

C — interruption in WORKING ​

The real executor reported WORKING for round-d4e80ba2-c0b3-4051-94e9-ccbacb248924, revision 6, then was closed. Resume retained the same Codex Runtime, replaced the executor and recorded the interrupted WORKING round for inspection. Before/after cooperation and Timeline JSON were equal. No work was automatically dispatched. A call replaying the old executor's actual Session/generation claims was rejected without a new event. Only a subsequent explicit Codex dispatch started the successful round above.

D/E — automated boundary and regression evidence ​

Deterministic native-record tests and real Session Host fixtures cover wrong UUID, wrong project/source, canonical Windows aliases, duplicate Resume, pending-startup adoption, uncertain dispatch, late old-runtime writes, old-round executor reports, partial failures, login continuation, archived/custom refusal, HTTP capability and project boundaries, and stale UI completion. Existing Web/shared suites cover Custom Terminals, Timeline, Health, notifications, Git Changes, PTY and adapters. The npm artifact is packed and installed in an isolated consumer/home; omission checks enforce both new runtime modules. These tests are separate from real model acceptance. Final counts and CI results are recorded on PR #139.

Capability boundary ​

This is manual live confirmation, not automatic native identity attestation. Local native metadata only preflights the exact target/project/source. Each new Codex runtime requires a fresh operator /status check; its runtime identity and original native Conversation ID remain separate. Arbitrary user wrapper internals remain trusted configuration and unsupported resume arguments fail closed. Windows compatibility tests do not establish Windows 10 real native CLI acceptance.

Local coordination. Explicit permissions. Independent verification.