|
| 1 | +# wp4 execution record |
| 2 | + |
| 3 | +Terminal outcome for this unit. PR #4510 merged into `dev` as |
| 4 | +`d7c7b493bfc8e9248b8bb98c692203d74c2ee6cc` on 2026-09-13T13:23:25Z. |
| 5 | + |
| 6 | +## 1. Live three-host proof |
| 7 | + |
| 8 | +All three runs were made at `c8c1c1a9cc`. Only `tests/` and `devlog/` changed between |
| 9 | +that commit and the merged head, so the runtime proven here is byte-identical to the one |
| 10 | +CI was green on. |
| 11 | + |
| 12 | +| platform | host | root pid before -> after | corroboration | |
| 13 | +|---|---|---|---| |
| 14 | +| Linux | `lidge` | 3425189 -> 3431091 | reparented to init, so the relaunch came from the detached `setsid` spawn and not a human; a live `--type=gpu-process --ozone-platform=x11` child confirms the app reached the compositor | |
| 15 | +| macOS | `macmini-cf` | 74114 -> 99260 | bundled app-server respawned under the new root; singleton lock released | |
| 16 | +| Windows | `mini` | 29132 -> 23136 | `{"attempted":true,"stopped":[29132],"surviving":[],"relaunch":"started"}` matching the observed tree | |
| 17 | + |
| 18 | +The Linux corroboration carries the most weight. `001` §2.2 measured the root zeroing its |
| 19 | +own environment block, so the session variables must come from a child. Had that been |
| 20 | +wrong, the pid would still have changed and the command would still have reported |
| 21 | +success, while the user got no visible app. The GPU process is what separates those two |
| 22 | +outcomes, and a pid table alone cannot. |
| 23 | + |
| 24 | +## 2. What running it found that ten audit rounds did not |
| 25 | + |
| 26 | +Five work phases and ten audit rounds found eighteen blockers by reading code. Windows |
| 27 | +held two more, and both surfaced within seconds of a real host. |
| 28 | + |
| 29 | +**A false success.** The ladder returned |
| 30 | +`{"stopped":[27788],"surviving":[],"relaunch":"started"}` while the app kept its original |
| 31 | +pid **and** start time throughout. Two causes in one helper: a stop was claimed on |
| 32 | +pid-liveness alone, and `stillSameProcess` returned a boolean over three distinct |
| 33 | +situations, so a re-probe that merely **failed** was read as "already exited". Reporting a |
| 34 | +restart that did not happen is worse than the stale picker this unit exists to fix. |
| 35 | + |
| 36 | +**Then its mirror image.** Requiring the enumeration to confirm exposed the opposite |
| 37 | +defect: `Win32_Process` lags after a kill, so the single confirming query still listed a |
| 38 | +process that was already dead. The ladder declared `targets_survived`, skipped the |
| 39 | +relaunch, and left the host with the app killed and never restarted. It was restored |
| 40 | +immediately. |
| 41 | + |
| 42 | +Both came from treating one weak reading as proof. Confirmation now polls until the |
| 43 | +platform's own process list agrees, with a final look after the deadline, and a probe that |
| 44 | +cannot run keeps the loop going rather than deciding either way. The kill and relaunch |
| 45 | +primitives were correct on Windows the whole time; only the confirmation was wrong, in |
| 46 | +both directions. |
| 47 | + |
| 48 | +The lesson generalises past this unit: a design that says "fails closed" is not evidence |
| 49 | +that the code does, and neither is a green focused test whose double models the world |
| 50 | +more simply than the world behaves. The doubles here modelled exit purely through |
| 51 | +`isAlive` and kept listing terminated processes, which is precisely why no amount of |
| 52 | +review could surface either defect. They now drop a process from the enumeration once |
| 53 | +liveness reports it dead. |
| 54 | + |
| 55 | +## 3. What hosted CI found that local runs could not |
| 56 | + |
| 57 | +- The macOS cases pointed at `/Applications/ChatGPT.app`, and discovery resolves the |
| 58 | + bundle through `realpathSync`, which touches the real filesystem and cannot be |
| 59 | + intercepted by the exec seam. They passed locally only because that machine has Codex |
| 60 | + installed. They now build a real bundle under a temp directory, realpathed there so the |
| 61 | + fixture and the adapter agree across the `/var` symlink. |
| 62 | +- `privacy:scan` found a second user's home path in two files of this unit. |
| 63 | + |
| 64 | +Neither was visible locally: the first because of what the machine happened to have |
| 65 | +installed, the second because the scan was never run there. |
| 66 | + |
| 67 | +## 4. Gate status at the merged head |
| 68 | + |
| 69 | +- Hosted CI at `156e10776b`: 24 pass, 0 fail, 2 skipping. |
| 70 | +- Local product suite, build and typecheck: **NOT RUN**, per `000` §2. |
| 71 | +- Focused desktop-restart files: 50 pass / 0 fail. |
| 72 | + |
| 73 | +## 5. Carried forward |
| 74 | + |
| 75 | +- **Windows app-server exclusion is a no-op.** The probe enumerates `ChatGPT.exe` while |
| 76 | + Windows app-servers run as `codex.exe` / `codex-code-mode-host`, so they still receive |
| 77 | + SIGTERM before the app quits. The restart is correct; the cost is one extra interrupted |
| 78 | + turn. Widening that query changes what may be killed, so it needs its own verification. |
| 79 | + Documented at the decision point in `src/cli/restart-scope.ts`. |
| 80 | +- **`ocx system codex-restart` refuses rather than handing off** when the proxy itself runs |
| 81 | + inside the Codex app, because a proxy never exits and the handoff waits for the caller. |
| 82 | +- **The local-macOS handoff path is unproven on a live host.** It is implemented and |
| 83 | + unit-tested, and proving it ends the session that issues it, so it is deliberately the |
| 84 | + last action rather than an omission. |
0 commit comments