# DSH 0.2.0-rc.1 compatibility

## Evidence

- `pnpm run test`: 521 passed, 0 failed, 0 skipped.
- `pnpm run test:behavior`: 406 passed, 0 failed, 0 skipped.
- `pnpm run test:delivery`: 87 passed, 0 failed, 0 skipped.
- `node --test tests/powershell-manager.test.mjs`: 28 passed, 0 failed, 0 skipped; the user PATH value matched before and after the run.
- `pnpm run build` completed; a subsequent build left `git diff --exit-code -- lib` clean.
- The 2.2.5 package was packed as `dsh-codex-subscription-2.2.5.tgz`.
- Official DSH `0.2.0-rc.1` acceptance passed: the candidate installed once, composed once, started Web and passed the readiness probe, was removed, and was reinstalled. The acceptance behavior run reported 406 passed and 0 skipped.
- Official DSH `0.1.7-rc.2` acceptance passed: published plugin 2.2.4 was installed first, then the candidate installed once, composed once, started Web and passed the readiness probe, was removed, and was reinstalled. The acceptance behavior run reported 406 passed and 0 skipped.

# DSH 0.2.0-rc.2 compatibility (plugin 2.2.10)

DSH 0.2.0-rc.2 was published on 2026-09-29. Two things stopped the plugin from working with it, and both were fixed in 2.2.10.

## What changed upstream

- The plugin manager rejects a plugin whose exact-version peer ranges do not list the host, so 2.2.9 could not be installed on rc.2 without `--accept-risk`. The peer ranges now include `0.2.0-rc.2`, including the optional Codex subtask component.
- rc.2 moves `@earendil-works/pi-ai` from 0.85.1 to 0.87.1. Its Codex request builder now reads the system prompt and tool declarations from the transcript instead of the context object. The plugin's audited-runtime gate refused 0.87.1 until it was reviewed.

## Evidence

- Diff of `openai-codex-responses` between pi-ai 0.85.1 and 0.87.1: the exported API is unchanged; only how the request body is built from the transcript changed.
- `pnpm test`: 537 passed, 0 failed.
- Real ChatGPT account on DSH 0.2.0-rc.2 with pi-ai 0.87.1, installed through the normal path (no `--accept-risk`):
  - plain chat, then a shell tool call and a follow-up that recalls an earlier fact: pass;
  - WebSocket mode: one connection, no failures, context kept across turns;
  - cloud compaction: a 194K-token message shrank to about 6K tokens and the fact stated in it was still answered.

## Not checked

- The optional Codex subtask component on rc.2 (the pinned component version moves to 0.2.0-rc.2 but was not installed or run).
- Image generation, the sketch canvas and quota reset flows on rc.2.

## Official desktop app acceptance (Windows x64, 0.2.0-rc.2, 2026-09-30)

Real ChatGPT account, isolated data directory, plugin installed as `dsh-codex-subscription@2.2.10` from the plugin dialog.

| Check | Result |
| --- | --- |
| Plugin installs by exact version, survives restart, "Codex 订阅" settings page appears | pass |
| Bare-name install picks 2.2.4 (24h release-age rule) and is rejected as incompatible | as documented in README |
| Device-code / browser login completes; account and quota show | pass |
| Model list is the account catalog (8 models); hiding two shrinks the count to 6/8 and the picker drops them | pass |
| Chat with `GPT-5.6-Luna`: shell tool call then answer | pass |

Not checked on the desktop build: WebSocket, cloud compaction, Codex subtasks, images and sketch (covered on the 0.2.0-rc.2 web host and by tests).

## 2.3.0 candidate on the official desktop app (2026-09-30)

The packed 2.3.0 candidate (installed from a local folder) on the official Windows desktop app, real ChatGPT account, isolated data directory.

| Check | Result |
| --- | --- |
| Plugin loads, sign-in, account and quota | pass |
| Models and search are separate cards; catalog status, 9 of 9 models listed | pass |
| Speed menu on GPT-6-Astra shows Standard and Fast only, no Ultrafast (this account's catalog has none) | pass, as designed |
| Chat plus image generation on GPT-6-Astra with Fast selected: 17+25 answered, 1254x1254 image generated, turn completed | pass |
| The short diagnostics report is produced; the full JSON is still in the details | pass |
| Ultrafast on the wire, and its menu entry | not run: no Pro 500 account; covered by catalog and transport tests only, so it is Beta |
| The `ultra` reasoning level advertised for several models | not adapted: DSH has no such level; documented in the guide |