Skip to content

samargt/auto-bean-review-probe

v0.1.0

Synthetic local MCP panel compatibility check; no ledger access.

GitHub

Local review UI compatibility probe

This is milestone 1 of the import review UI plan. It is an isolated synthetic plugin, not the auto-bean product UI. It never opens a ledger, changes import schemas, or changes Harbor grading. The product implementation must wait until the desktop checks below pass.

Run the protocol check

From the repository root:

UV_CACHE_DIR=/tmp/auto-bean-ui-uv-cache uv run --frozen --project experiments/review-ui-probe python experiments/review-ui-probe/check.py

This starts the official Python MCP SDK server over stdio and verifies tool discovery, HTML resource retrieval, a structured synthetic session, persisted idempotent actions, and rejection of an invalid action ID. It uses an isolated temporary directory and cleans its test data on exit. It does not prove that the desktop renderer or conversation bridge works.

The small HTML view uses the documented MCP Apps JSON-RPC bridge directly for the compatibility test. The full product UI remains planned in React/TypeScript with the MCP Apps SDK. The lockfile pins the probe's Python dependencies.

Installed test package

On 2026-10-21 an isolated copy of the package was installed as auto-bean-review-probe@auto-bean-ui-probe. Its temporary marketplace source is .cache/review-ui-probe-marketplace. It is separate from auto-bean@auto-bean. The package copy contains only the manifest, MCP configuration, lockfile, Python project, server, and HTML view; it excludes the development virtualenv.

Codex CLI 0.163.0 confirmed through plugin/read that the plugin is installed, enabled, and declares the review-probe MCP server. After a tool refresh, the current conversation exposed open_review_probe and opened its panel.

Required desktop check

  1. Select Auto Bean Review Probe in the existing conversation's plugin picker. Refresh the plugin list if necessary.
  2. Open Import Review Probe from the conversation's panel controls or call its open_review_probe tool.
  3. Confirm the view displays Host bridge connected.
  4. Choose Read synthetic session. Confirm example.csv and account_review.
  5. Choose Save test action. Confirm one synthetic action is persisted.
  6. Choose Continue this conversation. Verify the message arrives in the originating conversation and the agent can resume.
  7. Reopen the view and read again to confirm saved actions survive the view.

The continuation message contains only a synthetic action ID and test purpose. No financial data is read or transmitted. Saving alone never sends a message; the separate continuation button is an explicit test action.

Evidence and completion boundary

Verified on 2026-10-21:

  • Real stdio discovery, resource read, structured read, persistence, idempotency, and invalid-input rejection passed.
  • Installed plugin metadata reports enabled state and the local MCP server.
  • Existing product and Harbor host tests passed: 89 tests.
  • Root Mypy passed: 31 source files; the two standalone probe files also passed a separate check using the probe virtualenv as the Python executable.
  • The actual desktop host rendered the panel. Its initialization reported hostInfo.name: chatgpt, version 26.1016.51333, desktop platform, and serverTools, serverResources, and message capabilities.
  • The panel's Read synthetic session button fetched example.csv in account_review through the host bridge.
  • Save test action persisted da987fcc-6568-4978-8dbe-f19facb2f9c6; a subsequent model-side open_review_probe call returned that action.
  • Continue this conversation received a successful host response. The corresponding MCP app message arrived in the same conversation with the matching saved action ID; implementation resumed successfully.

Not yet verified:

  • State restoration after closing and reopening the panel.

The computer-use tool explicitly refuses to control com.openai.codex: Computer Use is not allowed to use the app 'com.openai.codex' for safety reasons. After tool refresh, the tool call opened the panel and the permitted MCP Apps browser surface could inspect it and click its controls. Native app control was not needed. Local rendering and bridge calls are supported on this build.

The full parsed-v2, staged approval, panel, skills, and Harbor changes are not implemented. Do not describe this probe or the existing 89 passing tests as acceptance of that feature.

References