Skip to content

LOPI_KTB3_ATTENDED_RUNBOOK.md

Latest commit

 

History

History
38 lines (31 loc) · 1.81 KB

LOPI_KTB3_ATTENDED_RUNBOOK.md

File metadata and controls

38 lines (31 loc) · 1.81 KB
Raw

KT-B3 attended runbook — real Claude Desktop, real render check

This is not a Claude Code session prompt on its own. This is the checklist a human executes, on their own machine, once a real .mcpb exists. Nothing here can be delegated to a sandboxed session.

Before you start

Confirm a real build exists: check the Actions tab for mcpb-release.yml's latest run, green not red, and that it post-dates bfe4d7bb (the timeout → perl -e 'alarm N; exec @ARGV' fix). Download the .mcpb artifact from that run.

Steps

  1. Install. Drag the .mcpb onto Claude Desktop, or double-click it. Review the install dialog like a normal user would, not like a developer who already trusts the source.
  2. Confirm the tools are visible. All eight tools should show up, including lopi_get_stack_status (new since MCPB-App-1), wherever Claude Desktop lists connected MCP tools.
  3. Trigger the widget. Submit a real lopi task if nothing's running, then ask Claude to check on it. Watch what happens.
  4. Pass looks like: an actual rendered panel, not a text block, showing real task/branch/stage data. Cross-check against lopi dock for the same task to confirm the data is correct, not just that something rendered.
  5. Fail looks like: the tool call completes, Claude describes the result in text, no panel appears, silent, no error. Capture Claude Desktop's version, any developer/MCP logs it exposes, the exact prompt used, and whether step 2's tool list was correct, that narrows down whether the failure is tool discovery or specifically the UI handshake.

Either way

Write the LEDGER.md entry. Pass: log the Claude Desktop version tested against. Fail: log everything from step 5, a real negative result is a complete, useful outcome here, not a failure to route around.