Correct the stage-dispatch check, and propose closing the gap it names

[VeryComplexAndLongName]

The instruction was outdated

Task 6.6 told the person doing it to set a stage to:

{ "agent": "claude-cli", "dispatch": "vscode-chat" }

That is the shape harness-config-strictness retired — chat dispatch became an agent id, and naming claude-cli beside it was the pointless entry that change removed.

The old shape still loads: it is migrated with a warning, deliberately, because this repository's own files used it. So following the original text was not broken — it exercised the migration path rather than the current one. Both reach the same dispatch; only one is what a reader should copy. It now reads { "agent": "vscode-chat" }.

The rest of 6.6 stays open, for a good reason

Reported precisely by the person attempting it:

The Extension Development Host started. Live integration suite: 10 passing. But the existing suite does not send a real webview command with the chat dispatch into the private AiPanel.dispatchToChat, so it does not confirm the Chat handoff or the panel state 6.6 asks for.

Correct. ExtensionTestApi exposes getRunners, runController, optionalServer, getDashboardContext and two tree providers — not the panel. Nothing in the suite can deliver the one message that reaches the dispatch.

So a shipped code path's only evidence is somebody remembering to look at a window, and that evidence lands nowhere a later reader can check. ADR 0016's started → handedOff — never completed — is a rule about the event stream: exactly what a test is good at, and what a person watching a chat window is not.

dispatch-to-chat-integration-coverage

The narrow fix: a way to deliver a webview command through the same handler a real message reaches, plus an integration case asserting that contract.

Explicitly not exposing AiPanel or dispatchToChat — a test that can call the private method tests this implementation instead of the contract. And explicitly not asserting anything about a visible window: the command that opens it can be observed; what appears on screen stays with the person.

What remains human afterwards is smaller and honest: that the chat window opens with the prompt in it.

Task 4.4 requires confirming the new case fails when the dispatch is broken — by breaking it, not by reasoning about it. Three tests found this week passed whether or not the feature worked.

Test plan

• openspec change validate --strict — both changes valid
• npm run lint:english — passed
• Documentation only; no source changes
• CI green

🤖 Generated with Claude Code