Start Run with Agentic Harness on the change it was opened for

[VeryComplexAndLongName]

Two commits: the archive move that was blocking npm run lint, and the prefill itself.

The prefill — one defect, three symptoms

Right-clicking a change and choosing Run with Agentic Harness opened a panel with nothing filled in: command list, no change selected, default agent.

Everything needed was already there. runWithHarness passes the change:

revealAiPanel({ ...dashboardContext(workspaceRoot, item.changeDir), startChain })

and AiPanel takes changeDir as a prop. It simply started from useState<CommandKind>("list") and useState<string>("") and never consulted it.

The agent is the same fact one step further on. AiPanel already pre-selects from stepAgents, keyed on the command kind through COMMAND_KIND_TO_HARNESS_STAGE. With the kind stuck at list there is no stage, so no entry is looked up and the picker keeps DEFAULT_AGENT_ID.

Seeding the kind fixes the agent as a consequence — the configured agent was never unavailable, only unreachable.

What the test caught

Seeding the selection alone was not enough:

expected '' to be 'demo'

A <select> cannot hold a value it has no <option> for, so the seeded change silently fell back to empty until a list had run. The option list is seeded too now, and the existing reconciling effect replaces it with the real list while keeping the selection.

Nothing is enforced — every seeded value stays editable, the same way the stepAgents recommendation already was. The chain path is untouched: runChange is set only for the picker target, since HarnessChainPanel has one button and nothing to pre-select.

The archive move

dispatch-to-chat-integration-coverage was implemented and its directory moved to archive/ on disk, but the deletion of the original path was left unstaged — so the old files stayed git-tracked while missing from the working tree, and check-english.mjs died with ENOENT before eslint ever ran.

That is the same failure that made npm run lint expected-red for days earlier this week, from the same cause: an archive move committed halfway. Staged both halves; 31/31 tasks complete, archive copy verified intact.

Staging it also surfaced a real no-useless-escape error in my own new test, which had been hidden behind that ENOENT — the precise reason this matters.

Test plan

• npm run typecheck — clean
• npm run lint — clean (0 errors), once the archive move was staged
• AiPanel 43/43, extension-context 7/7
• openspec change validate --strict — valid
• Only failing suite is git.push.test.ts, tracked as core-test-worker-contention
• CI green
• Human: right-click a change → Run with Agentic Harness → confirm the change, implement and the configured agent are already selected, and that changing any of them still works

🤖 Generated with Claude Code

[VeryComplexAndLongName]

Superseded by a branch rebased onto main. The original branch was cut from a sibling branch whose commit later landed on main as #200 (squashed), so the same content appeared twice and the pull request conflicted before CI could start.