E2E test: headless smoke recipe on the sandbox via fleet run #5232

Maintainer setup

None beyond #5230 (MINIMAX_API_KEY, E2E_SANDBOX_TOKEN, sandbox repository). No Anthropic credential.

Depends on: #5230.

What it proves

The headless path works end to end on MiniMax: autoskillit fleet run starts an orchestrator, which runs one skill session that edits the sandbox, runs its tests, pushes a branch, opens a PR and closes it.

Why a new recipe

The existing project-local smoke-test recipe is bound to this repository: it edits src/autoskillit/smoke_utils.py (recipe), runs task install-worktree and task test-check, and opens and closes a PR on its source repository. A sandbox variant must edit a sandbox file and run the sandbox's own test task.

Cost

Two concurrent sessions — the orchestrator and one smoke-task session — and no subagents (smoke-task declares none). Suitable for routine selection.

Scope

• Add the sandbox smoke recipe (in the sandbox repository's .autoskillit/recipes/ or as a harness fixture).
• Catalog entry with trigger paths covering the headless launch, MCP server, hooks and fleet dispatch code.
• Assertions: envelope success; a branch and a closed PR exist on the sandbox; the remote branch is deleted afterwards.

Each defect found gets its own bug issue; the catalog entry is marked expected-failure with the link until fixed.

Acceptance criteria

• Passes in CI on MiniMax with no Anthropic credential.
• Leaves no open PR or stray branch on the sandbox, on success or failure.