Dashboard: pick a per-task agent CLI in the n new-task flow

[tildesrc]

The per-task agent_cli override has had a full data path since #384/#385 (model, store, service, REST, client, MCP) but no way to set it interactively — the dashboard's n flow went repo → workflow → memo and always inherited the repo default.

The memo modal now carries an agent CLI dropdown that starts on the repo's own default. Leaving it alone sends agent_cli=None rather than a redundant explicit value, matching resolve_agent_cli's contract, so task creation stays repo → workflow → type → Enter and the CLI is there when you want it.

What's here

• KNOWN_AGENT_CLIS in core/models.py. The CLI name set is a control-plane concept (it drives the base-image variant + config mount), and terminal/taskservice can't import container/cli at runtime without breaking the determinism invariant — so the list is restated in the LLM-free core and guarded by a test.
• registered_agent_clis() on the adapter registry. _REGISTRY is populated lazily by _load_builtin_adapters(), so a drift test comparing against it directly would have passed vacuously or depended on import order. The new guard also pins the two long-duplicated DEFAULT_AGENT_CLI constants (core/models.py and container/cli/base.py) together — a pre-existing hole.
• MemoScreen takes agent_clis + repo_default and dismisses a 4-tuple whose added member is the picked CLI. Both params default to empty, so the screen still stands alone with no dropdown.
• Focus handling. Tab reaches the dropdown (TextArea's default tab_behavior is "focus"); picking hands focus back to the memo so the next Enter submits rather than reopening the overlay. That hand-back is deliberately unconditional — while the overlay is open, focus sits on the overlay rather than the Select, so a has_focus guard never fires on the path that matters.
• _selected_cli() falls back to the first known CLI when a repo names one this build doesn't offer, since allow_blank=False needs a value that's genuinely in the options.
• Repo default read from the payload in hand. action_new_task already fetches the repo dicts and threw everything but the id away; it now keeps the agent_cli too. self._repo_clis was the tempting source but it's only filled on a refresh pass, so it can be empty or stale at this point.
• Four hardcoded "claude" literals in dashboard.py swapped for DEFAULT_AGENT_CLI.

The fake-adapter registration test now cleans up after itself — it was leaking fake-cli into the module-global registry, which would have made the drift guard order-dependent.

Dropdown coverage: the repo default is shown and sent as no-override, an explicit override, every known CLI is on offer, an unknown one is rejected, an unofferable repo default degrades sanely, and the full tab → pick → focus-returns interaction.

Out of scope: a post-creation set_agent_cli operation, and validating Repo.agent_cli against KNOWN_AGENT_CLIS in the repo form.

Plan: the task's plan.md artifact.

🤖 Generated with Claude Code