feat(models): pick a profile interactively when switch is given none

[daoiqi]

Closes #919.

What

teamai models switch required a profile name, so a member who did not
remember the exact id had to run models list first, read the ids, then
retype one.

The profile is now optional. With none given, switch lists every profile —
team ones first, then personal, the order models list already uses — and
switches the one picked:

$ teamai models switch

  1. team:tokenhub — Tencent TokenHub
  2. local:mine — My gateway (personal)

Select a profile [1-2, or "none" to cancel]: 1

Behaviour

• Single pick. switch points each agent at exactly one gateway (Claude's
  settings.json holds a single ANTHROPIC_BASE_URL; the manifest records one
  profile per agent), so an answer naming several has nothing meaningful to do
  with the extras. Rather than silently take the first, the prompt says why and
  asks again:

  Select a profile [1-2, or "none" to cancel]: 1,2
  ⚠ switch takes one profile; you named 2. Enter a single number.
  Select a profile [1-2, or "none" to cancel]: 1
  

  An unusable answer is re-asked the same way. This is deliberately unlike
  init / push, where selecting several items is meaningful and the prompt
  advertises the syntax.

• none (or 0) cancels: prints Cancelled, changes nothing, exits 0.

• Non-interactive runs still require the profile. With no terminal there is
  nothing to pick from, so the command fails and names the argument rather than
  guessing — a hook or CI run must never land an agent on a gateway by accident.
  The one-profile case is not auto-selected, for the same reason.

• Naming a profile behaves exactly as before; models switch <profile> is
  unchanged and backward compatible.

Scope is switch only — models add/configure/remove keep their required
arguments. No new subcommand: [profile] also matches models list [profile],
and adding an optional argument needs no second entry point for one action.

Implementation notes

• chooseProfile resolves the profile (named, or picked) and findProfile now
  also accepts an already-resolved ProfileRef, so both paths run through one
  code path — the team identity stamping, key loading and the legacy-values
  checks stay identical whichever way the profile arrived.
• The prompt loops on askQuestion + parseSelection rather than using
  askSelection: the latter returns null for both "cancelled" and "unusable
  answer", and the two need different handling now that one of them re-prompts.
• models-cmd.test.ts gained vi.clearAllMocks() in beforeEach, because the
  new cases assert on log.warn call counts and would otherwise see the
  previous test's calls. Existing cases are unaffected.

Verification

npx tsc --noEmit and npm run lint clean; full suite 5496 passed. Three
push-env.test.ts cases fail, and I confirmed they fail identically on the base
commit (3a9a24a) without this change — pre-existing and unrelated.

Built CLI driven through a real pty (isInteractive() is false under a pipe, so
the picker needs one):

#	Case	Result
1	no profile, pick 2	lists both, switched claude switched to local:mirror
2	pick 1,2, then 1	rejected with the reason, re-asked; the second answer decides
3	pick banana, then 9, then 2	both re-asked; the third answer decides
4	pick none	ℹ Cancelled, settings untouched
5	no terminal	refuses, names teamai models switch <profile>, exit 1
6	explicit profile	unchanged behaviour

--dry-run writes nothing in either form (manifest and settings.json
unchanged after a preview).

Docs updated in both languages (docs/usage-guide.md,
docs/usage-guide.zh-CN.md); skill-data/core/references/commands.md
regenerated.

[github-actions]

No findings.

The PR description documents sufficient testing, including representative real-CLI/PTY verification for the runtime behavior change. The previously failing tests are clearly identified as pre-existing.

[jeff-r2026]

Please resolve the conflicts

[daoiqi]

Merged origin/main (up to bae48e5) to clear the conflict.

Conflict: src/models-cmd.ts, one import line. #880 extracted the piped-stdin reader into utils/prompt.ts as readStdin and dropped the inline loop from readSecretStdin, while this branch still imports askQuestion/isInteractive/parseSelection for the picker. Both are used, so the resolution is the union:

import { askConfirmation, askQuestion, askSecret, isInteractive, parseSelection, readStdin } from './utils/prompt.js';

readSecretStdin keeps main's readStdin(); chooseProfile keeps this branch's loop. The rest of the file auto-merged — main's findProfile(reference: string) and this branch's ProfileRef-accepting version are the same signature this branch already introduced.

Re-verified after the merge: npx tsc --noEmit clean; models-cmd model-switch commands-reference 88 passed. No behaviour change from the merge itself, so the pty table above stands.

[github-actions]

No findings.

The earlier “No findings” result remains valid. The PR description documents sufficient testing, including representative real-CLI/PTY verification for the runtime behavior change and identification of unrelated pre-existing failures.

[github-actions]

No findings.

The earlier “No findings” reviews remain valid. The PR description documents sufficient testing, including representative real-CLI/PTY verification for the runtime behavior change and clearly identified pre-existing test failures.