feat(tui): the CLIs turn on the arrows, and a step shows only its own question

Three things the walk was still carrying from when it was one sheet.

The CLIs turned on tab, because the sheet they were on was also reading a list
with the arrows. It is not that sheet any more: the first step asks one thing, so
up and down are the accounts under the tab and left and right are the tabs. A
chord for what a row of tabs is for is a key somebody has to already know.

The step that asks which model was reading the account back on its tuning line.
That step cannot change it -- it was the step before -- and a setting shown where
it cannot be changed is a setting somebody tries to change there. What the walk
has settled is on the line above the prompt when it is over. Where the agent works
stays, because for an isolated agent it is settled by the flow and there is no
step of its own that will say it.

And the row for the account nobody chose was a sentence. It is a row in a list of
accounts, beside `deepseek` and `work`, so it is two words: `as installed`.

Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>

Author: futrime

docs/tui.md

@@ -242,20 +242,21 @@ accounts, since an [account](providers.md) is one backend's — what signs in to
 not what signs in to codex:
 
 ```
-   claude · codex · kimi · mimo · opencode · pi   tab/shift+tab to switch
+   claude · codex · kimi · mimo · opencode · pi   ←/→ to switch
 
-   ❯ 1. as this machine is signed in ✔ nothing is redirected
+   ❯ 1. as installed              ✔ signed in as you signed it in
      2. deepseek                  gateway · ANTHROPIC_AUTH_TOKEN, ANTHROPIC_BASE_URL
      3. work                      login
 ```
 
-**tab** turns to the next CLI, **shift+tab** to the one before, and they wrap. One CLI on its
-own is a heading rather than a row of tabs: there is nowhere to switch to, so nothing says
-there is. Where the flow said that agent has to run a particular moment, only the CLIs that
+**→** turns to the next CLI, **←** to the one before, and they wrap. The arrows rather than tab,
+because this sheet asks one thing: up and down are the accounts under the tab, so left and right
+are the tabs. One CLI on its own is a heading rather than a row of tabs: there is nowhere to
+switch to, so nothing says there is. Where the flow said that agent has to run a particular moment, only the CLIs that
 run it are here — choosing one that does not is a flow that would refuse to start.
 
-The first row is what every agent ran as before there were any accounts: whatever this machine
-is already signed in as. It is asked here rather than hidden behind a chord because it decides
+`as installed` — the first row — is what every agent ran as before there were any accounts:
+the CLI signed in the way you signed it in, with nothing redirected. It is asked here rather than hidden behind a chord because it decides
 which credentials the turns run under, which is not a side question about anything: two agents
 of one CLI, one on a subscription and one on somebody's gateway, are two accounts running at
 once, each refreshing its own token and neither able to read the other's.
@@ -282,7 +283,7 @@ that runs it, and the CLI was settled a step ago, so it is named above them as a
      1. claude-opus-5           claude
    ❯ 2. claude-sonnet-5         claude
 
-   ◉ max effort  ←/→ to adjust · ◉ swarm mode off  ctrl+w to toggle · ◉ every skill  ctrl+s to choose · ◉ bypass  ctrl+p to change · ◉ as deepseek
+   ◉ max effort  ←/→ to adjust · ◉ swarm mode off  ctrl+w to toggle · ◉ every skill  ctrl+s to choose · ◉ bypass  ctrl+p to change
 ```
 
 Typing narrows the list — `cop` finds `claude-opus-5`, since nobody types a model id out — and
@@ -294,11 +295,14 @@ Under the models, the things that are adjusted rather than chosen: **←/→** t
 [what it may do](agents.md#what-an-agent-may-do). Those three really are side questions about
 the same agent, which is why they are keys here rather than steps of their own.
 
-The same line also **reads back** what a step of its own has already settled, with no key on it:
-the account from the step before, and where it works — `◉ in a container of python:3.12` for an
-agent the flow put in a container of its own, or `◉ on ssh://box` for one already pointed
-somewhere. An agent that works here says nothing there, which is what an agent nobody said
-anything about has always done.
+The account is **not** on that line. It was the step before this one, it is on the line above
+the prompt once the walk is over, and a setting shown where it cannot be changed is a setting
+somebody tries to change.
+
+Where the agent works is read back there, with no key on it, where a step of its own has
+already settled it: `◉ in a container of python:3.12` for an agent the flow put in a container
+of its own, or `◉ on ssh://box` for one already pointed somewhere. An agent that works here says
+nothing there, which is what an agent nobody said anything about has always done.
 
 Enter takes the row under the cursor and goes on: to where that agent works, if the flow says it
 may be pointed anywhere, and otherwise to the next agent the flow drives.

src/humanize/tui/SPEC.md

@@ -102,9 +102,14 @@ the cursor, and under them whatever is adjusted rather than chosen.
   are each a question nobody is being asked.
 - What an agent runs MUST be one choice rather than two, a model belonging to the CLI that runs
   it; and the CLIs MUST be read one at a time, a tab apiece, since every model of every CLI in
-  one list is a list that grows each time any of them ships a model. Tab and shift+tab MUST
-  turn between them and MUST wrap. One CLI MUST be a heading rather than a row of tabs: there
-  is nowhere to switch to, so nothing MUST say there is.
+  one list is a list that grows each time any of them ships a model. The arrows the list itself
+  is not using MUST turn between them, and MUST wrap: a sheet that asks one thing has its keys
+  free, and a chord for what a row of tabs is for would be a key to already know. One CLI MUST
+  be a heading rather than a row of tabs: there is nowhere to switch to, so nothing MUST say
+  there is.
+- A step MUST NOT read back what another step settles. What cannot be changed where it is shown
+  is a setting somebody tries to change there; what the walk has settled is on the line above
+  the prompt when it is over.
 - Typing MUST narrow the list being read and MUST belong to the tab it was typed into: a search
   that narrowed one CLI to one model would narrow the next to none, which reads as a CLI with
   nothing in it. Esc MUST clear what was typed before it leaves, on every sheet that is

src/humanize/tui/pick.py

@@ -661,11 +661,11 @@ def _fill(self) -> None:
         )
 
     def _tuned(self) -> str:
-        """The line under the models: the effort, and the rest of what this agent is.
+        """The line under the models: the effort, and the side questions about this agent.
 
-        What is adjusted here carries the key that adjusts it. What a step of its own settled
-        -- the account it runs as, the container the flow put it in, the machine it was
-        pointed at -- is read rather than adjusted, and says so by carrying no key at all.
+        What is adjusted here carries the key that adjusts it. Where this one works is read
+        rather than adjusted where the flow settled it, and says so by carrying no key at all;
+        the account it runs as is not here at all, being the step before this one.
 
         Returns:
           The line, as markup, or "" where the letters typed have left no model to say it of.
@@ -694,11 +694,9 @@ def _tuned(self) -> str:
             f"{_DOT}[$secondary]◉[/] {PERMISSIONS[self._permission]}  "
             f"[$text-muted]ctrl+p to change[/]"
         )
-        as_who = self._whose.provider
-        said += (
-            f"{_DOT}[$secondary]◉[/] as "
-            f"{escape(as_who) if as_who else 'this machine is signed in'}"
-        )
+        # The account is not read back here: it was the step before this one, it is on the
+        # line above the prompt once the walk is over, and a setting shown where it cannot be
+        # changed is a setting somebody tries to change.
         if where := self._where():
             said += f"{_DOT}[$secondary]◉[/] {where}"
         return said
@@ -1526,12 +1524,11 @@ class RunsAs(Sheet[Whose]):
 
     BINDINGS: ClassVar = [
         ("escape", "back", "back"),
-        # The tabs: one CLI at a time, forwards and back. Priority, or the list under the
-        # cursor would take tab as moving the focus and the interface below would take
-        # shift+tab as switching the flow -- which is not a thing to do from inside a sheet
-        # that is choosing what the flow runs on.
-        Binding("tab", "next_cli", "next CLI", priority=True),
-        Binding("shift+tab", "prev_cli", "previous CLI", priority=True),
+        # The tabs, on the arrows the list is not using: up and down are the accounts under
+        # the tab, so left and right are the tabs themselves. Priority, or the list under the
+        # cursor would take them as moving between rows it has none of.
+        Binding("left", "prev_cli", "previous CLI", priority=True),
+        Binding("right", "next_cli", "next CLI", priority=True),
         # Making one is asked for here rather than somewhere else: this is the moment somebody
         # finds out they have no account for this CLI, or that the one they want is not among
         # these, and sending them out of the question to answer it would lose the question.
@@ -1618,7 +1615,7 @@ def _tabs(self) -> str:
         )
         # Only where there is somewhere to switch to: one CLI is a heading rather than tabs.
         if len(able) > 1:
-            said += "   [$text-muted]tab/shift+tab to switch[/]"
+            said += "   [$text-muted]←/→ to switch[/]"
         return said
 
     def _turn_to(self, by: int) -> None:
@@ -1657,9 +1654,9 @@ def _ask(self) -> None:
         )
         needs = self._place.moments
         self.query_one("#about", Label).update(
-            f"Which coding agent takes this one's turns in {self._flow}, a tab apiece, and "
-            "under it that CLI's own accounts -- an account is one backend's. Its sessions, "
-            "its settings and its skills are the CLI's own whichever account it runs as."
+            f"Which coding agent takes this one's turns in {self._flow}, one per tab on the "
+            "arrows, and under it that CLI's own accounts -- an account is one backend's. Its "
+            "sessions, its settings and its skills are the CLI's own whichever it runs as."
             + (
                 f" This one has to run {', '.join(sorted(needs))}, so only the CLIs that do "
                 "are here."
@@ -1701,7 +1698,7 @@ def _fill(self) -> None:
             f"[$text-muted]{said}[/]" if said else ""
         )
         self.query_one("#keys", Label).update(
-            "ctrl+n to make one · Type to search · Enter to choose · Esc to go back"
+            "←/→ CLI · ctrl+n to make one · Type to search · Enter to choose · Esc to go back"
             f"{self.searching()}"
         )
 
@@ -1713,7 +1710,9 @@ def _accounts(self) -> list[tuple[str, str, str]]:
         if not backend:
             return []
         return [
-            ("", "as this machine is signed in", "nothing is redirected"),
+            # Two words for the account nobody chose: this is a row in a list of accounts, and
+            # what it is is the one the CLI is already signed in as.
+            ("", "as installed", "signed in as you signed it in"),
             *((one.name, one.name, _sets(one)) for one in providers.providers(backend)),
         ]
 

tests/tui/test_agents_walk.py

@@ -264,7 +264,7 @@ async def test_stepping_back_and_turning_to_another_cli_asks_about_that_one(
             "escape", "escape"
         )  # back to which CLI, and turn to the other
         await until(lambda: isinstance(app.screen, RunsAs), driver)
-        await driver.press("tab")
+        await driver.press("right")
         await driver.press("enter")
         await until(lambda: isinstance(app.screen, Models), driver)
         listing = app.screen.query_one("#choices", OptionList)

tests/tui/test_app.py

@@ -1108,7 +1108,7 @@ async def test_what_an_agent_runs_is_one_cli_at_a_time_and_an_effort_the_arrows_
         assert "claude" in tabs
         assert "codex" in tabs
         assert "kimi" not in tabs  # not installed, so not a tab
-        assert "tab/shift+tab to switch" in tabs
+        assert "←/→ to switch" in tabs
 
         await into_models(app, driver)
         sheet = app.screen
@@ -1150,15 +1150,15 @@ async def test_what_an_agent_runs_is_one_cli_at_a_time_and_an_effort_the_arrows_
         "codex": (Model("gpt-5.6-sol", ("xhigh",)), Model("gpt-5.5", ("high",))),
     },
 )
-async def test_tab_and_shift_tab_turn_to_the_next_cli_and_the_one_before(
+async def test_the_arrows_turn_to_the_next_cli_and_the_one_before(
     _installed: unittest.mock.MagicMock,  # noqa: PT019  -- `mock.patch` hands it over
 ) -> None:
     """Which is the point of the tabs: one CLI's accounts rather than everyone's at once.
 
-    They wrap, so the last tab is one press from the first however many CLIs are installed --
-    and shift+tab here is the tab before rather than the interface's own next flow, which is
-    not a thing to do from inside the sheet that is choosing what this flow runs on. What the
-    tab was left on is the CLI whose models the step after asks about.
+    They wrap, so the last tab is one press from the first however many CLIs are installed.
+    The arrows rather than tab, now that this sheet asks one thing: up and down are the
+    accounts under the tab, so left and right are the tabs. What the tab was left on is the
+    CLI whose models the step after asks about.
     """
     app = Humanize()
     async with app.run_test() as driver:
@@ -1170,18 +1170,20 @@ async def test_tab_and_shift_tab_turn_to_the_next_cli_and_the_one_before(
         await until(lambda: bool(listing.options), driver)
         flow = app._flow_named
 
-        await driver.press("tab")
+        await driver.press("right")
         await driver.pause()
         assert "[b $primary]codex" in str(sheet.query_one("#tabs", Label).content)
 
-        await driver.press("tab")  # round the end, back to the first
+        await driver.press("right")  # round the end, back to the first
         await driver.pause()
         assert "[b $primary]claude" in str(sheet.query_one("#tabs", Label).content)
 
-        await driver.press("shift+tab")  # and back the other way
+        await driver.press("left")  # and back the other way
         await driver.pause()
         assert "[b $primary]codex" in str(sheet.query_one("#tabs", Label).content)
-        # The interface's own shift+tab did not fire under the sheet.
+        # And the interface's own shift+tab is still not a thing to do from under a sheet.
+        await driver.press("shift+tab")
+        await driver.pause()
         assert app._flow_named == flow
 
         await driver.press("enter")  # as this machine is signed in, on codex
@@ -1226,7 +1228,7 @@ async def test_one_cli_is_a_heading_rather_than_a_row_of_tabs(
         assert "claude" in tabs
         assert "to switch" not in tabs
 
-        await driver.press("tab")  # nowhere to go, and nothing moves
+        await driver.press("right")  # nowhere to go, and nothing moves
         await driver.pause()
         assert [str(option.id) for option in listing.options] == ["="]
 
@@ -1271,7 +1273,7 @@ async def test_what_was_typed_belongs_to_the_tab_it_was_typed_into(
         await driver.pause()
         assert [str(option.id) for option in listing.options] == ["=deepseek"]
 
-        await driver.press("tab")
+        await driver.press("right")
         await driver.pause()
         # The next CLI is read from its whole list rather than through the last search.
         assert [str(option.id) for option in listing.options] == ["=", "=work"]
@@ -1313,7 +1315,7 @@ async def test_a_turn_is_said_to_run_hard_and_said_to_run_wide_separately(
         # Kimi has one. Which CLI it is was the step before, so that is where it is turned to.
         await driver.press("escape")
         await until(lambda: isinstance(app.screen, RunsAs), driver)
-        await driver.press("tab")
+        await driver.press("right")
         await driver.press("enter")
         await until(lambda: isinstance(app.screen, Models), driver)
         tuning = app.screen.query_one("#tuning", Label)

tests/tui/test_providers.py

@@ -342,9 +342,11 @@ async def test_the_first_row_leaves_the_agent_running_as_this_machine(
         await driver.press("enter")
         await until(lambda: isinstance(app.screen, Models), driver)
         tuning = app.screen.query_one("#tuning", Label)
-        await until(
-            lambda: "as this machine is signed in" in str(tuning.content), driver
-        )
+        await until(lambda: "effort" in str(tuning.content), driver)
+        # The account was the step before this one, so it is not read back here: a setting
+        # shown where it cannot be changed is a setting somebody tries to change.
+        assert "as installed" not in str(tuning.content)
+        assert "deepseek" not in str(tuning.content)
         await driver.press("enter")
         await until(lambda: not isinstance(app.screen, Models), driver)