chat: onboarding answers the pointer and a second enter, no telemetry notice or command, and free models only on a low account - #1720

Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>

Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>

Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>

…message box and on the setup screen

A balance read that OpenRouter answered with `API key expired` was a failed
read: nothing recorded, nothing said, and the next turn failed with a plain
auth error. The 401 that says expired is now a reading of its own
(credits.Reading.Expired, kept beside low in credits.json): `Your OpenRouter
key has expired — make a new one at openrouter.ai/settings/keys` stands at
the right of the keys row in a conversation and on Home for every model the
default service serves, free included, and under the chat model on the
setup screen, where the list is not cut because the free rows fail on the
same key. A turn refused as expired asks for a fresh read, through a fact
stamped on the provider's refusal (APIError.KeyExpired) so no surface reads
the status for itself. A new key clears it before it is read.

Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>

… window allows

The example was a second column beside the form, 36 cells wide, drawn
from 112 columns up and level with the first field: on a tall window the
lower half of the screen stayed empty while the panel wrapped every
sentence three ways. It is one column now — the form, its keys line, one
blank row, then the panel at up to 92 cells, whole or not at all. A window
with no rows to spare under the form draws the form alone and the keys
line stops naming the arrows. The form is 64 cells on every window.

Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>

…ts title on its edge, and shows a /senior-dev hand-off

Under the form, the legend stood between the form and the panel and read
as a caption for the wrong one. The order is now the panel, two blank
rows, the keys line, and the form directly under it. The panel's top edge
reads `○ Example · <the example's title>` and the title is no longer a row
of the body. A fifth example, `Hand off complex coding tasks`, asks
`/senior-dev Add retries with backoff to the HTTP client, with tests.`
beside `Start a conversation`, and a command in any example's request
wears the composer's chip. The keys line says `enter sets the limit` on
the limit row and `↑↓ moves` in place of `tab moves` everywhere; the tmux
suite's needle follows it.

Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>

…t, the panel loses its foot line, and the account's warning rides the frame's foot

`Models and spending` over `Keep these choices or change them.` is one
line, `Basic settings`, and the keys line is the row directly under it
rather than the foot of the form. The keys line no longer names the
example's arrows: the panel's own bottom edge carries them. The panel no
longer ends on `An illustration. Nothing here has run.`; its `Example`
label says that once. The low-credits and expired-key lines leave the
chat-model row and stand right-aligned on the frame's last row, where the
keys row carries the same warning outside the setup; the expired-key line
is shortened to fit any window the panel fits.

Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>

…grey once answered, white until then

`→` on the last example is the first again, and `←` on the first the
last. Every row's name — the limit, the chat model, the review, the way
out — is painted by one rule: the accent while the focus is on it, dim
once enter has acted on it (the limit set, a model taken from the list,
the review shown), and the body ink until then. Opening the list and
leaving it with esc answers nothing.

Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>

Each row was the catalog's friendly name with the exact id drawn under
the cursor's row, which read as a heading with a subheading and made the
one row a person could act on two rows tall. Each row is the id now, one
row per model; typing still finds a model by its friendly name and the
field above still shows it.

Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>

…t points at /settings

`Other settings use defaults · review` showed three settings and let
nobody change them — a door painted on a wall, and people pressed on it.
The form is three rows now: the limit, the chat model, the way out. Under
`Start a conversation` one dim line reads `Everything else is in
/settings`, with the command painted as the composer's chip.

Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>

…ut carries no loose enter

The limit's line ends `/budget changes it later.` and the chat model's
`/model changes it later.`, each command painted as the composer's chip
the way the note's /settings is. The dim `enter` at the right of `Start a
conversation` is gone: the keys line under the heading already says
`enter starts` when the focus is there.

Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>

…and never follow the focus

The panel followed the focused row, which read as it jumping about under
a person walking the form. It opens on the first example and turns to
the next every three seconds, round the ring; ←/→ browse by hand; any
key — typing, walking the rows, browsing, a press — holds the clock for
a full three seconds from that key; the screen-reader tier never turns
by itself. Each arriving example plays its demonstration once.

Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>

…n's ceilings

`?` on the Daily limit row now says what the two ceilings are and how
they meet: /budget 50 changes the day's and /budget none removes it;
/budget conversation 20 gives the open conversation a smaller ceiling of
its own and keeps it as the default for new ones; both hold at once and
whichever is reached first stops the work — the day's everything until
midnight, the conversation's just that conversation; the running turn
always finishes. Each /budget wears the composer's chip.

Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>

…on the list

TestModelNamesReadAsNamesAndKeepTheirLevel still expected the open list to
draw "DeepSeek V4 Flash", which the flat list of exact ids stopped doing on
purpose; it now asserts the two halves that hold — the field row and the
typed search speak the friendly name, every row of the open list is the
exact id — and leaves the flat shape itself to the test that owns it.

Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>

… the setup's prose catches up

The setup's foot said "esc to paste a new one" on every frame, while the keys
line two rows up said "esc skips setup" whenever the controls step stood
alone — a key already in the shell or the profile, only the controls asked.
The line is now chosen by the same rule as the keys line: with a connect
step before it, esc goes back there; standing alone, it names /connect as
the door to a new key. A test pins the alone case on the same frame.

The manual stops listing the review row's enter verbs (`shows them`, `goes
on`) and stops saying the setup shows memory, permissions and the countdown
under `Other settings`; both pages that explained esc on an expired key say
the conditional thing.

The comments the redesign left behind say what is true now: the example
turns on a clock and the focus never moves it, the panel stands above the
form, the edge label — not a foot line — is what says it is an illustration,
and `codeaf telemetry show` is spoken of as the command that went.

Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>

Quote boundaries had reached the send scan, so asking about /task could start work.
Keep the original word boundaries for action; resting words and follow-up plain ranges use the transcript's quote-aware painting scan.
Leave quoted door names plain after an ordinary send and in guard/revive wrappers, while live tags retain their chips.
Scope the bash guard to action and draft/transcript paint; informational lines beginning with ! still chip known commands.
Cover straight and curly quotes, whole-message delivery, unchanged unquoted tags, ordinary-send transcript paint, informational ! lines and the manual answer.

Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>

…g models

The engine flattened refusals into text and lost the expiry-triggered balance read.
Recognize that text beside provider status handling, and keep expired readings from moving untouched conversations between defaults.
Cover structured and flattened refusals, excluded statuses and direct services, and document both launch roads.

Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>

Keep the no-match model list open with its row unanswered and its focus and model unchanged.
Hold example turns behind one timer, including a tick queued before the latest key.
Build the free-id set once per catalog pass and prove the 400-row order and membership; omit the wall-clock benchmark comparison.

Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>

…d off

The settings row persisted off while the running process kept sending usage and exit events.
Close the existing process gate after a successful off write; enabling waits for startup to resolve all opt-outs.
Name replacement switches in the removed-command error and make the disclosure, hints and manual agree on timing, defaults and relay.

Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>

The controls explanation exceeded a standalone manual read, and the change entry still claimed removed review and example-key behavior.
Split self-contained setup topics, add retrieval probes, and describe current credit, model-list and telemetry outcomes.
Use the required first-run title and describe only origin/dev-to-PR invalidations, including the empty model list and telemetry switches.
Split the empty-task-page answer too because the changed corpus ranking otherwise hid an existing probe; leave every existing probe unchanged.
Keep the standalone section-size test; drop the transient change-entry test and duplicated exact-sentence assertion.

Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>

The engine sends the session's unauthorized ending without its typed provider refusal, so the ordinary launch missed the fresh balance read.
Export one unauthorized sentence for both session spellings and match its trimmed prefix after the engine wire, under the existing refusal debounce.
Keep typed expiry recognition and the direct-service exclusion; prove the real 401 ending, arbitrary key-source tails, excluded statuses and unchanged conversation defaults.
Document that the read decides expiry and an invalid key leaves the previous reading alone.

Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>

A landing retained its sendable pool configuration through judging, so turning the settings switch off still allowed a later POST.
Resolve the live pool ladder before adding outgoing scores and at each HTTP request, including later batches and retries; keep local scores and leave older unsent rows pending.
Preserve the pool's existing behavior on source and test builds through its own configuration ladder.
Prove a switch changed inside the judge sends no new request, a switch changed during the first batch stops the next one, and ordinary pushes and quiet modes still work.

Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>

Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>