Skip to content

fix(cockpit): a waiting tab's question skips Claude Code chrome, its ready answer goes out with Enter - #702

Merged
mboiman merged 1 commit into
bks-lab:mainfrom
mboiman:fix/cockpit-tab-answer
Nov 21, 2026
Merged

mboiman merged 1 commit into
bks-lab:mainfrom
mboiman:fix/cockpit-tab-answer

Conversation

@mboiman

@mboiman mboiman commented Nov 21, 2026

Copy link
Copy Markdown
Contributor

What was wrong

On a live Claude Code tab the cockpit read the wrong words as the question. It took the session recap (※ recap: ...) and the right-aligned hint line (new task? /clear to save ... · ✔ Update installed · /restart to apply) instead of the agent's last words. The question then did not end with ?, so the page offered no yes or no. The text in the input line (often Claude Code's own grey suggestion) counted as a blocking draft, and the page still offered buttons that type words, which the server then refused (tab_draft).

What changes

  1. The question is the agent's words, never chrome (asks.py). The block starts at the last ⏺ line and ends at the first line Claude Code draws itself: the turn status line (✻ ... for Ns), the recap with its wrapped lines, hint lines (/clear to save, /restart to apply, Update installed, turn off recaps), or any line pushed to the right edge. The fallback for a screen with no ⏺ drops the same lines. Blank lines inside the agent's words stay part of the block.
  2. The input line is a suggested answer. question_of returns it as suggestion. draft is kept, and suggestion is not part of the hash. A line with no word character in it (❯ #) gives no suggestion. The new verb tab_accept (needs the driver's key right) presses Enter alone. It is guarded by guard_fresh with the question hash, the tab name and the state, and it refuses with tab_moved when the input line read fresh no longer holds exactly that text. In replyOptions that answer is the first, primary button (Send "...", de: „...“ senden), and the hint now says the tab already has an answer ready.
  3. No button the server would refuse. While the input line holds text, the page offers no Yes or No as words, no "Keep going", no text field and no guide answers that type words. The slim tab card and TabControls hide go on as well. On the server, tab_go_on now always checks the input line, even when no question hash comes with it, so a draft can no longer be appended to through the tab controls.

Docs (docs/bridge-cockpit.md), the demo copy (docs/cockpit-demo/) and the mock follow.

Tests

  • New failing-first tests: the real screen tail (with and without a ⏺ head), the update-hint screen, the suggestion and its scrub, tab_accept (press, changed line, empty line, new question, missing args, dialog, no key right), go on without a hash on a draft, and in replyOptions the ready answer plus no refused buttons in any draft state, rendered through TabReply.
  • Server suites: 2950 passed, 4 skipped. Frontend suites: 1338 passed. build-cockpit-demo.py --check and validate-bridge.py are clean.

🤖 Generated with Claude Code

…ready answer goes out with Enter

The question of a waiting tab is the agent's last block of words (from its last
⏺ line to the first line of Claude Code's own): no turn status line, no session
recap with its wrapped lines, no right-aligned hints such as "/clear to save"
or "Update installed · /restart to apply". A question wrapped over two lines
now ends with its "?" again and reads as yes or no.

The input line may hold Claude Code's grey prompt suggestion or typed words;
the screen cannot tell them apart. question_of returns that text as
`suggestion` (`draft` stays), and the new verb tab_accept presses Enter alone,
guarded by the question's hash and the very words of the line read fresh. The
page leads with that answer ("Send \"...\"", de: „...“ senden) and offers
nothing that types words while the line holds text, since the server refuses
it (tab_draft). tab_go_on now always reads the input line, also without a
question hash.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Signed-off-by: mboiman <mboiman@gmail.com>
@mboiman
mboiman merged commit c978252 into bks-lab:main Nov 21, 2026
21 checks passed
mboiman added a commit that referenced this pull request Nov 21, 2026
…the bucket's answers (#703)

* fix(briefing): an item waiting on the Bridge is delegate, and every printed row has a number

An inbox item put off waiting on the Bridge (inbox.py defer --on Bridge) stood under
"Waiting on others", also through a section forced to bucket: waiting (8 rows on
21.11.2026). It is the Bridge's to do (delegate), or the person's (do) when only they can
take its next click (gate only-you, a login). task_who reads "the Bridge" in any spelling.

The numbers a briefing prints come from one rule (briefing_view.numbered) and stand on
each row as `nr`; calm now numbers its waiting rows too, each with its topic. A cockpit
pass (every row, --max-items) numbers as the printed briefing would: the profile's style
and max_items, without what was settled in the cockpit. The view names every shown bucket
with its title and answers (`bucket_defs`), an empty one too.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Signed-off-by: mboiman <mboiman@gmail.com>

* feat(cockpit): Work through by the briefing's buckets, one card with the bucket's answers

With cockpit.flow.buckets (feature flow_buckets, on by default) every point of Work
through stands in exactly one bucket of the briefing profile (the server's bucket rule;
names, order and answers from view.buckets), the left column lists the buckets with their
counts. One card at a time says one sentence Next (the tab's question, the card's step, or
a fixed sentence of its bucket) and offers the bucket's answers as big buttons, the first
one the Bridge taking the next step; each answer is the act the briefing's answer is
(do_it, later, settle/shelve, a day picked, runs_on, task_keep, a tab's own answers
including tab_accept). Everything else stands under More, there is no right column, an
answer moves on by itself, and the card names its number in the printed briefing. The
Aufgaben columns and Jev's squares say the buckets' names while it is on.

The classic card of a tab that asks shows the question itself (the last sentence that
asks, across wrapped lines), no longer its screen's first line (#702 follow-up). A task a
page names beside the briefing no longer carries the default bucket do.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Signed-off-by: mboiman <mboiman@gmail.com>

* chore: the new suites ship as CORE, the explore page counts them

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Signed-off-by: mboiman <mboiman@gmail.com>

---------

Signed-off-by: mboiman <mboiman@gmail.com>
Co-authored-by: Claude Opus 5.5 <noreply@anthropic.com>
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

1 participant