Skip to content

Fork: no Claude footer as the 'no summary' message - #6

Merged
jbutton merged 12 commits into
mainfrom
work/20-fork-no-claude-footer-as-the-n
Nov 12, 2026
Merged

jbutton merged 12 commits into
mainfrom
work/20-fork-no-claude-footer-as-the-n

Conversation

@jbutton

@jbutton jbutton commented Nov 12, 2026

Copy link
Copy Markdown
Owner

npm run check -- --all now passes on the branch as it stands: 1303 tests, and asking again answers "The check already passed for this exact code." The branch holds all of main (d03cf4d), so the board has nothing new to merge before its run.

As in the two earlier rounds, no test failed. Each time, another task had just landed on main and the board merged it into my branch first (this time PR #4 from task 9, e7b1d4e). That changes the code, so every test has to run again. With other tasks' whole checks running at the same time, the run goes past the board's 15-minute limit, check.timeout. My runs took 8 to 11.5 minutes. The warnings in your output were already there before my change, and my fix, tests and docs are untouched.

If it happens again, raising check.timeout for herdr-projects would let the board's own run finish. That setting is yours to change, so I've left it alone. Nothing of mine is left running, and the closing note has this round's result.

The fix itself is unchanged. With a tall draft in Claude Code's input box, a task that stops without a summary now shows the agent's last words, not the footer. Claude Code's own lines just above the box are left out too: its ✻ … · done … line after each turn and its notice there.

Also noticed:

  • tests/real_terminal.rs: when the test panics, the board it started keeps running.
  • tests/real_terminal.rs:45,68: keys() waits only 150 ms for a whole frame, so a very busy machine can fail the test.
  • tests/coordinator.rs:35, tests/coordinator_chat.rs:32, tests/coordinator_chat_more.rs:37, tests/board_list/pinned.rs:230: the compiler warnings in your output, printed on every test build.

Task: PRJ-20 (herdr-projects/20)

jbutton and others added 12 commits November 12, 2026 21:06
Two screens recorded from a real Claude Code 2.1.322 (tests/fixtures/screens). With ten lines typed into its input
box, more than 8 lines stand under the prompt, agent_words finds no box and keeps the whole screen: the task's
"Stopped without a summary" message is the status line and "⏵⏵ bypass permissions on". At rest, the message ends
with Claude Code's own "✻ Cooked for 0s · done 9:03 PM" line instead of the agent's last words.

These tests fail here; the fix follows. The case was left open in round 3 of d8b21f3's review (upstream
eliasstravik/shepherdr #406).

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
…out a summary

agent_words looked for the input box's prompt only within 8 lines of the screen's end. Claude Code's box grows with
what is typed into it, so a draft of more than 8 lines left no box found, and the whole screen was kept: the task's
"Stopped without a summary" message became the status line and "⏵⏵ bypass permissions on" (the shape seen on the
board's stopped tasks). Now, with no prompt near the end, the box is found from its bottom rule near the end, up to
the prompt line right under its top rule; a bare ">" counts there only between a box's edges, so an agent's own
quote under its own rule is not taken for one.

Claude Code's own lines over the box are left out too, as on 10 of 22 live screens read on the board they were one
of the message's two lines: the line after each turn ("✻ Cooked for 0s · done 9:03 PM") and its notice pushed to the
right over the box ("new task? /clear to save 599.8k tokens", "✔ Update installed"). With nothing left, the message
stays "Stopped without a summary".

Spec: eliasstravik/shepherdr #406 ("Stopped without a summary" shows the agent's status line as the message); the
case was left open in round 3 of d8b21f3's review. Shepherdr itself has no such state, so there is no upstream code
to port. Tests from real Claude Code 2.1.322 screens (tests/fixtures/screens).

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
…aude's notice

Review found that the notice rule (8 spaces in, ending near the rule's end) also dropped an agent's own indented
line right over the box: with an 80-column rule, "        curl --retry 3 … /health" was cut and only "⏺ The command
failed:" was left. A notice is pushed to the right: on the 10 live screens that had one it started at column 71 or 112
of a 152-column rule. It now has to start past the first third of the rule's width as well; the reviewer's probe is
a test, and all 10 live notices are still left out.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
…s place

Second review round: an agent's indented command right over the box was still taken for Claude Code's notice in
narrower panes (a 40-column rule with 16 spaces, a 24-column rule with 8), and a joined row longer than the rule
passed the "ends at the rule's end" test too, so the task's message lost the command. Position alone cannot tell a
notice from an indented line in every pane width.

A notice is now told by the words Claude Code writes there (seen on the 10 live screens that had one: "/clear to
save", "Update installed", "Restart to update"), in a row that ends at the box's top rule, within 4 columns and never
past it. An agent's own line there is kept in a pane of any width; a notice with other words stays in the message,
which loses nothing of the agent's. The reviewer's probes are tests, with a notice in a narrow pane still left out.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
…t out

Third review round: an agent's own "echo 'Restart to update'", 14 spaces in, right over a 40-column box, was cut
from the task's message, since a row holding a notice's words counted as the notice.

The row right over the box now counts as Claude Code's notice only when it is nothing but its notices, each whole,
joined by " · " ("new task? /clear to save <n> tokens", "✔ Update installed", "Restart to update", as on the 10
live screens that had one), and it ends 2 columns before the box's top rule ends, as every one of those did. A line
of the agent's own there is kept, even one that holds a notice's words; a notice worded otherwise stays in the
message. The reviewer's probe is a case of the narrow-pane test, with more look-alike rows in the notice test.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
@jbutton
jbutton marked this pull request as ready for review November 12, 2026 22:59
@jbutton
jbutton merged commit 8e405dc into main Nov 12, 2026
0 of 2 checks passed
@jbutton
jbutton deleted the work/20-fork-no-claude-footer-as-the-n branch November 12, 2026 22:59
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