Repository navigation
v0.21 bugs: wrapped-row redraws (#388), command colors after a slow zshrc, width in /modules explain (#365) - #423
Merged
Merged
Conversation
…ever erases earlier output The live region (#412) moved the cursor by stored lines. Stored lines are logical, but programs count terminal rows: a progress display whose line is wider than the window (two rows at 40 columns) moved up three rows to redraw itself, NMSh moved up three lines, and `CSI J` erased the program's own log lines above the display, one more per frame. Reproduced in the parser, in a real zsh PTY at 40 columns, and through the frontend; the new tests fail on the previous parser. - The frontend reports the PTY's width (setScreenColumns). A line wider than it occupies several rows: cursor up/down, `ESC M/D/E` and line feed move by rows (the next row of a wrapped line, or the next line), within the same bounds. - Carriage return, backspace, `CSI G/C/D` and `CSI K` act on the cursor's row within its line; erasing never removes the row the cursor is on. The pending wrap after a row is filled exactly is tracked from the character that filled it. - A resize re-wraps stored lines at the new width, as a reflowing terminal does. - The contract stays deliberately limited and is now explicit and tested: absolute rows (`CSI H/f/d`) are not mapped (no cursor position reports are answered, so an inline program cannot know its rows); scroll regions are not emulated, except that text drawn below a region the program set (an installer's reserved status row) is not stored, so it cannot leak into the log rows. NMSh's own lines are never dropped. - Unknown width keeps the previous line-based behavior. Refs #388
…om when the helper reads them Found while re-verifying #408. The semantic helper (a zsh that sources the user's configuration and answers `whence` lookups) reads requests only after its startup files finish, but each lookup's 1.5 s limit started when it was sent. A configuration that takes longer to load (nvm, a large framework: 2 s reproduced) made the first lookup time out, which ends the helper; every later executable or builtin then stayed uncolored (neither known nor unknown) for the whole session. The helper now says when it reads requests; lookup limits (and raw source/facts lookups) start then, and startup itself has its own 20 s bound. The new test (a 2 s .zshrc) fails on the old service: `ls` came back unknown after 1.5 s and the helper was dead. #408's own synchronization was re-checked: newer shell knowledge bumps the generation and clears pending lookups, so a late answer finds no pending entry and is never cached; superseded lookups ask again (at most three times). Refs #408
…ortened it, for width (#365) #365 asks why a module is visible, hidden or compacted. Explain covered routing, visibility, conditions and source state; width was missing. The Context Rail fits module groups by priority (place, compact, then drop) and the Status Strip narrows forms, then drops the lowest priorities; both now fill an optional fit report on their live render (previews of a draft never overwrite it), and `/modules explain` adds "left out for width in the last frame (higher-priority modules kept)" or "shortened to fit the width in the last frame". A strip too narrow to show at all reports every routed module as left out. A module turned off is still explained as off. Nothing is collected or measured to produce it. The Main Prompt and Right Context truncate text rather than drop modules; a per-segment truncation trace there is left as future work (documented). Refs #365
This was referenced Oct 18, 2026
Merged
5 tasks
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
1 participant
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
Stacked on #422. Bug fixes from the v0.21 stabilization review of every open
type:bugissue, plus two defects found while verifying them.#388: a redraw over a wrapped line erased earlier output
The live region (#412) moved the cursor by stored (logical) lines, but programs count terminal rows. At 40 columns, a progress display whose line wraps to two rows moved up three rows. NMSh moved up three lines, and
CSI Jerased the program's own log lines above it, one more per frame. Reproduced in the parser, in a real zsh PTY at 40 columns, and through the frontend.ESC M/D/Eand line feed move by rows. CR, BS,CSI G/C/DandCSI Kact on the cursor's row. The pending wrap after an exactly-filled row is tracked. A resize re-wraps.CSI H/f/d) are not mapped, because an inline program cannot know its rows (cursor position reports aren't answered). Scroll regions aren't emulated, except that text drawn below a region the program set (an installer's status row) is not stored, so it can't leak into log rows.Command colors stopped for the session after a slow
.zshrc(found verifying #408)The semantic helper reads lookups only after sourcing the user's configuration, but each lookup's 1.5 s limit started when it was sent. A 2 s configuration (nvm, a large framework) made the first lookup time out, which killed the helper: every later executable or builtin stayed uncolored. Lookups are now timed from when the helper reads them, and startup has its own 20 s bound. The test fails on the old service. #408's own generation/re-ask synchronization was re-checked and holds.
#365: why a module isn't shown, including width
The Context Rail and Status Strip record how their width fitting treated routed modules in the last live frame.
/modules explainadds "left out for width in the last frame" or "shortened to fit". Previews never overwrite the record. The Main Prompt and Right Context truncate text rather than drop modules; a per-segment trace there is documented as future work.Investigated, not reproducible
fish_readhandover fixes hold.fish -cuses never start the generator. The Flaky CI at the cause: task exit verdict, race-free private-root removal, stalled-file teardown #422 teardown preload is a further guard.Also checked: 99 adversarial input/recognizer combinations for Semantic Output (up to a 2 MB single line). None exceeded 150 ms.
Refs #388, #408, #365, #356, #403, #410