Skip to content

0.24 work, module split, and bug fixes - #4

Merged
tomeyer97 merged 292 commits into
mainfrom
fix/bugs
Nov 18, 2026
Merged

tomeyer97 merged 292 commits into
mainfrom
fix/bugs

Conversation

@tomeyer97

Copy link
Copy Markdown
Owner

Brings main up to date with three things, in order:

  1. fix/remote-perf (already on GitHub): the Remote-SSH and big-repo speed work and the understanding-check tour.
  2. Module split (two review-and-cleanup rounds): src/extension.ts went from ~4,800 lines to a list of boot steps plus a command table per area. Code now lives in src/ask, check, notes, tour, journeys, sidebar, write and webview. The CLI source is modules under cli/understanding/, and cli/bundle.mjs builds them into cli/understanding.mjs. Repeated helpers are now defined once, and the tests are one test per behaviour, in a file per module. Behaviour is unchanged: each step compared old and new output on the same inputs (random and real journeys and notes) and found no differences.
  3. Bug fixes found during the cleanup. Each is its own commit with a test that fails before the fix. See the CHANGELOG's "Fixes" entry. The ones a user notices:
    • The sidebar's "Failed to resolve tree node" errors are gone. src/sidebar/viewSync.ts waits to reveal a row until the view holds it, and spaces redraws so the window applies each one.
    • Undo of Mark All Read makes resolves and reopens unread again.
    • Notes queued for Claude no longer stay "Queued…" when the CLI isn't found.
    • The CLI and the extension now place notes the same way. A shared fixture runs against both, and a random comparison went from 3,159 disagreements to 0.
    • Step text: bare #L refs, line and range labels, and a step link opening the journey of its own workspace folder.
    • With reduced motion, journey scenes don't animate at all. The scenes harness now asserts this.

Behaviour change: the default branch is one rule for the extension and the CLI: origin/HEAD when known, else main or master, plus the compared-with branch. In a repo whose origin/HEAD is develop, pushing main now gets background questions, and pushing develop doesn't.

Tested

  • npm test: 285 tests pass.
  • All five screenshot harnesses pass (panel, scenes, check panel, check tour, journey list).
  • npm run test:integration: two full runs pass, with no tree-view errors in the log.
  • An end-to-end replay suite in VS Code for the web passes: create, reply, resolve, delete/undo, file and folder notes, sessions, journey walk, journey after code changes, keyboard chords, journey map.

🤖 Generated with Claude Code

tomeyer97 and others added 30 commits November 16, 2026 14:24
…er the grade moves on

A pass no longer jumps to the next question, so its feedback is read where it
is. The reply box is empty once the answer is in. Moving on is the link under
the last grade (shown while another question is open), or Ctrl+' D.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
… the Code Notes tree

Notes asked from a question (Ask Claude, a fix) stay in the tree: they are
ordinary notes.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
… it has been answered

Before, '⚠ May be a bug: fix it or explain' sat under the question and gave the
answer away. It now appears (as 'May be a bug.', with the actions) only once
an answer was graded.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
… line, said in the thread

Add TODO no longer un-flags a concern. It goes above the code the concern
names (else the question's first line), is at most 80 characters, and the
thread then says where it went ('TODO added at refunds.js:56') in place of
the action.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
…k that Apply puts in

The fix note now sits on the code the concern names, and Claude is asked for
the smallest change as one suggestion block for exactly those lines (the
notes' existing Apply; Undo takes it back), with no headings or essays.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
…ops between its steps

The tour (the check file's journey) is shown in the journey panel like a journey but
never listed with them. The storyline has a row per question after the step that
shows its answer; Next past that step opens the question's thread (the card says
"Question ›"), a pass walks on, Previous goes back. Check my understanding and
"Walk the tour" in the Check view open it where it was left; the view's rows say
their step and go to it. The tour is written right after the questions in the same
session, without a notification: the questions can be answered meanwhile. See also:
a journey of the repository on the same code, under an open question and on the
tour's last step (ranked like In this file, no model call). The grader gets the tour.

Tests: CLI and unit tests, the Check view and storyline harnesses at 280-1000 px, the
VS Code suite walks a tour end to end with the fake Claude.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
…nts before focusing the thread

A just-opened editor fetches its comment threads before drawing them; focusing
earlier left the keyboard in the code. The commenting-range request is the
signal (the half-second wait alone was not always enough).

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
…ong to its branch

They carry the check they came from and are hidden on other branches with its
threads, so they never show 'file not found' or offer Move.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
…ted next to linguist-generated

GitLab ignores linguist-generated. init now writes
'understanding/*.json gitlab-generated linguist-generated' and replaces the
line an earlier version wrote.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
…urn into GitLab references

'(~5 min)' rendered as a label link. Besides @mentions, numbered references
in quoted text (#12, !3, ~5, %2, $7, &4) get a word joiner, so they read the
same and link nowhere.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
…ething is partial and worth discussing

- Pass feedback confirms the answer's own point (a bug it names, confirmed)
  and adds at most one fact about that point, never an unrelated one.
- A pass whose feedback corrects a detail that matters ('retried later' when
  the retries stop) is 'partial': 'Mostly right.' + the correction; the
  reviewer's note is the correct point; on the merge request it is listed
  under 'Worth discussing' with that point instead of as a plain right answer.
- Two real-model eval cases (P1, P2) with a restock fixture.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
…n the line it names

The diff thread goes on the line the concern's grade linked (found again by
its text when it moved), not the last line of the question (a closing brace).
The bot comment, posted after the threads, lists each concern as a one-line
link to its thread instead of repeating it.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
VS Code can't put the caret in a thread's reply box, so the first keystroke was
lost. The selected question now has a box in the Check view (our webview):
opening a question from anywhere focuses it, Ctrl+Enter sends the answer by the
same path as a thread reply, Esc goes back to the code. Its grade shows under
the box (and in the thread); the grade's code links light the code and leave
the keyboard in the box; after a pass the box closes and 'Next question ›'
appears. Unsent drafts are kept per question in workspace state. The code is
revealed after the box has the keyboard, so typing never lands in source.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
… after Esc or a caret move

The webview reports whether its box got the keyboard; when a busy window
focused the view late, the extension focuses it once more, unless the person
pressed Esc or moved the caret meanwhile.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
…box and the tour of the change; a pass stays on its question in the tour too

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
… whatever is checked out meanwhile

The writer's script read the branch, commit and files from the working tree at
each step, so a switch before or during the writing put the questions (and a
--replace) on the other branch. Start now passes --branch/--head; the script
reads files at that commit and, when its branch is not checked out, finishes
into the git-ignored cache that is put in place on return. The Check view of
the other branch names the writing branch instead of showing it as its own.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
…ment and in the threads

An answer could fake the bot's own status line ('### ✅ … waived'), turn into
a heading over a '---', or break out of its quote with '</blockquote>'. Text
from people and from the grader is now escaped line by line (HTML as entities,
markdown punctuation with backslashes, line-start list and heading markers),
each answer line stays in its quote, and check threads show answers as typed.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
…em as a commit of their own

Untracked .gitattributes/.ignore written on one branch blocked 'git checkout'
of a branch where they were committed. The extension no longer runs init. While
the lines are missing, the view says 'Collapse check files in diffs: Set up ·
Hide'; Set up writes them and commits only those two files ('Add understanding
check setup'). CI adds a one-line tip to its log (never the comment).

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
…l, or where the removed code was used

The writer refused every file of a deletion-only change, so no check could be
written while CI demanded one. Removal points (the line after removed code, in
files that still exist) now count as changed lines, and when the change only
removes code, a question may also sit on the code that used it.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
…rds, with Retry

No more toast with the CLI's own advice ('Run --model …') or the writer's
jargon. The view says e.g. "Claude couldn't use the model. Retry · Settings"
(Retry replaces Start), the reason is logged, and a retry that works clears it.

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

A branch made from a branch that did its check was asked about the parent's
code again (9 lines counted for a 1-line change). When the tree holds another
branch's check whose code commit lies between the merge base and the head,
status, the question writer and CI start the change there. After the parent
is merged the merge base is past it, so nothing special remains.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
…ction around the cursor

Functions come from src/codeBlocks.ts: the language's symbols (never what is nested in a function, so a
destructuring or a callback is part of it), else the file's text while a language server is still starting.
A step on the comment just above a function belongs to it; a file's top-level consts share one lens. The lens
picker lists one row per journey. Every line of a function gives the same In this file ranking, led by the
journeys with the most steps in it.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
…80 characters wide; settled tree reveals; search order

- Refresh: Claude Code with the code-journey skill updates only the stale steps of a staged copy; once journey.py
  check adds no new problem, the copy replaces the journey file and a notice offers the diff. No note, nothing in
  the reader's name. From the preview or a row it refreshes that journey (it opened the picker before); the
  command's picker preselects only the journey in front.
- The lenses and In this file use the function blocks (and retry the symbols once a language server is up).
- The walk panel opens at the reader's last width (else 35 %, at most 40 %) and never leaves the code under ~80
  characters; too narrow for both: the step card only. Dragging the panel is remembered.
- Tree reveals go only to nodes that stay (Continue, In this file once the editor settles), and a step's question
  is never revealed in the Notes view: no more "Failed to resolve tree node". Tree redraws are coalesced.
- While searching, the sort reads "Best match"; another order picked applies to the results.
- Tests: refresh through the fake writer, one lens for top consts and one per function, the panel's share, no tree
  view errors in the extension host log; the test window is zoomed out so code and panel fit side by side.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
… about the move itself

git paired a renamed file only when both paths were in the diff's pathspec;
only the new one was, so a pure rename counted its whole content as new code
and got questions about lines nobody touched. Both paths are passed now. A
change that only moves files says 'Renamed, 0 lines changed' and still needs
one question (whole file) about what the move changes; CI agrees.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
After 'git branch -m' (or pushing the same commits under a new name) the check
file of the old name is in the tree and covers every commit, so status says
"Covered by feat/x's check." and CI passes with the same reason instead of
asking for a new check. More commits on top get questions about them only.

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

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

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
… everything else collapsed

- Worth discussing items (explained questions, partial passes) are posted as
  resolvable diff threads like flagged concerns (same marker, never twice,
  resolved when the item goes); the comment links to each.
- Concerns nobody flagged are listed once under 'Also noticed while grading'.
  In VS Code a concern can be dismissed with a one-line reason
  (concern.dismissed) or marked fixed once flagged; both take it off the MR.
- A flagged concern whose line is gone from the file says 'possibly fixed',
  in the comment and in its thread.
- The answers go into a collapsed 'All answers': each answer once; the
  grader's line, labelled 'Check:' and outside the quote, only when it
  corrects something.
- Tests follow the stacking of feat/second and feat/third on feat/limits.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
…-approval is named; settings changes are called out

- Advisory: 'Advisory: this check doesn't block merging.' is the comment's
  first line; no skip label instructions in the comment or the log.
- After the author approves their own MR the comment says 'Your own approval
  doesn't count: ask a reviewer to approve, then re-run this job.'
- An MR that changes understanding/config.json gets 'This merge request
  changes the understanding check's settings: mode required → advisory.',
  also when no code changed.
- The size is honest: '3 questions, ~2 min' instead of a fixed ~5 min.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
tomeyer97 and others added 27 commits November 18, 2026 09:12
walkDiff read every "--- x" / "+++ x" line as a file name, so after a removed
SQL/Lua/Haskell comment line (or an added Haskell "++" line) the change's code
lines, changed ranges and merge request line pairs went to a file named "x".
It now counts each hunk's lines and reads those prefixes as file names only
outside a hunk.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
A sentence under More was dropped when any one of the places it names was already on screen, so the other
places it named were lost. It is now dropped only when every place it names is on screen.

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

The check of whether an applied fix was still unsaved looked only at the
first open document with the file's path; a git: diff view of the file opened
before it hid the unsaved fix, so Commit & push did not offer it. One rule
(src/check/unsaved.ts unsavedIn) now serves both places: every open document
is looked at, and only the file itself (file: scheme) counts.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
A ref to a renamed file that is gone now was struck out under its old name, and a line past the end of the
step's own file under its written (unshifted) line, while what was checked and found gone was the new name and
the moved line. The struck-out label now reads as the place looked for, like a live link would.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
It was meant for the tour's position text ("3/13", "at 3/13", "13 steps"); that is drawn elsewhere: the
status bar item (renderStatusItem, on every updateTour), the Journeys view rows (journeyRows, 'step 3/14') and
the map (resume in mapMessages). Nothing read this copy, so it is removed rather than wired twice.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
The Check view's "no check here" used origin/HEAD (else main/master), while
background questions skipped main, master and the compared-with branch, so in
a repository whose default is develop or trunk the two disagreed. The CLI's
isDefaultBranch is now the one rule (origin/HEAD when known, else main/master,
plus the branch compared with); the status carries it as onBase and the
background writer uses that. With origin/HEAD at develop, a push of main or
master now gets background questions (unless it is the compared-with branch);
develop never does.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
"Ask about this step" used to open an empty comment thread on the step and remember (card.asks) what the
note should say it was about. It now asks in its own input box (askStep/askWithDraft) and saves through
noteOnStep, which adds the "About the step ..." line and the {journey: id, step, n} link itself (covered by
the journeyTour integration test). Nothing filled the map any more, so its lookup in noteFromNewThread could
never hit: both are removed.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
With several workspace folders holding a journey with the same id, a step lens, a question's back-link or a
check's See also could open the other folder's journey (openStepLink took the first journey with that id).
The links now pass their folder; openStepLink prefers the journey there (journeyById), else the active
editor's folder, else the first with that id (older links without a folder).

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
contextScore bounds-checked the lines before only at the top of the file and the lines after only at its end; with the
tracked start past the end of a shorter file, anchor() read past the end ("Cannot read properties of undefined").
Both sides are now checked at both ends.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
show() built the card's markdown for the new thread and then again to compare with what it shows, so a new
card computed its text twice and presented its code links and value hints twice (both dedupe, so the cost was
work, not visible). The markdown is now made once, after the settle checks, and used for both.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
folderName kept backslashes and a leading ./ (only a trailing / was dropped), while folderProblem accepts them, so the
notes were looked for and written under a path that doesn't match the folder. Both folder settings now go through one
helper, settingFolder (paths.ts): slashes, no trailing one, no leading ./, as the journeys folder already did.

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

runAsk's no-CLI way out reported the failure on the note it was starting but did not move the queue on (the
model-problem one did), so the notes queued behind it kept "Queued…" for good. It now runs the next one too: each
waiting note says Claude Code was not found.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
The filter picker has offered only All, Unread and Stale since 0.23 (who spoke last is each row's icon), and nothing
else sets the filter, so listedUnder's needsMe and waiting cases could never run. (cancelWriting, the other dead code
reported, was already removed in eff6871: a journey being written is cancelled from its progress notification.)

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

lenses() defaulted to activeTextEditor!, which is undefined while the journey panel has the focus (the flake); the
caret was moved in it the same way. Each call now names its file (lenses/codeLenses take the URI), and the caret moves
in the code group's editor.

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

A step animation with an actor type, id, status or screen named after an Object.prototype
key stopped drawing ("ACTORS[a.type].size is not a function") or gave a status chip the
class "function Object() ...". The same plain-object maps keyed by data held answer drafts
(question id), journey previews and stale steps (journey key) and box title heights (node
id). Lookups in fixed tables now go through Object.hasOwn; maps built from data have no
prototype. The scenes harness draws a scene full of such names.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
…"Failed to resolve tree node")

The Journeys and Code Notes views revealed rows VS Code's extension host did not hold yet or any more: the Continue
row of a journey just walked before the view was redrawn for it, a group right after a redraw had dropped every row
(its re-fetch then raced the window's own fetch of the same rows), and redraws fired in the middle of a reveal. Each
logged "[TreeView:codeNotes.journeys] Failed to resolve tree node".

viewSync.ts (ViewSync, one per view) notes which rows VS Code was given since the last redraw (through the provider
it gives VS Code; the tests' snapshots don't count), and reveals a row only when VS Code holds it and its parents, or
holds the parents and asks for its siblings only for this reveal (a folded parent not drawn yet); else it tries again a
few times. A redraw due while a reveal is under way waits for it. The Code Notes view's getParent no longer answers for
rows it never lists (journeys, a step's questions: the Journeys view lists those).

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
With prefers-reduced-motion, chat, queue and app messages still slid in, gauges grew,
file lines, queue slots, app rows, pods and the inner box still faded, a marked pod kept
pulsing, and the splitter faded on hover. The scene's reduced-motion rule now covers every
part of the scene (one rule instead of a list per part), and the splitter's fade is off.
The scenes harness steps through every beat once more with reduced motion emulated and
fails on any running animation or any non-zero animation or transition.

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

"A slow refresh never holds the event loop" failed under load: it asserted
the worst single stall stayed under 30 ms (26-360 ms measured, refactor or
not), and it could not see a stall at the very end of the refresh (the ticker
stopped before the late tick fired). It now compares the share of time the
loop was held with the same share measured idle in the same run, counts how
many slow-git-call lengths (0.25 s) the loop was held in long stalls, and
lets one more tick in before stopping. Checked: five synchronous git calls
after the refresh now fail it (held 1300 ms); the old check passed them.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
The flaky /Done\. Pushed\.$/ came from reading the status item right after
view() said "Pushed.": view() is the model computed live, so it can say so
part way through the refresh after the push, while the status item (like the
view on screen) is drawn at the end of that refresh. No stale status for a
user. The wait added in 6d28d12 is the fix; the comment now gives the reason.

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

The far-view check parsed colours itself and read color(srgb ...) values (0-1 numbers,
what color-mix() computes to: the outer names' and the blocks' backgrounds) as 0-255, so
those backgrounds came out near black. It now uses the harness's shared parser, which
scales them. Measured with it, every far name in the dark theme is at least 8.59:1.

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

VS Code's extension host applies a tree redraw that comes within 200 ms of the one before only when that time is up,
and drops the rows it gave the window meanwhile: a reveal of one of them still logged "Failed to resolve tree node"
(a walked journey's Continue row). ViewSync spaces its redraws further apart than that.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
…their context without racing the browser

The CLI reads GITHUB_HEAD_REF / CI_COMMIT_REF_NAME for the branch, so in GitHub Actions the tests' temporary repos
took the pull request's branch name (47 failures). The scenes harness's page-close handler closed its context
unawaited, and browser.close() could dispose it first: an unhandled rejection at exit.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
@tomeyer97
tomeyer97 merged commit f24f042 into main Nov 18, 2026
0 of 2 checks passed
@tomeyer97
tomeyer97 deleted the fix/bugs branch November 18, 2026 10:14
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