Repository navigation
Fix/remote perf - #3
Merged
Merged
Conversation
…ill that writes the questions ci never calls a model: it checks the file belongs to this change, every question is answered and graded, the recorded verdicts meet the threshold and match their hashes, and keeps one merge request comment up to date (GitLab and GitHub). Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
…s a design choice Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
…threads, graded answers, a small panel The questions are notes of a session "Check · <branch>" on their lines (a file note for a design question). Ctrl+Enter in a question's thread saves the answer to understanding/<branch>.json and has Claude grade it; the reply lands in the thread. The panel lists the questions with their state and, once the threshold is met, one next step with a Stage button. A status bar item offers the check or shows 2/4 ✓. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
… built; 0.24.0 Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
…ll dropped the bit Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
… from one list model src/journeyList.ts computes the rows (area, age, In this file, search hits), the panel's sections and the sidebar's groups in one call. The Journeys view (its own, beside Notes) lists In this file, Continue, Checks and All by area, with Search and Sort. The panel's home gets a search field, kind and area chips, a sort, sections, and a preview of the selected row (a small map and its steps); a click selects, Enter or Walk starts it; ↑↓, →, Esc and / work from the keyboard. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
… lines read, a miss never gives the answer away A pass is "Right." plus at most one new fact (25 words); a first miss says which part is right and what to re-check, with a hint that points at lines; an explanation is at most 60 words. Every claim about the code must come from lines the grader read, and it re-reads its draft for contradictions before replying. test/grader/eval.mjs grades fixed answers with a real model on a refunds change built from two cases where a grader once said wrong things (a sent-at stamp that is never read; an idempotency lookup before the lock with a UNIQUE constraint). Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
… a comment when the check is missing, Commit & push, "also" ranges
- relocate(): question lines (and their also ranges) follow their anchors; status shows today's lines, CI's
comment uses the head's.
- MR comment refs link to the merge request diff at that line (GitLab line codes, GitHub diff anchors); the file at
the head commit when it is not in the diff.
- No check yet: the job posts (or updates) the MR comment in two lines; the log adds what it is for, a path without
VS Code and "then re-run this job" for the skip label. A pass says the room left: "2 lines since the check
(top-up at 40)".
- commitAndPush(): commits only the given files, pushes with -u when there is no upstream, and says a rejected push
or a failed sign-in in one line (git without a shell, no credential prompt).
- checkState(): where the check stands in one word, with drift from uncommitted lines too.
- Skill: --also path:A-B for unchanged code an answer needs; lists ("X, Y and Z") and "…, and could …" refused.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
…ements; the grader's note, topic and concern - The rule: every question done (pass, explained or restated; a "try again" is not done) and ceil(60 %) passed. summarize() gives done/passed/met; ruleText() says it in one phrase: "Answer all 5 · 3 must pass". - checkState(st, cfg): nothing | none | invalid | drift | answering | short | done, with the counts. stateWords(state, st): the same short words for the status bar, the Check view, `status` and CI. localStatus() returns `state`; the file's result.verdict follows the rule. - Verdict "restated": after an explanation the author may say the point back in one sentence; a right one counts as passed. gradeMode() picks the restatement prompt. - The grader also returns "note" (15 words for the reviewer), "topic" (explained) and "concern" when the code itself looks wrong. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
…ements; the grader's note, topic and concern - The rule: every question done (pass, explained or restated; a "try again" is not done) and ceil(60 %) passed. summarize() gives done/passed/met; ruleText() says it in one phrase: "Answer all 5 · 3 must pass". - checkState(st, cfg): nothing | none | invalid | drift | answering | short | done, with the counts. stateWords(state, st): the same short words for the status bar, the Check view, `status` and CI. localStatus() returns `state`; the file's result.verdict follows the rule. - Verdict "restated": after an explanation the author may say the point back in one sentence; a right one counts as passed. gradeMode() picks the restatement prompt. - The grader also returns "note" (15 words for the reviewer), "topic" (explained) and "concern" when the code itself looks wrong. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
…t never grades in public
- No minimum size: every change with code lines gets a check. The question count follows the size in code lines
(1 under 20, 2 under 60, 3 under 150, 4 under 300, else 5; config "questions" bounds it, or a number fixes it).
questionsFor() gives the count, the skill's `finish` holds to it.
- Code lines only: blank and comment-only lines (by file type: // # /* */ -- <!-- -->) count neither for the change
nor for drift. Config "exempt" globs (empty by default) leave files out. "✓ No check needed: only exempt files
changed (…)" / "no code lines changed".
- The merge request comment follows the job: in progress it is one line ("2 of 4 done"); drift, a damaged file or
changed answers say what is missing in the first line; once done, the Q&A: explained questions as "Worth
discussing: <topic>" with the correct point in one line, others with the passing answer and at most 15 words of
the grader's reasoning. Wrong guesses and retries are never posted, no ✗. Flagged concerns
(questions[].concern { text, flagged }) come first as "⚠ Found during the check".
- Skip: any approval by someone other than the author waives the check ("✓ Understanding check waived (approved by
@bob)"); a self-approval and a missing approval say what to do; GitHub messages don't ask for a re-run. A label
without an approval no longer fails a check that passes by itself.
- Skill script: never asks about comment-only lines; --suspect marks a question whose honest answer is "a bug";
"…, or does…", "…; what…" and "as well as" are refused like "…, and could…".
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
…in the comment - Drift compares the branch's own change against today's merge base with the change the questions were written for (same file, sign and code text). Merging or rebasing onto the target branch counts nothing; lines resolved differently in a conflict do. Without the questions' commits (a fresh CI clone after a rebase) the sizes in code lines are compared, as before. - A top-up asks only about those new lines of the branch's own change (listed as NEW LINES), never about code that came in from the target branch, and records today's merge base. - The merge request comment says "answered by <name>" when a question was answered by someone other than the check's author. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
…e comment that follows the job, concerns
- Design doc: the ci rules and messages in order, the comment (progress line, Worth discussing, Found during the
check, answered by), drift on the branch's own change, the questions[].concern { text, flagged } and suspect
fields, the question count by code lines, exempt globs.
- SKILL.md: the count comes from `start`; never ask about comments; --suspect when the honest answer is a bug;
a top-up asks only about the NEW LINES.
- README: every change with code gets a check; exempt example; any reviewer approval waives.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
…e either Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
…e everywhere, restatements, concerns, Commit & push - The check list is a Check view in the Code Notes sidebar (above Journeys), not an editor group: the code keeps its width and the list survives a reload. When questions are ready only question 1 opens (its code and thread). - Status bar, view and CLI say the same thing (checkState / stateWords): "Understanding 2/5", "✓ Understanding", an amber "Needs top-up: +46 lines" before the push (uncommitted lines count too). The view follows the branch: "No check yet for <branch>" with Start. - After a grade the editor stays on the question and its thread stays open (a check thread is never "resolved" in the editor, which collapsed it). A pass moves to the next open question; Ctrl+' D / A walk the open questions; view rows work with the arrow keys. - After an explanation the author may say it back in one sentence: a right one counts as passed. - A concern from the grader (or a question marked as a possible bug) shows on its row and in its thread with Fix now (Ask Claude, a note of its own), Add TODO and Flag for reviewer. - Commit & push commits only the check file (and .gitattributes when it offered that line) and pushes, -u without an upstream; a rejected push or a failed sign-in is one line. - Ask Claude on a check question asks what you want to know and answers in a separate note; Claude is told to explain the code and never write the check's answer. - Journey lenses step aside in a file with an open question; another branch's check threads are hidden. - Questions finished after a branch switch wait in .code-notes/.cache (git-ignored) and are put in place when the branch is checked out again. Before writing: one line about uncommitted or new files (Commit first, Continue). - First check: offer to collapse understanding/*.json in merge request diffs (.gitattributes). - Someone else's check: "Ada's check: your answers are recorded as yours." - Integration suite: real clicks and typing over DevTools, the workbench console checked for errors. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
…keeps the folder out of diffs and search - pruneAttempts(question): every attempt but the last loses its answer text and hash (verdict, times, by, feedback, note, topic and concern stay). addAttempt calls it, so `grade --answer` and the extension never keep a wrong guess. validateCheck needs an answer only on the last attempt; hashProblems verifies only last attempts. Older files, with every answer and hash, still validate. - `understanding.mjs init`: adds `understanding/*.json linguist-generated=true` to .gitattributes and `understanding/` to .ignore in the repository's root, idempotent, never a line twice. - Doc and README: the format with superseded attempts, the "Out of diffs and search" section. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
…exempt globs, question count by size The extension follows: the question-level concern (one per question, its flag kept across grades; Add TODO clears it), no "small" status, a test change large enough for three questions. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
…mpts (checked lightly) Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
…itory quietly; superseded answers are not kept - Merged: initRepo, pruneAttempts, links on questions and attempts in validateCheck. - The first check calls initRepo (.gitattributes collapses check files in merge request diffs, .ignore keeps them out of search tools; both are committed with the check by Commit & push) and adds understanding/** to this workspace's search.exclude. No prompt; the Check view says once: "Answers are kept out of search." - An answer a later one superseded is dropped from the check file (pruneAttempts) and from its thread's note. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
…y step cards
- The grader names code ("the catch block in tick()") and links it: links [{phrase, file, start, end}], 1-3, never
line numbers in prose. A link whose phrase isn't in the text, or whose lines aren't in the file, is dropped; the
words stay. Stored on the attempt.
- The skill's add takes --links "phrase@path:A-B;…" (checked the same way); SKILL.md asks for linked phrases. An
"also" range a phrase links to needs no "Also read" line.
- Threads render the question's and each grade's links like a step card's: a tinted phrase; a click lights the lines
and brings them into view; "← Back to question" returns to the thread. One markup helper for both.
- Check view rows show file names only.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
…hread, never into the code under it
VS Code's API can't place the caret in a thread's reply box, so the thread itself takes focus ("Focus Comment on
Current Line"): typing can no longer edit the file. Tested with real typing over DevTools.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
…e branch changes A step was located when the window opened and then only while its file was open in an editor. After a checkout, pull or an agent's edit, a step stale on the previous branch stayed stale, and one the new code changed stayed fresh, so the list showed a mix of both branches' staleness. Files with steps that change on disk now re-locate those steps; a branch switch re-locates all. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
… only for the question it was shown with; rows keep their question - Each thread records a hash of its question's id and text. When the file is rewritten (questions written again, a pull, a checkout), a thread whose question changed is made again and one whose question is gone is removed. - An answer is stored only if its question is still the one the thread showed; else "This question changed: reopen it from the Check view." - In the notes list a check row is always its question's text; ✓ only for passed (or said back), an icon for try again and explained. No Resolve/Reopen on check threads (grading does that); Ask Claude is offered on them. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
…ir lines - GitLab: a discussion with a diff position (diff_refs; the question's last line, else the first code link of the attempt that raised it). GitHub: a review comment on the line (side RIGHT). - "⚠ Found during the understanding check (flagged by <name>)" and the concern, mentions inert, a hidden marker with the question's id: never posted twice, updated when the text changes. - Un-flagged or its question gone: GitLab resolves the discussion (reopens it when flagged again); GitHub gets one "No longer flagged." reply. A line outside the diff falls back to a plain note with a file:line link. API errors are logged and never fail the job. The concerns stay at the top of the bot comment too. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
…resented as a safeguard; eval cases for a restatement and a concern The eval (test/grader/eval.mjs, a real model) gains R1 (an explanation said back in the author's words: restated) and C1 (low-stock alerts mailed to an OPS_EMAIL the worker's deployment never sets: the grader must raise a concern), and checks that no reply points at code by line numbers. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
- Config "mode": "required" (default) | "advisory"; UNDERSTANDING_MODE overrides it. Advisory: wherever ci would
fail it exits 3 and its first line starts "⚠ Advisory:"; the merge request comment is the same plus one muted line
"Advisory: this check doesn't block merging." Exit 2 stays 2.
- GitLab recipe: allow_failure: { exit_codes: [3] } ("passed with warnings"). On GitHub Actions, which can't allow one
exit code, ci prints a ::warning:: annotation and exits 0.
- README and design doc: start advisory for a month, then switch to required.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
…rted window; the shot before the measure Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
…rch, reload) Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
…ch switch; plain words when Claude can't help
- Done: the button sits right under "✓ Done" (it used to be below the questions, out of sight) whenever something
is left to commit or push ("Push" when only the push is), naming what it commits. When only the tour is new, the
line says so and the status item doesn't push it by itself.
- A fix applied from a concern is said on its question ("fix applied · not committed") and goes with Commit & push,
ticked and saved first; that commit marks the concern fixed.
- After a push the status bar says what went, the status item uses the view's words, and a branch without a merge
request gets the link the push printed (the CLI's commitAndPush returns it).
- A merge: Keep both keeps a flag either side set (said in one line); Commit & push concludes the merge; a rejected
push offers Pull; a stale push line goes once HEAD moves.
- The tour reads the code checked out: a branch switch stops it and it goes on once the branch is back; the log has a
short reason, never Claude's reply. An untracked, unfinished check no longer follows to other branches.
- Claude Code not found: Install, its path, Retry, and that an approval waives the check. Grading outages use the
writing words ("your answer is saved") with Try again, also under the answer box. A broken file names the line and
Open goes there; its tour stays. Nothing committed yet: Commit… first, the check starts once it is in. A reload
while writing says "Writing questions", not Start.
- Add TODO writes the whole concern, wrapped, inside the block a "} catch {" line opens.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
… stops VS Code marks the document dirty only after the change event, so the edit is recorded either way and dropped when the document is still clean (a reload from disk). The main suite's step code link expects the caret to stay put. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
…applied Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
# Conflicts: # test/integration/check.ts
…ck to where you were; vscode://…/link for terminals Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
…tour of a change Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
… answering in Claude Code; commit from the CLI - Documentation and a change that only removes code get 1 question; a removal's is an evidence question from a search of the repository (confirmed or corrected in one sentence) - Tours: 2-3 steps for docs, a removal or under 10 code lines (docs that describe no code: none), then 2-4, 4-7, 6-10; journey.py checks the tour's own count - grade --prompt / --reply, next, flag: the skill asks one question at a time in a chat and records each grade as the extension does; no model call from the CLI - commit: the check file, plus the .gitattributes line on a repository's first check - init also puts the skill in .claude/skills (and the CLI beside it unless the repository has one) Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
…arate Set up offer; no tour for docs that describe no code
- Commit & push writes the .gitattributes line on a repository's first check and says so on its
line ("also folds check files in diffs (.gitattributes)"); the Done state's Set up · Hide is gone
- A docs check whose text describes no code has no tour, and nothing offers to write one
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
…k commit folds check files in diffs; /understanding-check in Claude Code Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
…14, so any journey.py version works The extension may hand the skill an installed code-journey skill whose journey.py has no --steps. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
…hing changed The check's refresh ran about a dozen git commands through execFileSync in the extension host. Every one of them stopped the whole host (every extension sharing it; over Remote-SSH the editor waits on it too): blocks of 120-180 ms on a busy repository, over a second on a big one. - The CLI's functions that run git (localStatus, codeStat, diffText, identity) run on a worker thread (cliWorker.ts, cliThread.ts). The same CLI file, the same code: the terminal and CI keep running it as before, synchronously in their own process. The worker never takes git's index lock (GIT_OPTIONAL_LOCKS=0), so reading the status writes nothing another watcher sees. - A refresh first works out a key of what it reads (HEAD, the refs, the working tree's changed files with their size and time, the check file, its settings and cache, the files of an applied fix or a TODO, git's config): the same key and nothing is read again. Every trigger goes through it; a direct refresh (an answer, a write) is always in full. - No watcher on .git/index: any `git status` rewrites it (VS Code's Git view, a terminal, an agent), and the check kept refreshing itself. VS Code's git extension says when the working tree changed instead. - Refreshes are coalesced: while one runs, the calls made meanwhile share one more. - The default branch and git's identity are kept while the files they come from are unchanged. - The status bar item, the checkOpen context key and the Check view are only sent what changed. - The grader's diff and the check that a parked check file is tracked no longer block either. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
At activation (and when the setting or the trust changes) every candidate's `claude --version` ran through spawnSync, a few hundred ms each. It now runs as separate processes, all at once (findCliAsync), filling the cache the synchronous findCli answers from. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Each setContext and each setDecorations is a message to the window (over Remote-SSH, over the network), and they were sent again with the same value all the time: the caret's context key on every caret move, seven context keys on every refresh of the notes, and every part's decorations on every editor of every visible-editors change (mostly nothing, on editors that had nothing). - setCtx sends a key only when its value changed. - setDecorations (decorations.ts) sends only what changed per editor and type; after an edit of the document everything is sent again (the window moves decorations with the text). - The function around the caret (the Journeys view's "In this file") no longer asks for the document's symbols on every caret move in a file that has none: asked again after 10 s or an edit. - The workspace file watcher's per-event check for journey step files looks the path up in a set (rebuilt when the journeys change), not through every step of every journey. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
…inks, questions Over Remote-SSH each awaited call into the window is a network round trip, and the actions people repeat all the time awaited several in a row, plus fixed sleeps. - A step (next, previous, Back to step N) no longer waits for its place to be stored (two round trips): the stored value is in memory at once, and the write goes on in the background. - A code link into a file already on screen, with the keyboard kept where it is, shows nothing (no call at all), only lights and scrolls; otherwise the caret goes with the show (one call, not a show and then a caret). - The 120 ms sleep before deciding whether the way back is needed is gone: a link into another file needs it at once; in the card's own file the decision waits for the scroll only when there is one. - Chat links, a note's thread and the check file open with their caret in the same call. - Reading a note stores both read markers at once and no longer holds up opening it. - The code lenses are refreshed only when what they show changed (each refresh makes the window ask every editor). - Add Note waits until the window has asked for the commenting ranges again (still at least 120 ms): on a slow link the fixed wait was too short and the box didn't open. test/integration/roundTrips.ts counts the calls into the window per action (CODE_NOTES_RTT_MS=<ms> also delays each and prints how long the actions take) and checks each action waits for at most one. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
…pening a question - `git status` (five times) refreshes nothing; ten triggers at once check the key at most twice and read nothing; ten forced refreshes at once are at most two; a commit is picked up without being asked. - With git made slow (a `git` that sleeps 0.25 s a call while .git/slow-git exists), a refresh takes seconds and the extension host's event loop is never held up more than 30 ms. - Opening a question: one show, at most three calls into the window. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
… a command A link from a Claude Code chat is recognised by its selection: whole lines, column 0 to column 0, set by another extension. VS Code up to 1.14x reports such a selection with no kind; newer versions report Command, so no link was ever lit there. A Command selection now counts too, but only for more than one whole line in an editor that has just come to the front (Go to Definition and the like select a name, not whole lines). The extension's own selections (a step's caret, a note opened, a beat, a code link) are marked and never count. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
…window From a remote extension host (Remote-SSH) every vscode.workspace.fs call goes to the window and back: one network round trip each, one after the other. Activation read every note and every journey that way (175 notes and 17 journeys: 91 s and 145 s at a 250 ms round trip), and a step read its file, its references and their state the same way before anything showed. - wfs.ts: the same calls, with Node's fs for `file:` URIs (the extension host runs where the files are); other schemes still go through workspace.fs. Every workspace.fs call of the extension uses it: notes and journeys read and written, the README and .gitignore, the trash, moves, stats, the folder change. - A step's code links and value hints read their files' lines without opening documents (an open document, else the disk): no round trip per file. - A code link names where its Back goes in its tooltip, and has no Back row when there is nowhere to go back to. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
…t says - Each step opened gets a number: one superseded by a step opened since (another step or journey: Next, Next, Next at a slow link) stops at its next wait and never shows its file, its caret or its card. The same step opened twice still finishes both (its file had moved). - The status bar's step and the map move first, before the step's file is read. - A code link's way back is the step being walked when it was clicked: "Back to step N" names it (and the journey, in its tooltip), goes only there, and goes away once the walk moves on or ends. A link clicked outside a walk has no Back row (a bare "Back" stayed for good and later led to whatever step was being walked then). - The step card's thread is made with its comment (never an empty thread for a moment). Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
- A reply, a new note and a check answer are taken once: a second Ctrl+Enter while the first is being saved (the box
keeps its text until the command returns, up to seconds over a slow link) posted it twice.
- A reply is in its thread at once, before it is saved (taken back if the save fails); reading it and marking new
notes read are stored in the background.
- A check answer says "Checking…" (the view and the thread) before anything is awaited; the box empties at once and
gets its text back if the answer was not taken. Its refresh first is gated (reads git again only on a change).
- A note opened from the list while its editor is still opening (its thread is made only once the editor has
settled) is made expanded, not collapsed.
- The list's row is revealed after the list was redrawn for the note, and the list redraws once per burst of changes:
VS Code asked for a row of a drawing the next one had dropped ("Cannot resolve tree item").
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
The zero-width space after every word of a link's phrase was copied with the text, so copied words no longer matched a search. VS Code's comments allow <wbr>. Also in the tests: the round-trip count includes reading every note and journey (no call to the window for files); a link from a chat is driven as the chat opens it (its lines selected with the show), and whole lines selected in an editor already in front are not a link; R waits until the file it starts from is settled in front. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
- The round-trip probe walks a journey with steps in several files, sends three Nexts at once and checks the walk ends on the last one with its file and its card. - A unit test fails on any vscode.workspace.fs call outside wfs.ts (each one a round trip from a remote extension host), and on a document opened only to read lines in the code links and value hints. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
…chat links Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
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
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.
No description provided.