Chat font and max-output-token settings, plus a settings save that no longer drops keys - #225
Open
leonidasbarkas98-cpu wants to merge 26 commits into
Open
leonidasbarkas98-cpu wants to merge 26 commits into
leonidasbarkas98-cpu wants to merge 26 commits into
Conversation
Adds a fork-issue-27 context indicator to the status bar, together with the Claude subscription usage: five-hour window and weekly utilization percentages with reset times, fetched from the unofficial OAuth usage endpoint using the CLI's own OAuth token, throttled to one request per five minutes and failing silently (stale or hidden display) on any error. A new rate_limit_event case additionally captures the five-hour reset time for free from the CLI stream. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
The usage indicator's _usageLimits type gains sevenDayOpus/sevenDaySonnet fields, and the oauth/usage response parser now reads the per-model-tier weekly buckets seven_day_opus (shown as Fable) and seven_day_sonnet through the existing defensive parseWindow, alongside the five-hour and week windows it already handled. getUsageIndicatorHtml renders both new buckets in the status-bar usage indicator when the account's usage response includes them. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
…ue-35) The oauth/usage endpoint now returns per-model weekly windows only as limits[] entries (kind "weekly_scoped" with a model scope, field "percent", 0-100 scale); the legacy seven_day_opus/seven_day_sonnet fields are null. Fall back to those entries, preferring is_active ones. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
…ssue-38) WSL-to-Windows path mapping, case-insensitive workspace-relative path computation (undefined for the root itself and anything outside), NUL-based binary detection and the deterministic claude-diff URI build/parse pair. vscode-free so the 21-case mocha suite (test:diff-utils) runs headless. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
…point (fork-issue-38) Replaces the racy fileContentBefore diff (button hidden on tool_result, one-slot pending data, per-click Date.now tabs, global renderSideBySide mutation, dead openDiffByIndex path) with vscode.diff against the shadow backup repo: left side is the checkpoint before the turn (git show via execFile, 16MB buffer, 2MB/binary guards, BOM strip), right side is the real editable file; stable per-(sha,path) URIs dedupe tabs and the provider re-resolves restored tabs itself. The Open Diff button stays visible permanently (dataset-based, works after history load via the persisted showRestoreOption sha). Auto-opens once per file per turn after a successful Edit/MultiEdit/Write with preserveFocus, silently skipping files outside the workspace (scratchpad/memory edits are routine); gitignored paths never fake an all-new baseline. Configurable via claudeCodeChat.diff.autoOpen (default on) in settings and the modal. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
…the 5-minute cache (fork-issue-54) Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
… check (fork-issue-50)
_commits is cleared when a history load switches conversations, but the
replayed showRestoreOption messages keep their Restore buttons alive, so
_restoreToCommit answered every click with "Commit not found". On a
_commits miss, confirm the sha directly against the shadow backup repo
(execFile git cat-file -e <sha>^{commit}) and rehydrate the display info
from the replayed entry; shas are format-validated before reaching any
git command. Live-session behavior is unchanged, and a genuinely missing
commit still reports "Commit not found". New pure helpers in
restore-commit-utils.ts with 15 unit tests (test:restore-commit-utils).
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
extension.ts's _usageLimits comment still described the opus-lineage bucket as covering "the Fable model shown to users" -- there is no Fable model in this branch (_setSelectedModel only accepts opus/sonnet/default), so that clause is just wrong; dropped it. Also fixes the usage-limits status line (src/script.ts): the German "Wo" abbreviation becomes "Week", and the sevenDayOpus bucket's "Fable" label becomes "Opus" (it's the Opus-tier weekly window, mirroring the existing "Sonnet" label for sevenDaySonnet) instead of naming a model that isn't selectable in this branch. Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>
getContextWindow() claimed a 1M-token native window for opus/sonnet (and a fable entry that isn't even selectable in this branch) instead of the real 200K, so the status-bar Ctx percentage under-reported usage. Removed the fable entry and reset opus/sonnet to 200000. Also cleaned up references that only made sense inside the private fork: two comments pointed at shell-utils/auto-model-switch, a module already removed from this branch and one that never existed here; a few inline code-review shorthand notes were shortened; example paths using a real name became generic placeholders; and bare #NN issue markers now read fork-issue-NN so they aren't mistaken for this project's own issue tracker. Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>
…line The fork-issue-27 context indicator (getContextIndicatorHtml) was reading a currentContextTokens variable that was never assigned anywhere, because the extension.ts half of the plumbing (the _currentContextTokens field, computing it per turn, and sending it on the updateTokens postMessage) had been dropped. The function silently always returned '', which had quietly replaced the previous "N tokens" status-line display with nothing for non-subscription users. Restores the field, the per-turn ctx calculation (input + cache read + cache creation tokens), the three reset points (compact boundary, new session, loaded conversation), and the webview-side assignment/resets in script.ts. Also adds the missing .ctx-warn/.ctx-crit CSS rules the 80%/95% warning classes rely on, which existed nowhere in ui-styles.ts.
A first pass missed several inline code-review shorthand notes sprinkled through the fork-issue-38 turn-diff code. Neutralized all 15 spots, keeping the substantive explanation in each comment.
_permLog (fork-issue-15) writes unrotated plaintext to os.tmpdir(), so the auto-skip diagnostic for _openTurnDiffFallback was leaking the full, potentially sensitive file path on every skip. path.basename(filePath) is enough to diagnose which file was involved without exposing the path.
The diff-content-provider comment justified skipping per-panel routing by pointing at a "New Claude Chat (Separate)" (fork-issue-24) feature that doesn't exist in this branch (no such command in package.json, no second panel type) -- reworded to describe what's actually true: a single shared ClaudeChatProvider instance backs both the panel command and the sidebar webview, so there's only ever one provider to resolve against regardless of who opened the tab. Also removes two blank lines that an earlier commit in this branch introduced incidentally (one in extension.ts before the rate_limit_event usage-limits refresh, one in script.ts before renderEnvVariables) and that don't match the surrounding code's spacing.
…eak, fix stale cat-file comment Review-round follow-up: KNOWN_ENDPOINT_MARKERS was declared but never read -- _isOpenCredits() kept checking the same two literals by hand. Now _isOpenCredits() actually uses the constant, so there's a single place to add a future endpoint marker. _openTurnDiffFallback's auto-skip perm-log line still leaked the full backup-repo/file path through its reason= field even after the earlier basename-only fix for file=: _resolveTurnDiffBaseline wraps raw execFile failures whose message starts with the full command line. reason now goes through the same basename collapsing before logging. Also corrects a comment that misattributed a "plain cat-file, exit 1" contrast to _isPathIgnoredInBackupRepo (which uses `git check-ignore -q`, not cat-file) instead of the actual unpeeled `git cat-file -e` case it meant to describe. Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>
…ssue-42) New setting claudeCodeChat.advanced.maxOutputTokens (integer, 0 = CLI default) feeds the env variable on both the native and the WSL spawn path. When the CLI reports 'API Error: ... output token maximum' - it arrives as assistant text, verified against the live CLI - the chat shows a hint with a button that opens the settings UI filtered to the new setting. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Two new settings under claudeCodeChat.ui apply via CSS variables set with style.setProperty (no markup interpolation). The .messages and .input-field font-size rules, and the per-bubble font-family rules (user, tool-result, thinking), respect the variables so the custom font stays consistent, and the input height re-measures when the size changes. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
…modal (fork-issue-44, fork-issue-42) The three settings already exist in the manifest; this surfaces them in the gear modal via the existing updateSettings/settingsData roundtrip, clamps them like the manifest does, and adds claudeCodeChat.advanced to the config-change broadcast so a second open surface cannot overwrite a fresh value with a stale one. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
_updateSettings ran the whole settings batch from the webview in one loop wrapped in a single try/catch, so a rejected config.update() -- e.g. a setting not yet registered right after a version bump -- aborted the loop and silently dropped every key that came after it in the same batch. In the order the webview sends them, a rejection on e.g. advanced.maxOutputTokens would have dropped environment.variables, router.enabled, diff.autoOpen, ui.fontFamily and ui.fontSize with no indication in the UI. Each key now gets its own try/catch through the new vscode-free settings-batch module: failures are collected instead of aborting, _sendCurrentSettings() and the balance refresh run even on partial failure, and one summary error message names every failed key plus the first error's reason. Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
…rk-issue-56) applySettingsBatch's onSettled callback was never wired up by any real caller (extension.ts's _updateSettings only ever passes two arguments) -- only its own tests exercised it. The "review follow-up" comment that justified it was also wrong: _updateSettings writes no perm-log line, and the refreshSettingsOnConfigChange function it names does not exist anywhere in this codebase. Dropped the parameter, its tests, and the misleading comment.
…tings-modal commits Bare #NN in comments/suite titles auto-links to this fork's own GitHub issues instead of our private tracker. Requalified as fork-issue-NN in extension.ts, script.ts, settings-batch.ts and its test suite (including the suite title, which shows up verbatim in the mocha reporter output). Genuine "upstream andrepimenta#150" references are left untouched.
That commit's message says it "adds claudeCodeChat.advanced to the config-change broadcast", but it only adds a key to the plain settingsData payload _sendCurrentSettings already sends. There is no onDidChangeConfiguration/affectsConfiguration listener for it -- the only one (activate(), above) still filters on claudeCodeChat.wsl only. 6257620 isn't the tip of this stack, so a code comment corrects the record instead of an amend.
…ling-file refs The module comment and toErrorMessage's comment both claimed extension.ts's injected updateSetting callback (and _updateSettings) call _permLog as a side effect; neither does -- _permLog's only call sites (extension.ts:3893, 3971, 4304) are the unrelated usageLimits/turndiff logging. Named the real caller (_updateSettings) instead. Both files also pointed at "shell-utils" and "perm-log-redact"/ "markdown-restore" as sibling test-pattern files; none of those exist in this branch, only restore-commit-utils does. Corrected the references, and fixed the malformed-input test's title, which named a "extension.ts's key=<batch> marker" that doesn't exist -- the actual outer catch in _updateSettings just logs and shows an error message.
settings-batch.ts's module comment and the matching test fixture both invented a "real-world hit on 2026-07-26" involving ui.renderMath and six settings (font family/size, completion popup/sound, send-on-enter) that don't exist anywhere in this branch's package.json -- the only settings registered here are wsl.*, thinking.intensity, permissions.yoloMode, executable.path, advanced.maxOutputTokens, environment.*, router.enabled, diff.autoOpen, ui.fontFamily and ui.fontSize. Replaced the illustration with the actual key order _updateSettings receives from the webview, so a rejection on advanced.maxOutputTokens is shown dropping the real keys that follow it (environment.variables, router.enabled, diff.autoOpen, ui.fontFamily, ui.fontSize). Updated the test fixture accordingly.
…setting settings-batch.test.ts's first suite plumbed 'ui.compactMode' through applySettingsBatch, but that setting doesn't exist in this branch's package.json (only wsl.*, thinking.intensity, permissions.yoloMode, executable.path, advanced.maxOutputTokens, environment.*, router.enabled, diff.autoOpen, ui.fontFamily and ui.fontSize do). Swapped it for advanced.maxOutputTokens, already used the same way in the second suite below. Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>
…s never opened
enableYoloMode() (the button on every permission-error banner) flips
the yolo-mode checkbox and calls updateSettings(), which now also
reads max-output-tokens, chat-font-family and chat-font-size. Those
three fields are only populated by the settingsData roundtrip that
runs when the gear modal opens; clicking Enable Yolo Mode without ever
opening the modal left them at their empty '' default, so
parseInt('') -> NaN got clamped to 0 above and was sent as
advanced.maxOutputTokens/ui.fontSize, silently resetting a real value
back to 0.
updateSettings() now only adds these three keys to the settings batch
it sends when their DOM element actually holds a value.
Co-Authored-By: Claude Sonnet 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
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 top of #223 — please review/merge that one first; only the last 12 commits (from
9c03a2eonwards) are new here. Everything below9c03a2ein the diff belongs to that PR.This branch bundles three settings the chat UI was missing and the fix for a settings-save path that silently dropped keys:
claudeCodeChat.advanced.maxOutputTokenssetting (integer,0= CLI default) feedsCLAUDE_CODE_MAX_OUTPUT_TOKENSon both the native and the WSL spawn path. When the CLI reportsAPI Error: ... output token maximum— which arrives as ordinary assistant text, so it is matched there — the chat shows a hint with a button that opens the settings UI already filtered to that setting, instead of leaving the user to guess which env variable the CLI wanted.claudeCodeChat.uiapply through CSS custom properties set withstyle.setProperty, so nothing is interpolated into markup. The.messages/.input-fieldfont-size rules and the per-bubble font-family rules (user, tool-result, thinking) all read the variables so a custom font stays consistent across bubble types, and the input box re-measures its height when the size changes.updateSettings/settingsDatamechanism and are clamped the same way the manifest declares.claudeCodeChat.advancedis added to the config-change broadcast so a second open surface cannot overwrite a fresh value with a stale one._updateSettingswrote the whole batch from the webview in a single loop inside one try/catch, so a rejectedconfig.update()— for instance a setting not yet registered right after a version bump — aborted the loop and silently dropped every key after it. In the order the webview sends them, a rejection onadvanced.maxOutputTokenswould have takenenvironment.variables,router.enabled,diff.autoOpen,ui.fontFamilyandui.fontSizewith it, with no indication in the UI. Each key now gets its own try/catch through a new vscode-freesrc/settings-batch.ts: failures are collected rather than thrown,_sendCurrentSettings()and the balance refresh still run on partial failure, and a single summary error names every failed key plus the first error's reason. Covered by a mocha suite (npm run test:settings-batch).enableYoloMode()— the button on every permission-error banner — flips the checkbox and callsupdateSettings(), which now also reads the three fields above. Those are only populated by thesettingsDataround-trip that runs when the gear modal opens, so clicking Enable YOLO Mode without ever opening the modal left them at their empty''default:parseInt('')→NaNwas clamped to0and sent asadvanced.maxOutputTokens/ui.fontSize, resetting a real configured value.updateSettings()now only includes those keys when their DOM element actually holds a value.The remaining commits are review follow-ups on the above, not new behaviour: a
settings-batchcallback parameter that no real caller ever passed was removed together with its tests and the comment that justified it (the comment described a log line_updateSettingsdoes not write and a function that does not exist in this codebase); an overstated claim about the config-change broadcast in a code comment was corrected against the one listener that actually exists; and test fixtures/comments referencing setting keys and sibling modules that do not exist in this branch were replaced with real ones. Comments referencing our fork's internal issue tracker are written asfork-issue-NNso they do not auto-link to unrelated issues on this repository.This is one of several PRs from our fork submitted together as part of a broader, thematically grouped series (permission/session handling, settings/UI, security hardening, rendering, etc. split across separate branches/PRs). Note that
src/settings-batch.tsalso exists, with a different API and a different purpose (workspace-then-global write fallback), in the security-hardening PR of this series — an add/add conflict is expected if both are merged, and the two are straightforward to reconcile since neither imports the other.tsc --noEmitclean, no new dependencies; full unit run at this tip is 80 passing, 0 failing.🤖 Generated with Claude Code