perf/fix: startup performance, subagent discovery reliability, and sidebar polish#3
Open
TheEndless wants to merge 97 commits into
Open
perf/fix: startup performance, subagent discovery reliability, and sidebar polish#3TheEndless wants to merge 97 commits into
TheEndless wants to merge 97 commits into
Conversation
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
…model/turnCount/contextPct
Adds buildActivityHistory to collect the last ≤3 meaningful events (newest-first) as ToolCallSummary[], using the same isNoise/labelFromEvent logic as deriveActivity. Updates test suite to cover the new function and align Agent construction with the current type shape. Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
…tPct Adds extractSessionMeta to extract model name (with claude- prefix stripped), assistant turn count, and context window usage percentage from tail events. Updates buildAgent to populate Agent.model/turnCount/contextPct fields and builds activityHistory. Updates webviewProvider to serialize the new fields. Adds 5 unit tests for extractSessionMeta: model prefix stripping, turn counting, contextPct calculation, empty input handling, and latest-model-wins behavior. Note: Tests fail due to pre-existing vscode module mock resolution issue in vitest that affects all test files equally. The test suite infrastructure needs fixing (vscode alias configuration) but feature implementation is complete and correct. Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
…overwrote helpers
…eline + meta-bar + filter bar Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
Without a CSP meta tag, VS Code's webview applied a default restrictive policy that blocked the inline <script> from executing. The script never ran, so the 'refresh' postMessage was never sent, and the sidebar stayed on 'Loading agents…' indefinitely. Added a nonce-based CSP (recommended by VS Code docs) and threaded the nonce through to the script tag to allow execution. Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
…r hang
esbuild compiled the emoji characters (💬📄📁) in renderActionBtns() to
\u{1F4AC}-style Unicode code point escapes. These are valid ES2015 but
cause a SyntaxError when VS Code loads the HTML via document.write(),
preventing the inline script from ever executing and leaving the sidebar
stuck on "Loading agents…".
Replacing with 💬 etc. keeps them as plain ASCII in the JS string,
bypassing the code point escape issue entirely. Also simplified CSP back
to 'unsafe-inline' since the nonce approach didn't help and the transcript
panel works correctly without it.
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
/\/ in a TypeScript template literal compiles to /\/ in the output
string. The browser then parses /\/g,'/' as a regex that consumes the
surrounding quote, producing a SyntaxError ("missing ) after argument
list") that prevents the entire <script> block from executing.
/\\/ in the template literal compiles to /\/ — the correct regex
matching a backslash — so projectKey/projectLabel work correctly on
Windows paths and the sidebar renders.
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
- Strip XML/HTML tags from agent titles before display so prompts like '<command-name>/clear</command-name>' show as clean text - Fix 'Scanning…' text node persisting alongside rendered agent cards by tracking empty state and clearing root.innerHTML on transition - Hide done/ended sessions behind a collapsible 'N ended' toggle per project group; active (running/idle) sessions always visible Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
applyFilter ran after reconcile and unconditionally set style.display='' on all matched cards, undoing the display:none set on done/ended cards. Now applyFilter respects the endedOpen flag per group before restoring card visibility. Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
body.appendChild on already-attached cards moved them to the end, placing them after the toggle on every render cycle. Fix: update all cards first without touching DOM order, then set correct order in a separate pass (active cards → toggle → done cards). Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
Inactive projects (all sessions ended) are now hidden behind a single collapsible 'Archived (N)' section at the bottom of the sidebar, keeping the active project groups uncluttered. Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
Archived section was appended to root before active groups, so active groups ended up below it. Move root.appendChild(archivedEl) to after the group loop so it always sits at the bottom. Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
2 segments was often ambiguous (e.g. 'CBS / git-ent' vs 'Clients / CBS / git-ent'). Using 3 segments adds enough parent context to distinguish groups without overflowing the sidebar. Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
path.normalize() on Windows converts slashes to backslashes, causing
the webview's split('/') to fail on Windows paths. Normalize to forward
slashes in normalizeCwd() instead of fighting the encoding in the webview.
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
…cards - normalizeCwd: restore \a from BEL (u0007) before stripping other control chars, so Windows paths like d:\Development\ai\agent-viewer display correctly - resolveCwd fallback now calls normalizeCwd so the same sanitization applies - Ended session cards no longer render the meta bar or activity timeline — the section header already communicates the session is done Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
Previously the refresh command only re-posted cached agent data. Now it sends a reset to the webview (clears DOM + reconcile maps), then triggers a full agentService.refresh() re-scan. Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
…ents Chokidar can miss 'add' events for files in directories created after the watcher starts, particularly on Windows. Every 30s, scan for any .jsonl files not yet in the agents map and process them. This ensures subagents spawned during a session are eventually picked up without a manual refresh. Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
…tics - Use forward slashes in chokidar watch glob (PROJECTS_ROOT on Windows has backslashes which break glob matching for recursive subagent directories) - Log subagent count at init and each emit so failures are visible in output Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
Previously refresh() only re-processed files without clearing the agents map or title cache, so the webview received a reset followed immediately by the same stale data. Now it wipes both maps, resets _ready, and re-runs initialize() so the view fully rebuilds from the filesystem. Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
…e get titleCache.get(sessionId) was called after await tailEvents(), so a concurrent refresh() could clear the cache in between, returning undefined and crashing buildAgent(). Capture the TitleCache in a local variable before the tailEvents await so the local reference survives any clear(). Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
The result section was always fully expanded inside an open tool entry, making the layout noisy. Now the result body starts hidden - only the label row (with icon, name, and preview) is visible. Clicking the label toggles the body open with a rotating caret indicator. Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
- entry-body-content: less vertical padding (6/8px -> 5px each), tighter line-height - result-section: smaller margins/padding throughout - result-body: smaller font, tighter line-height (1.5 -> 1.35), white-space:pre so line-numbered Read output renders as compact monospace without word-wrap Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
The result-section was nested inside entry-body-content (the 180px scrollable div), making expanded results appear inside the input scroll container and adding extra vertical space. - Move resultSection to be a sibling of entry-body-content, not a child - Give result-section its own horizontal padding (10px) to align with content - Switch result-body.raw from white-space:pre to pre-wrap so plain-text results (e.g. Edit success messages) wrap instead of overflowing - Add word-break:break-all to result-body.json for long JSON lines Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
Findings from the three review agents, fixed:
Memory / efficiency:
- evictedSessions now insertion-order bounded at 10k entries
- Added evictSession() helper that clears per-session state from all
six tracking maps (was: 3 manual lines per call site, missing entries)
- checkStaleOperations early-exits when no in-flight ops
- Removed unconditional per-read I/O log (was log spam — kept only the
>1s slow log which is the actually useful signal)
Dead code / redundant state:
- Dropped processFileActive counter — derived from processFileInFlight.size
- Dropped inflightAtStart variable — was always true at assignment point;
the finally{} guard via .has() was already correct
- Renamed processFileLastReadMs → processFileNextReadAt; stores the
absolute time read is next allowed instead of a shifted past-timestamp.
Clearer semantics, simpler cooldown check.
Duplication:
- Extracted clearTransientQueues() helper shared by refresh(),
flushQueue(), and dispose() (was three near-identical drain loops
that had already drifted — refresh wasn't clearing waiters)
Comments:
- Removed leaky "Defender exclusion TIP" log from initialize
- Trimmed war-story comments referencing specific past freeze durations,
file sizes, and incident counts (those belong in commit history)
- Consolidated field-group comments for the processFile-related fields
All 33 tests passing; lint clean.
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
…detection Two related fixes for slowdowns observed during unfocused periods: 1. processFile now self-enforces shouldBatch() — every caller honors batch mode, not just chokidar handlers. Previously scanPendingSubagentDirs, watchAndDiscoverSubagents, and discoverNewFiles all called processFile directly, bypassing batching and firing immediate reads while the window was unfocused. That defeated the purpose of batching for subagent discovery (which is when the heaviest load comes in). drainPendingPaths passes bypassBatch:true so its calls actually run instead of looping back into _pendingPaths. 2. runIoControlProbe now triggers throttle detection directly when its timing exceeds 500ms. Previously throttling required 3 consecutive tick lags above 2s, which can miss real saturation that shows variable lag. The control probe is the cleanest signal we have for libuv pool health — if our own 4KB source-file read is slow, the pool is starved. All 33 tests passing; lint clean. Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
Two fixes for the freeze that user just hit:
1. Focus debounce was comparing the incoming value against the COMMITTED
_windowFocused state, not the pending target. A rapid blur→focus
sequence (screenshot tool overlay, brief click-aways) would:
- lose focus → schedules timer to set false
- regain focus → matches committed (still true), early-exits, timer NOT cancelled
- timer fires → state incorrectly flips to false
Now tracks _pendingFocused and compares against that during the
debounce window, matching expected debounce semantics.
2. flushQueue was useless when the slow operations were live on libuv
threads (user just reported "flush did not help" with all counts=0).
Now also clears processFileInFlight, processFileStartedAt, and
processFileNextReadAt — frees our concurrency gate so new work can
run even while orphaned operations continue on their threads. The
"next read" cooldowns also reset so chokidar events can re-read.
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
Wraps every fs operation we issue (stat, open, read, close, readdir, readFile, writeFile) in a counter so we can compare our actual contribution against the global process._getActiveRequests count. The ioProbe log now shows: requests=X ours=Y (stat=N,open=N,...) external=Z where Z = X - Y, telling us how many active fs ops aren't ours (VS Code's own file watching, other extensions, etc.). Goal: if we see 9502 requests but ours=8, we know definitively the problem isn't in our code and the solution must come from elsewhere (reducing what other code is doing, raising UV_THREADPOOL_SIZE, etc.). If we see ours=2000, we found a real leak to fix. Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
Regression introduced by the central batch gate. When VS Code's window started unfocused (very common in development workflows), every processFile call during initialize() short-circuited into _pendingPaths instead of actually loading the agent. The user reported "Initial scan complete — 10 sessions loaded" when their disk had 3000+ sessions, then agents trickled in over minutes as the batch drainer caught up. Side effect: as drained agents added subagent dir watchers, chokidar fired 'add' for their subagent files, which also batched, drained, etc. — turning a one-shot startup scan into a 5+ minute cascade. Fix: the batch gate now skips bypass-checking during this._initializing. All processFile calls (direct from initialize, indirect via watchAndDiscoverSubagents during init) run immediately. After initialize() flips the flag, normal batching resumes. Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
The 8s minimum cooldown was being set after every read, including fast ones. For actively-written files that complete reads in 2ms, this locked out reads for 8s, so the agent's mtime stayed stale and the transcript panel saw no change to render. User reported having to click the agent to force a refresh. Original cooldown purpose was preventing compounding slow reads under Defender. Fast reads don't need it — the in-flight guard plus chokidar debounce already prevent thrashing. Also clear any stale cooldown entry on a fast read, so a file that recovers from slowness immediately becomes responsive again. Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
…dates A 1.6s read on a 12MB transcript was triggering a 3.3s cooldown, blocking the next read until 4.9s later. That's not real-time. The transcript panel saw no mtime change in between and skipped updates entirely. The cooldown's actual purpose is protecting against catastrophic >10s reads (Defender saturation, file locks). Normal reads of 1-5s on large files don't need throttling — the in-flight guard plus chokidar's debounce already prevent thrashing. Now: <5s reads have NO cooldown. >5s reads get 2× cooldown as safety net. Always clear stale cooldown on a fast/normal read so a file that recovers from slowness immediately becomes responsive again. Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
User reported the agent list cleared and the transcript stopped updating after a freeze recovery — no postAgents/emit lines in the log for 6+ minutes. Root cause: scheduleEmit only fires on chokidar events or state transitions. After a freeze (or chokidar buffer overflow on resume), chokidar can go silent. Sessions sit at 'running' state with no transitions, so no emit, so no postAgents, so no UI refresh, so the webview's zombie-detection path never runs. Two fixes: 1. Force-fire onDidChange every 6 ticks (~30s) in onTick(). Gives the webview a heartbeat and updateTranscriptPanels a chance to pick up transcript changes chokidar may have missed. 2. Force-fire onDidChange when throttle clears. Recovery from saturation is a known point where chokidar may have missed events during the slow period. Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
…chers
Architecture change recommended by 4 independent Opus agent diagnoses.
Replaces the chokidar glob watcher (~1290 directory handles for the
entire projects tree) with:
1. ONE root watcher (depth: 0) on PROJECTS_ROOT, just for 'addDir'
events when new project directories appear.
2. Per-active-project-directory watchers, one per non-done session's
project dir. These watch recursively, so subagent files in nested
subdirs are picked up automatically — eliminating the need for the
dynamic watcher.add('subagentDir/*.jsonl') calls that were generating
chokidar internal fs traffic we couldn't see in our trackedFs counter.
3. A reconciler that runs every reconcileIntervalSeconds (default 5s):
- Adds watchers for newly-active project dirs
- Removes watchers for dirs whose sessions have all gone done
- Every activeStatScanIntervalSeconds (default 15s), stats every
in-memory agent file to detect resumed-done sessions or any file
change chokidar may have missed during a buffer overflow.
Both intervals are configurable via VS Code settings:
- agentViewer.reconcileIntervalSeconds
- agentViewer.activeStatScanIntervalSeconds
Net effect: handle count drops from ~2775 to ~1+N where N is the
number of currently-active sessions (typically 1-5).
Removed the watcher.add('subagentDir/*.jsonl') calls in
scanPendingSubagentDirs, watchAndDiscoverSubagents, and the matching
watcher.unwatch in pruneSubagentState — all were adding globs to a
root watcher that should stay at depth:0.
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
Findings from the three review agents, fixed: DRY: - Extracted onJsonlAdd/onJsonlChange/onJsonlUnlink as private methods. addProjectWatcher now binds them instead of inlining 30 lines of handler logic. Single source of truth if any of them need tweaking. - Promoted the inlined 500ms literals to CHANGE_TRAIL_MS and UNLINK_GRACE_MS constants. Efficiency: - scanKnownAgentsForChanges now filters to !parentSessionId. Subagents are covered by their parent's recursive watcher; stat'ing them was redundant work every 15s. Cuts the resumed-detection scan from ~80 stats to ~20 (in user's typical case) without losing coverage. - scanNewProjectDir now adds the project watcher BEFORE processing files. Previously the watcher wasn't added until the next reconcile tick (up to 5s), so file changes in that window were silently missed. - Config change to agentViewer.reconcileIntervalSeconds now takes effect immediately via onDidChangeConfiguration rather than requiring a reload. Cleanup: - Deleted three stale comments narrating the diff itself (the "No watcher.add here — per-project recursive watcher covers it" variants). The architecture has changed; the absence is the norm. - Trimmed the orphaned discovery-tick comment that referenced the removed watchAndDiscoverSubagents claim. All 33 tests passing; lint clean. Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
CRITICAL — Windows path-separator mismatch was breaking the watcher Map. path.dirname(agent.transcriptPath) produced backslashes on Windows when transcriptPath came from path.join (initial scan), but chokidar emits forward-slash paths. Once a file event came through chokidar, its transcriptPath became forward-slash. On the next reconciler tick, the Map.has() lookup would fail because the key was backslash, causing close-and-reopen of every watcher every tick. Fix: added projectDirOf(transcriptPath) helper that normalizes to forward slashes and walks up to the top-level project directory. addProjectWatcher also normalizes its input. Map keys are always forward-slash regardless of where the path originally came from. HIGH — Subagent orphaned when parent goes done. The reconciler skipped subagents (logic: "parent's recursive watcher covers them"), so when a parent transitioned to done its dir watcher closed, even if a subagent under it was still running. The 15s stat scan also skipped subagents, so no safety net either. Fix: runReconciler now includes ALL non-done agents (parent and subagent) keyed by their top-level project dir. Since subagents map to the same project dir as their parent, the Set dedupes — but a still- running subagent keeps its project dir's watcher alive even when the parent has gone done. HIGH — scanKnownAgentsForChanges could overlap with itself if a slow scan took longer than the configured interval. Added _scanInFlight guard with try/finally so a slow scan can't pile up concurrent copies. All 33 tests pass; lint clean. Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
CRITICAL — addDir handler's strict === comparison against PROJECTS_ROOT would never match on Windows because chokidar emits forward slashes while PROJECTS_ROOT from path.join has backslashes. Same separator bug class as the first review caught for projectDirOf, just missed in this spot. Normalized both sides before comparing. HIGH — Eviction → chokidar race: between evictSession running (which removes from agents Map and adds to evictedSessions) and the next reconciler tick (up to 5s), a chokidar 'change' event for the evicted file would re-enter processFile and re-add the session, since processFile does not check evictedSessions (only discoverNewFiles does). Fix: call runReconciler() synchronously right after eviction so the project watcher is closed before chokidar can fire another event. HIGH — Reconciler timer could be created twice if the user changed agentViewer.reconcileIntervalSeconds while initialize() was still running. The config-change listener would create the first timer; then initialize() would create a second at its end. Both run forever. Fix: config listener now only re-arms an existing timer. If no timer yet, initialize() will create one with the current config value. MEDIUM — chokidar.watch can throw synchronously on Windows if the directory disappeared between desired-set computation and watch creation. The reconciler had no try/catch around the call, so the error would propagate out of setInterval. Wrapped addProjectWatcher's chokidar.watch + listener attachment in try/catch. All 33 tests pass; lint clean. Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
…ckoff H1 — runReconciler called from inside scheduleEmit's eviction path would also run the stat scan (since _lastActiveStatScanMs=0 on the first eviction). The async scan eventually calls processFile → scheduleEmit again, which schedules another debounce. Not infinite (re-entry only goes one level), but messier than necessary. Replaced with a targeted closeOrphanedProjectWatchers() that only closes watchers for dirs whose sessions all went done — no stat scan side effect. H2 — refresh() did not await per-project watcher closes before calling initialize(). On Windows, re-opening a chokidar watcher on the same path before the prior close drains can fire duplicated 'add' events or throw EBUSY/EPERM. Now awaits Promise.all over all per-project closes. H3 — addProjectWatcher had no backoff on chokidar.watch failures. A permanently-inaccessible dir would retry every 5s forever, spamming logs. Added _watcherFailBackoff Map with WATCHER_FAIL_BACKOFF_MS=60s. M1 — addDir's normalized comparison against PROJECTS_ROOT didn't handle drive-letter case mismatch (USERPROFILE=c:\… vs chokidar emitting C:\…). Case-insensitive comparison on Windows. M2 — On a fresh install with no Claude projects yet, ~/.claude/projects/ doesn't exist; chokidar.watch on a missing dir silently no-ops on Windows and never recovers. mkdir -p at the top of initialize() so first-run works without a manual Refresh. All 33 tests pass; lint clean. Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
When a session ends, Claude Code writes a 'summary' event that isTerminator() matches. buildAgent set terminated = events.some(isTerminator), and classifyState returns 'done' whenever terminated is true regardless of mtime. When the user resumes a done session, Claude appends new events AFTER the old summary. Our 64KB tail read picks up BOTH the new events and the still-present old summary. events.some(isTerminator) sees the summary and incorrectly marks the session terminated, so it stays locked at 'done' even with fresh mtime. User reproduction: 1/3 stat scan was finding the resumed session's new mtime, processFile was reading the tail, but the state never transitioned out of done because of the stale summary. Fix: check whether the LAST event is a terminator, not whether ANY event is. JSONL is append-only chronological, so a summary followed by more events means the session was resumed and continues. Also expanded scanKnownAgents logging to always fire (not just when stale>0) so we have visibility into whether the scan is running. Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
Eviction at 3×/10× doneAge was added earlier in this session during the freeze investigation, on the theory that fewer agents in memory would reduce I/O pressure. The actual freeze cause was chokidar's invisible internal I/O, not agent count. Now that chokidar is fixed, eviction is just a regression that hides archived sessions from the user. Default is now never-evict — all done sessions stay in memory and remain accessible via the webview's existing "Show N hidden" filter, which already has pagination (DONE_PAGE_SIZE=100). New setting agentViewer.evictionAgeMultiplier (default 0 = never) preserves the old behavior for users who want it. Setting to e.g. 10 restores the just-prior 10× doneAge eviction. Cost of not evicting at ~3700 agents: - Memory: tens of MB heap, negligible - Per-tick classifyState: linear, sub-ms - Per-emit serializeMs: previously observed 0ms even at 3696 agents Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
Subagents already have a 'model' field in their serialized payload, but it wasn't surfaced in the UI. Now each sub-row shows a small model chip between the task name and the timestamp, using a slightly tighter variant of the existing .model-chip style for the more compact context. Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
A parent with many subagents (e.g. 49) was showing all of them in the sub-list, even though most were done. Active subagents (running/idle) now render inline; done subagents collapse into a nested toggle that defaults to closed. Click 'N done' to expand the done section. Open/closed state is persisted via openSections keyed by 'proj:<project>:done-subs:<sessionId>', mirroring the existing pattern for the parent sub-toggle. Scoped the existing .card.subs-open .sub-caret rule to direct-child .sub-toggle only, so the nested .done-sub-toggle's caret rotates independently based on its own state. Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
User reported new sessions taking ~2 minutes to show up. Root cause: the new reconciler architecture only watches dirs of currently-active sessions. A new session in a previously-quiet project dir has no watcher, so the only detection path is the discoverNewFiles tick which was running every 5 minutes (average wait: 2.5 min). Reduced default discovery interval from 5 min → 30s, made it configurable via agentViewer.discoveryIntervalSeconds, and wired up the config listener so changes take effect without reload. Cost at 30s default: ~1290 readdirs/30s = ~43 readdirs/sec average, done in batched groups of 50 = ~250ms wall time per cycle = ~0.8% CPU. Trivial compared to the responsiveness win. Hidden sidebar scales up to 6× the visible interval (3 min at default) so we don't waste cycles scanning a tree no one's looking at. Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
User reported that clicking 'Load 100 more ended sessions' showed no obvious effect — the remaining count dropped, but the sessions appeared in collapsed project groups inside the Archived section, multiple levels of nesting deep. Two changes: 1. Renamed button text: - was: 'Load 100 more ended sessions (1921 remaining)' - now: 'Show 100 more in Archived (1921 still hidden)' Explicit about where the sessions appear. 2. Clicking the button now auto-expands the Archived section so the new project groups are immediately visible (each can still be expanded individually to see its sessions). Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
When the user clicks 'Show 100 more in Archived', the new sessions distribute across many similar-looking project rows. Even when a row's count goes up by 99, it's nearly invisible without doing a careful diff. Track per-group session count in lastGroupCounts. On each render, if a group's count grew, briefly flash its badge with a blue highlight that fades back to its normal state-color over 1.6s. The animation reset trick (remove → force reflow → re-add) ensures consecutive load-more clicks each trigger a fresh flash. Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
Drops healthy-state noise (~400 lines / 22min → ~5-10) from drain, scanKnownAgents, ioProbe, discoverNewFiles, emit, and postAgents. Heartbeat (every 5min) still proves liveness; slow paths, structural changes, and delivery failures still log. Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
Replaces the antenna-on-stick with an 8-spoke asterisk burst above the robot head to evoke the Claude/Anthropic mark. Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
Transcript panel: - Search: next/prev buttons (↑↓), match counter, Enter/Shift+Enter navigation, re-apply highlights after live DOM updates, multi-match per text node, more visible highlight style with active-match accent - Diag bar moved to sticky footer (always visible), session ID added with user-select:all for easy copy, token count in footer - Single atomic prepend for history chunks eliminates content dancing during lazy-load - Activity timeline hidden for idle/done agents (shown only when running) Sidebar: - Model badge: inline-block + symmetric padding, no emoji icon - Hide "Thinking"/tool trail for idle/done agents Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
C--Users-skean-AppData-Local-Temp was left undecoded when events lacked a cwd field, producing a different projectKey than sessions with a real cwd event. Both now resolve to C:/Users/skean/AppData/Local/Temp and land in the same group.
…menu queue-operation: use enqueue content as latestUserPrompt/firstUserPrompt fallback when no clean user-event prompt exists — fixes JSON blob titles on subagent sessions that receive tool data as their first user turn. Transcript: right-click context menu with 'Format as code' — copies selected text (or full body if nothing selected) wrapped in a code fence. Also disable markdown-it typographer so quotes copy verbatim from source.
…menu queue-operation: use enqueue content as latestUserPrompt/firstUserPrompt fallback when no clean user-event prompt exists. Transcript: right-click context menu with 'Format as code' — copies selected text (or full body if nothing selected) wrapped in a code fence. Also disable markdown-it typographer so quotes copy verbatim from source. Fix: escape \n as \n inside template literal to avoid literal newline in embedded JS string (was causing SyntaxError on document.write).
queue-operation: use enqueue content as latestUserPrompt/firstUserPrompt fallback when no clean user-event prompt exists. Transcript context menu (right-click in scroll area): - Copy: copies selected text via extension host clipboard - Format as code: re-renders selected text as an inline code block Both items grey out when nothing is selected. Fixes: escape \n as \n in template literal, route clipboard through vscode.postMessage (navigator.clipboard loses focus on right-click), save Range at contextmenu time so click doesn't clear selection. Also disable markdown-it typographer so quotes render verbatim.
- Format as code: detect JSON and pretty-print with syntax highlighting - Format as code: falls back to entire response body when nothing selected - Fix typographer:false (was lost in revert) so quotes copy verbatim
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.
Summary
This PR brings a substantial round of performance improvements, reliability fixes, and UX polish developed against the extension on a machine with 4000+ Claude Code sessions tracked.
Startup performance
initialize()now skips the 32 KB title read for sessions older than the done-age threshold. A background pass (backgroundScanTitles) fills them in progressively after the sidebar first renders, usingsetImmediateyields so live events are never starved. Cold-start time on a 4000-session machine dropped from ~22 seconds to ~2 seconds.Subagent discovery reliability
_agentsWithSubagentsset — once a parent session is seen to have Agent tool_use calls, it's remembered permanently. Previously,watchAndDiscoverSubagentswas gated onagentCallDescs, which ages out of the 64 KB tail window on long sessions — causing subagents to silently stop being discovered mid-session.watcher.add— chokidar does not reliably watch a glob pointing at a directory that doesn't exist yet.watcher.addis now deferred untilreaddirconfirms thesubagents/directory exists, preventing silent misses when the directory is created after the parent session starts.agentService efficiency
buildTreeskipped on content-only updates — in-place field mutation preserves parent→child refs without a full two-pass rebuild over 3800+ agents on every file write.MIN_CHANGE_EMIT_MS(2 s) prevents rapid writes from multiple active sessions triggering a full rebuild on every line.setTimeoutchain replacessetIntervalso concurrent scans can never pile up.agentViewer.runningWindowMinutesandagentViewer.doneAgeHoursexposed as VS Code settings.transcriptPanel
parseTranscriptWithStateyields viasetImmediateevery 500 lines, keeping the event loop responsive during large transcript loads and preventing "Loading…" hangs.drainPanelUpdateprevents pile-up under rapidonDidChangebursts; mtime guard skips redundantstat+readFilecalls.openJsonlreverted to the simplervscode.opencommand.Code quality
readFileSliceextracted tosrc/fileUtils.ts— removes identical copies fromagentServiceandtranscriptPanel.latestUserPromptmerge fix —backgroundScanTitlesused!x(overwrites empty string) instead of??, inconsistent withbuildAgent.'background'emit reason — background title scans no longer reset_lastChangeEmitMs, which was delaying live-session updates by up to 2 s per batch.align-items: centerfor reliable vertical alignment regardless of font metrics.Test plan
agentViewer.doneAgeHoursandagentViewer.runningWindowMinutestake effect🤖 Generated with Claude Code