Replace the hardcoded MODEL_PRICING table and costUsd() function in
sessionAnalyzer.ts with calculateMessageCost() and getDisplayPricing()
from the shared pricing module. Re-export getDisplayPricing as getPricing
for backward compat with CostSection. Replace the ModelPricing interface
with a type re-export of DisplayPricing.
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
Create a unified pricing module in src/shared/utils/ that wraps
resources/pricing.json with typed lookup, tiered cost calculation,
and display pricing helpers. This is the foundation for consolidating
the dual cost calculation systems (main jsonl.ts + renderer sessionAnalyzer).
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
Shared pricing module replacing dual cost calculation systems
(jsonl.ts hardcoded + sessionAnalyzer.ts LiteLLM) with a single
src/shared/utils/pricing.ts used by both main and renderer processes.
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
- Introduced MoreMenu component to consolidate less-frequent actions (Search, Export, Analyze, Settings) behind a dropdown menu.
- Removed individual action buttons from TabBar for a cleaner interface.
- Implemented functionality to close the menu on outside clicks and Escape key press.
- Extract SCROLL_THRESHOLD (300px) constant so auto-scroll hook and
scroll-button visibility logic stay in sync
- Extract CONTEXT_PANEL_WIDTH_PX (320px) constant to avoid layout drift
if the context panel width is ever adjusted
- Gate drag spacer on isElectronMode() && isLeftmostPane to match the
TabBar drag region logic and prevent unintended drag regions in
split-pane layouts
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
- Add toggle in settings to auto-expand AI response groups
- Auto-expand new AI groups on live session refresh
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
Includes scroll improvements:
- Scroll to bottom on session open and live auto-scroll
- Make auto-scroll StrictMode-safe via needsInitialScrollRef
- Add floating scroll-to-bottom button in chat view
Window drag fix:
- Apply drag region to leftmost pane TabBar regardless of sidebar state
- Cap tab list at 75% so drag spacer always has room
- Add explicit flex-1 drag spacer between tabs and action buttons
- Set WebkitAppRegion: no-drag on tab items for reordering
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
Show "no commits" / "no lines changed" instead of "total cost ÷ 0"
when commitCount or linesChanged is zero.
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
- Recompute totalInputTokens, totalOutputTokens, totalCacheCreation,
totalCacheRead, and grandTotal after injecting the "Subagents (combined)"
row so tokenUsage.totals stays consistent with byModel rows
- Fix CostSection import order: external packages before path aliases,
merge duplicate import type statements
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
- Compute subagent cost from token breakdown instead of relying on
unpopulated proc.metrics.costUsd; extract actual model from subagent
messages for accurate pricing and mismatch detection
- Add aggregated "Subagents (combined)" row to cost table with arrow
navigation to the Subagents report section (no misleading breakdown)
- Filter <synthetic> model from token/cost tracking to eliminate zero rows
- Fix parseTestSummary to treat missing pass/fail count as 0 so clean
all-passing test runs are not dropped from trajectory analysis
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
Intercept Ctrl+R/Cmd+R in before-input-event to prevent Electron's default page reload. Replace hardcoded macOS shortcut symbols with platform-aware helpers that show Ctrl+ on Windows/Linux.
Skip indexing and lookup when tool call/result IDs are missing to
prevent collisions on the empty-string key in toolCallIndex.
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
- getPricing: use tokenized order-insensitive matching so Claude 3 era
models (e.g. "claude-3-opus-20240229") correctly match pricing key "opus-3"
- sessionAnalyzer: skip isSidechain messages in parentCost loop to prevent
double-counting with processSubagentCost
- CostSection: disable row expansion when token stats are missing
- tabSlice: guard against undefined sessionId/projectId in openSessionReport
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
- Add AssessmentBadge component with hover tooltips explaining thresholds
- Add Key Takeaways summary section surfacing top actionable findings
- Add cost attribution stacked bar and per-token calculation breakdowns
- Add clickable navigation from takeaways and tool errors to detail sections
- Add theme-aware assessment colors via CSS variables for light/dark mode
- Collapse lower-priority sections by default for progressive disclosure
- Replace all hardcoded color hex values with CSS variable references
- Fix missing Fragment key in CostSection model table
- Add defensive division-by-zero guard in stacked bar calculation
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
Add interpretive assessment layer to session reports. Every metric section
now shows color-coded severity badges (green/amber/red) computed from
configurable thresholds, replacing raw numbers with instant signal.
Includes centralized reportAssessments utility, model mismatch detection,
switch pattern recognition, and 44 new tests.
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
Add 7 missing sections and fix 4 incomplete ones that the Python
analyzer generates but the TypeScript port was missing:
New sections: skillsInvoked, bashCommands, lifecycleTasks,
userQuestions, outOfScopeFindings, agentTree, subagentsList
Fixed: compaction (was just count, now has summaryCount + note),
serviceTiers removed (not available in parsed types),
compactionCount replaced with compaction object
Adds InsightsSection UI component and 8 new tests.
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
Port all analysis logic from the Python script to TypeScript, running
entirely in the renderer process. Produces a typed SessionReport from
a SessionDetail with: token/cost analysis, friction signals, idle gaps,
conversation tree metrics, test progression, and thrashing detection.
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
The HTTP_SERVER_GET_STATUS handler returned a raw { running, port } object
while all other HTTP server handlers (start, stop) return { success, data }.
invokeIpcWithResult() in the preload expects the IpcResult<T> envelope and
treats a missing success field as failure, throwing 'Unknown error'.
This caused an unhandled promise rejection every time the Settings page
opened, because GeneralSection calls api.httpServer.getStatus() on mount.
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
When the sidebar is collapsed and the user clicks a project, the session
list was silently updated behind the hidden panel with no visual feedback,
leaving the user confused about why nothing appeared to change.
Fixed by setting sidebarCollapsed: false in both selectProject and
selectRepository at the moment of selection. Both entry points are covered:
- selectRepository: used by the dashboard project cards
- selectProject: used by the command palette and keyboard shortcuts
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
Two performance regressions introduced in recent PRs:
1. readAgentConfigs blocked transcript rendering (PR #50)
The agent config IPC call was awaited on the critical path of
fetchSessionDetail, preventing any transcript data from rendering
until the filesystem read completed. On macOS this was especially
noticeable due to security checks on first directory access.
Fixed by making the call fire-and-forget: the transcript renders
immediately and subagent color badges update asynchronously.
Also set the project cache key optimistically before the async call
to prevent duplicate in-flight requests on rapid navigation.
2. SessionSearcher stat()-called every session file on each search (PR #53)
LocalFileSystemProvider.readdir() did not populate the optional
mtimeMs field on FsDirent entries. The new SearchTextCache-based
SessionSearcher fell back to an individual fsProvider.stat() call
per session file when mtimeMs was missing, adding N extra filesystem
round-trips on every search in local mode.
Fixed by statting all entries concurrently inside readdir(), so
mtimeMs is always populated and the stat fallback is never triggered.
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
