v2.28.1 — fix: pair orphan tool_calls on history load

Fix

Tool cards from chat history were rendering as forever-running amber spinners
because the tool_result data was never linked back to its originating tool_call.

Root cause

OpenClaw gateway persists tool results in a shape Silos wasn't reading:

• Anthropic-style: `{type: 'tool_result'}` content blocks inside the next user message
• OpenClaw-native: separate top-level messages with `role: 'toolResult'`, with `toolCallId` + `toolName` + `content` as siblings on the message

`loadChatHistory` was dropping `role: 'toolResult'` messages via filter, and the link was lost.

Fix

• `buildToolResultMap` scans raw history for both shapes
• `injectToolResultBlocks` inserts synthetic tool_result blocks adjacent to each tool_call so `MessageContent`'s existing pairing catches them inline
• `extractToolMessagesFromBlocks` accepts the map so extracted tool ChatMessages (visible in bottom Tools panel) also get their `result` populated
• True orphans (run aborted mid-tool, no result anywhere) now render with an "aborted" notice instead of spinning forever

Tests

17 tests in `message-normalizer.test.ts` cover both shapes + orphan path.

v2.28.0 — perf pass: browser freeze fixes + latency tab hardening

Browser performance (fixes freezes under streaming load)

• Narrow zustand selectors in ChatView, AppSidebar, SessionDetailView, SessionTasksKanban, UsageBar, MainShell, App — unrelated store mutations no longer re-render these components
• Batched delta writes via new recordDelta() action (4 set() per delta → 1)
• Coarse hasAnyActivity boolean subscription instead of raw lastActivity timestamp — stops per-delta ChatView re-renders
• Stable messageGroups useMemo deps + module-level EMPTY_AGENTS fallback — MessageGroup memo now actually skips re-renders during streaming
• React.memo on MessageGroup, MessageContent, ToolCard, ThinkingBlock
• Virtualization threshold lowered 500 → 80 (typical mid-sized conversations now benefit)

Latency tab correctness

• Skip effectiveTokensPerSecond when tool time dominates (>90% of gen time) — prevents absurd rates from tiny residual divisor
• finalizeRunLatency fallback to caller-provided sessionKey when markRunStart was missed
• TTL reaper for orphan runTimings (10min)
• Strict toolCallId dedup on both call and result counters — tolerates duplicate phase events
• Streaming gap threshold 1s → 3s (accommodates batched-delta providers like MiniMax)
• Gap attributed to toolTimeMs only while a tool is actually in flight
• localStorage-backed per-session latency history (7-day cap, max 200/session)
• UI: cap notice when showing last 200, tokenization estimate tooltip, corrected gap-attribution hint

v2.23.0

Highlights

Session telemetry + control pass focused on the CMO-style workflows that live on tools.

Agent control

• Activity-based working signal — dashboard no longer loses track of the agent when OpenClaw's compaction retry transparently swaps runs after a context overflow. 15s recent-activity window keeps the Stop button visible across short gaps.
• Abort fallback via lastKnownRunId — Stop still aborts when activeRunId has been cleared by an intermediate error event.

Latency breakdown

• Per-run tool count badge, gap-based tool time detection (>1s gaps = tool/wait, not generation), and effective tok/s that excludes those gaps.
• Aggregate stats split avg latency by w/tools vs plain runs.

Errors panel

• New reusable JsonTree component with syntax colors, depth-based collapse, long-string previews, and auto-unwrap of nested JSON strings (tool-result content[0].text renders as a sub-tree, not an escaped blob).
• Copy button produces the flattened form you see on screen, not the wire-format with \n\" escapes.
• Detects OpenClaw's {content, details} tool-error shape and shows details by default with a toggle for the raw wire format.

Logs

• Fixes tslog's log.child({subsystem}).info(msg) serialization: parser now correctly extracts the message from 1 instead of rendering the tag stringified in 0 as the body. Unwraps _meta.name when it arrives as a JSON string. Fallback summary when no primary content field is populated.
• LogDetailPanel hides the Payload section when it's just the Message repeated, and renders string payloads as preformatted blocks so CLI-table alignment survives.

Chat history stripping

• Replaced the regex-based stripInboundMeta with a line-by-line parser ported from OpenClaw's canonical stripInboundMetadata. Handles all 6 sentinels + trailing untrusted-context suffix + timestamp prefix. Silos-specific addition strips async System (untrusted): [ts] ... lines for cleaner chat.

Task pipeline

• Dropped the isBackgroundProcess heuristic that substring-matched pid , still running, background in tool results and created false-positive "Subagent" entries in the kanban.
• Sidebar Tasks badge filters stale runs (status=running + startedAt >1h) so it aligns with what /tasks shows.

Included commits:

• 18a2b2b feat(session): activity-based agent-working signal + abort fallback
• dda8309 feat(latency): tool count, gap-based tool time, effective tok/s
• c02cc3b feat(errors): reusable JsonTree + auto-unwrap + tool-result simplification
• e040683 fix(logs): tslog tag unwrapping, preformatted payloads, dedup empty sections
• 3164291 refactor(chat): port stripInboundMeta canonical parser from OpenClaw
• eb170cd fix(tasks): stale-aware sidebar Tasks badge + drop isBackgroundProcess false-positive

v2.22.1

Patch: fix build failure in v2.22.0 from unused LayoutGrid import in TasksFlowsPage.tsx.

See v2.22.0 for feature notes.

v2.22.0

Summary

• Logs: semantic color tokens for light/dark themes — --log-* and --syntax-* CSS variables exposed via Tailwind as log.{error,warn,info,subsystem} and syntax.{string,number,boolean,key}. AA contrast in both modes without dark: modifiers on each class.
• Session telemetry: new bottom tab bar in ChatView with Tools / Errors / Latency, persistent selection and auto-switch on new events. Errors panel captures tool failures, provider errors, rate-limited 429s (RPM/TPM/BUDGET), chat errors, and network errors with per-session badge. Latency panel estimates TTFB, total latency, and tokens/s per run with avg/p50/p95 aggregates (client-side, no gateway changes).
• Crons tab in top panel next to Pipeline: lists cron jobs scoped to the agent; expanding a card lazy-loads runs, clicking a run opens TaskRunDetail with cron-specific info and conversation filtered by runId.
• Tasks redesign (/tasks): cancel/abort/retry actions across subagent, ACP, CLI, and cron runtimes.
• Pipeline kanban: converted to single-expanded accordion — all rows collapsed by default, click to expand one (collapses others), persisted per user.
• Docs: README polish with tighter architecture diagram and Quick Start.

Included commits:

• 120e161 feat(tasks): redesign /tasks + add cancel/abort/retry actions across runtimes
• 73689ad docs: polish README
• 0a32a08 feat(logs): add semantic color tokens for light/dark themes
• 0f4be26 feat(session): add Errors, Latency, and Crons tabs with telemetry
• 6c27d19 refactor(tasks): convert vertical kanban to single-expanded accordion

v2.21.0

Summary

• Always show pipeline kanban rows (Queued/Running/Waiting/Done) even when empty
• Add inline LaTeX→Unicode rendering for chat messages
• Decompose dashboard-store monolith into 9 focused Zustand slices
• Fix tsc -b strict build errors

v2.20.1

Summary

• fix: browser viewer auto-reconnects after VNC disconnect (3s retry loop instead of staying stuck)
• Add postMessage bridge between noVNC iframe and React panels for accurate connection status
• Connection status dot now shown in floating browser panel too

v2.20.0

v2.20.0 — Chat & Settings Refactor

Bug Fixes

• Fix tool call/result matching using toolCallId (aligned with OpenClaw official UI)
• Fix duplicate stripReasoningTags — unified to lib/reasoning-tags.ts
• Fix streaming double-strip bug (preserve mode in StreamingMarkdown)
• Fix tool_use content blocks lost in loadChatHistory
• Fix invisible URLs in light mode (text-white → text-primary)
• Fix rate limit state lost on component unmount (moved to Zustand store)
• Increase chat history limit 100 → 200 (matches OpenClaw)
• Deduplicate tool messages on history reload

Refactoring

• Extract 10 components from ChatView.tsx (1816 → 689 lines) into src/components/chat/
• Extract 7 sections from SettingsPage.tsx (2764 → 108 lines) into src/components/settings/
• Split 730-line handleEvent into 3 focused handlers in chat-event-handlers.ts
• Extract isSilentReply to shared lib/reasoning-tags.ts
• Add toolCallId field to ChatMessage type

Cleanup

• Delete 6 unused components
• Net reduction: -2,116 lines

Commits

ca71925 fix: remove unused imports (EventFrame, describeBrowserAction) for strict tsc -b
46d9e45 refactor: extract chat and settings components, fix 7 chat bugs

v2.19.5

Changes

• Fix context window display: Resolve correct context window per model from the model catalog instead of always showing OpenClaw's 200K fallback
• Strip reasoning tags: Remove internal reasoning tags from message content in task detail, task run detail, and sessions views
• UI cleanup: Add vertical layout to session tasks kanban, remove unused Layout/Sidebar components

v2.19.4

fix: hide misleading 200k context window when provider doesn't report real value