v0.1.19 — UX, protocol & build fixes

User-experience and protocol-correctness release on top of v0.1.18's
security hardening. Two visible new pieces (background task
indicator, CSV / text-blob preview), one structural fix (tool-event
bubble restoration on session reload), and a sweep of LLM-pipeline
and macOS-build issues exposed during data-analysis runs.

Designs in
docs/en/background-task-indicator.md,
docs/en/tool-event-restore.md, and
docs/en/tool-call-roundtrip.md.

Added

• Background task indicator. When the agent kicks off
  post-response work (title generation, memory compaction,
  pinned-fact extraction), a small badge appears in the
  input-status-bar naming what's running. The agent stays Busy on
  the backend until those tasks complete; the input field stays
  disabled and the Sidebar's New / Load / Delete actions are
  greyed so a quickly-typed second message can't race them. Abort
  still cancels post-tasks. Logs are symmetric — every task
  records start and exactly one of done / canceled / error
  at INFO/ERROR level so an operator can correlate "session
  never got a title" with "user typed before title-gen finished".
• CSV / text-blob preview in the Data pane. Clicking a
  text/csv or text/tab-separated-values blob opens a real
  HTML table (first 200 rows, 30 columns shown, RFC-4180-shaped
  quoting); other text MIMEs (JSON, plain text, Markdown, HTML,
  XML, NDJSON, JS) drop to a fixed-width <pre> view. Source is
  clipped to 100 KB before reaching the modal. Binary blobs stay a
  click no-op, with Export still available.
• Tool-event bubbles persist across session reload. Tool
  calls used to be invisible in restored sessions because
  LoadSession dropped every tool record on the floor. They
  now come back as compact name + status bubbles, matching the
  live look. Required adding a Status field to memory.Record
  (omitempty, populated from the executeTool return value
  already used for tool_end activity events). Legacy records
  without the field default to success so older chats stay
  readable.

Changed

• Post-response tasks now hold Busy until they finish. An
  earlier iteration tried auto-cancelling them at the start of
  the next Send, but local LLMs are slow enough that
  pinned-fact extraction never completed during rapid
  conversation, and the extractor only sees the latest 4 hot
  records — cancellation drops those facts permanently rather
  than deferring. Live behaviour now matches the design: state
  stays Busy from the moment a user message arrives until all
  three post-tasks finish, and the frontend (input field, Sidebar
  session ops) keys off that.
• Sticky table headers in the Data pane. Both the new
  BlobPreview CSV table and the existing Tables-tab DB preview
  used --bg-hover (≤0.1 alpha in every theme) for the sticky
  <th> background, so scrolled rows bled through. Added a
  per-theme opaque --bg-sticky-header and routed both rules
  through it.
• Build artifact integrity on macOS. make build used cp -r to publish dist/shell-agent-v2.app. macOS strips
  extended attributes (and with them the ad-hoc code-sign
  resource fork) on cp -r, producing a bundle that crashed at
  launch with SIGKILL "Code Signature Invalid" until manually
  re-signed. Switched to ditto to preserve the signing
  resources end-to-end.
• Tool-call assistant turns are no longer restored as chat
  bubbles. Live chat surfaces tool-call narration through the
  activity stream (the transient progressTool "thinking" banner),
  not as a chat bubble. Restore was inconsistent — any
  non-empty Content came back as a bubble regardless of
  attached ToolCalls, exposing thought-style preambles that the
  live view had hidden. Skip those records on restore so the two
  views match.

Fixed

• Vertex parallel function calls returned HTTP 400. When the
  assistant emitted multiple FunctionCall parts in a single turn
  (parallel tool calls), the matching FunctionResponse parts were
  emitted as separate user Content blocks; Gemini requires them
  packed into one. vertex.go buildContents now coalesces
  consecutive RoleTool messages into a single Content block.
• /model could appear frozen during a 429 retry. The
  slash-command parse used to live behind postTasksWg.Wait(),
  so a /model local typed while a previous turn's
  pinned-fact-extraction was sleeping in retry-backoff blocked
  for minutes. Slash commands now parse before the wait, and
  Abort also fires the post-task cancel func — previously Abort
  only cancelled the in-flight Send, leaving the post-task
  goroutines unkillable.
• Bindings could panic on startup. GetLLMStatus,
  LoadSession, and Send had no nil-guard for the brief
  window between Wails' frontend mount and the backend
  finishing agent.New. The frontend's first poll could race
  in and trigger a nil-pointer panic in every GetLLMStatus
  tick. Added defensive checks.
• Gemini 2.5 Flash thoughts leaked into the assistant text.
  Some responses arrived as THOUGHT\n…\n\nactual reply or
  思考\n…\n\n本文 or シンクタイム: 3秒\n\n本文 — sometimes
  in their own Part with Thought=true, sometimes inline. Set
  ThinkingConfig.IncludeThoughts=false explicitly (default
  behaviour, but worth being explicit), filter Parts with
  Thought=true in parseResponse, and log per-Part shape at
  debug level so the rare inline-text case can be diagnosed
  without guesswork.

Notes

The remaining inline-text-thought case (model writes 思考\n…
into a regular text Part on a fraction of replies) is not
addressed in this release. The Part-shape filter doesn't reach
it, and a heuristic strip risks over-deleting legitimate user
content. The follow-up option is to set
ThinkingConfig.ThinkingBudget = 0, which trades reasoning
quality for an end to inline preambles; deferred to a future
release once the trade-off is evaluated against representative
analysis tasks.