v0.1.1

Patch release — first-turn fix, docs refresh, CI hardening.

Fixes

• First-turn message loss — SDK can deliver content_block_start before message_start on the first turn. Server now auto-initializes message context, so the first response and tool calls render correctly.
• Multi-turn message loss — forceFlushPendingMessage() runs at turn boundaries, preventing silent drops when message_start resets state.
• SDK field names — file_path not path, content not contents, pattern not glob_pattern. Tool pills now show actual file paths and inputs.
• Session end ordering — force-flush before session_end. Frontend SESSION_END and MESSAGE_START force-finalize orphaned messages.

Docs

• README rewritten — same information, half the lines, no AI slop
• CLAUDE.md updated with v2 protocol details and SDK field name gotchas
• Design docs marked Implemented with post-implementation notes
• Enterprise plan v1.3 with current LOC baselines (chat.ts 468, ChatView 337)

CI

• Test count guard — fails if tests drop below 200 (currently 209)
• Bundle size reporting in job summary

v0.1.0 — Stable chat experience

First tagged release. The v2 message protocol is complete and the chat experience is stable.

What's in this release

• v2 message protocol — explicit block lifecycle (start → delta → end) for text, thinking, and tool_use blocks. No more inferred state from timing or absence of events.
• Thinking blocks — render with expand/collapse, survive across turns
• Tool pills — show names immediately on block_start, file paths/inputs on block_end, results after execution
• Multi-turn streaming input — follow-up messages queue natively, interrupt mid-generation
• iOS reliability — message snapshots on reattach after silent WS drops, connection pooling with heartbeat and reconnect
• Session management — restore from localStorage and API, swipeable session list, session dismiss
• Mobile-first UI — compact layout, mode pills (Ask/Agent/Auto), model selector, sandbox toggle

Bug fixes in this cycle

PR	Fix
#35	Session restore crash — v1→v2 shape mismatch in getMessages and localStorage
#36	Tool results not paired with tool calls in restored sessions
#37	Deferred message_end until all content blocks are closed
#39	SDK field names in tool summaries (file_path, content, pattern) + session_end ordering
#40	Multi-turn message loss — flush pending message_end on new turn, finalize orphaned current

What's next

• Plan Mode (Ask mode produces structured plan, Execute button starts Agent run)
• Tool permissions v2 (deterministic tier model with mode matrix)