- Fix company isolation: reset chat state when switching companies,
clear stale comment cache, fix Board Operations issue creation
(status: todo instead of in_progress to avoid assignee requirement)
- Optimistic user messages: show user's message instantly before server
round-trip for a natural chat feel
- Live status indicators: forward tool-use events from Claude CLI as
SSE status messages (Running a command, Reading a file, Searching, etc.)
shown as a separate bar below the streaming response
- Paperclip SVG thinking animation (slowed to 1s loop)
- Chat bubble styling: blue user bubbles, shaped corners via raw CSS
to bypass --radius:0 design system
- Agent system prompt template: add Model field (default: sonnet)
- Add beginner guide: doc/BOARD-CHAT-GUIDE.md
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
Add a board-member skill that teaches Claude how to manage a Paperclip
company via chat — covering onboarding, hiring plans, approvals, task
monitoring, cost oversight, and agent system prompt management.
Phase 1 (Claude Code surface):
- Board skill at skills/paperclip-board/SKILL.md with full API reference
- CLI bootstrap command `paperclipai board setup` that installs the skill
and prints env exports
Phase 2 (Web UI surface):
- New /board/chat/stream endpoint that spawns Claude with the board skill
as system prompt, passing PAPERCLIP_API_URL and PAPERCLIP_COMPANY_ID
- BoardChat page with streaming responses, status indicators, and
conversation persistence via Board Operations issue
- Sidebar nav link and route registration
The skill is a portable knowledge layer — same document powers Claude Code
(Surface 1), web UI chat (Surface 2), and future MCP server (Surface 3).
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
Layer 1 (instant): detectUserIntent() on user message creates work product
+ fires background artifact generation immediately.
Layer 2 (CEO confirm): parseStructuredActions() on %%ACTIONS%% signal in
CEO response, with regex fallback. Deduplicates against Layer 1.
Also adds: canned openers for common messages (hybrid with real streaming),
message action metadata bars, cleanAgentMessage strips action signals,
generate-artifact now posts CEO comment and updates task to in_review.
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
Replace adapter-based chat relay with lightweight claude CLI streaming
endpoint. CEO responses now arrive in 2-5s instead of minutes.
Key changes:
- POST /agents/:id/chat/stream — spawns claude -p directly, SSE streaming
- POST /agents/:id/chat/canned — persist welcome/approval messages
- POST /agents/:id/chat/generate-artifact — background doc generation
- Server-side detectArtifactCommitments() replaces unreliable AI observer
- Frontend: optimistic user messages, typewriter streaming, observer events
- Onboarding: separated name/mission substeps, back navigation
- Dev script: full server restart + fresh company + mission setup
- Removed: canned responses, heartbeat polling, inline plan detection
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
New POST /api/agents/:id/chat/relay endpoint that calls the adapter
directly and streams stdout back via SSE, bypassing the heartbeat
queue. Comments are persisted normally so conversations stay durable.
Frontend tries the relay first, falls back to poll-based flow if
unavailable.
Backend: 1 new file (agent-chat.ts), 1 line in app.ts.
Frontend: streaming fetch in CEOChatPanel with fallback.
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
