Add agent response rendering specification and design

[johannesjo]

Summary

This change adds comprehensive OpenSpec documentation for a new agent response rendering capability. The specification, design document, and task breakdown define how the app will render structured agent responses (from Claude Code, Amp, and other agents) in a dedicated panel alongside the existing terminal view.

Changes

• spec.md — Complete capability specification with 10 requirements covering:

  • Panel opt-in behavior and lifecycle (enable/disable/restore across restarts)
  • Session ID binding at agent spawn time
  • Forward-only JSONL file tailing with fs.watch + stat-poll fallback
  • Event adapter normalization (Claude Code → shared AgentResponseEvent shape)
  • Live event push to renderer
  • Markdown rendering pipeline reuse with extended sanitization
  • Collapsible tool/thinking blocks
  • Panel coexistence with terminal (no interference with pty)
  • Watcher lifetime management

• design.md — Architectural rationale and implementation strategy:

  • Why a parallel reader (tailing session JSONL) rather than headless mode
  • Claude Code session JSONL format and fixed session ID injection
  • Adapter registry pattern for multi-agent support
  • Watcher lifecycle and IPC contract
  • Parser choice (@anthropic-ai/claude-agent-sdk for type safety)
  • Normalized event shape (AgentResponseEvent union)
  • Renderer surface (component structure, markdown extensions, image sanitization)
  • Streaming and performance considerations
  • Explicit v1 scope boundaries

• tasks.md — Detailed implementation checklist organized by subsystem:

  • Shared contract (IPC channels, types, AgentDef extension)
  • Backend watcher and adapters (fs.watch, line buffering, adapter dispatch, session ID injection)
  • Frontend store and panel (event log, UI components, markdown rendering, persistence)
  • Component and store tests
  • Persistence migrations
  • Acceptance criteria and validation steps

Implementation Notes

• The panel is opt-in per task and only shown when the agent has a configured responseSource
• Session IDs are fixed at spawn time via CLI flag injection, enabling deterministic file tailing
• The watcher uses forward-only reading with byte-offset tracking and fs.watch with stat-poll fallback
• Adapters normalize agent-specific events into a shared shape, keeping renderer code agent-agnostic
• Markdown rendering reuses existing marked + shiki + DOMPurify stack with extended sanitization for images
• The panel never blocks or replaces the terminal; input still flows through xterm
• Initial support targets Claude Code and Sourcegraph Amp; adapter registry enables future agents (Codex, Cursor, etc.)

This is a specification-only change; no source code is modified. The spec is ready for implementation planning.

https://claude.ai/code/session_01PNd94wGpSPm4bxcQsjWw3V