feat(cli): PTY-based Claude backend for interactive billing

[Jacky040124]

Summary

• Adds ClaudePTYBackend that spawns Claude Code in an interactive PTY session instead of claude -p, so usage is classified as "interactive" rather than "programmatic" under Anthropic's June 15 billing split
• Reads all events (text, thinking, tool-use, tool-result) from Claude's session JSONL file, preserving full streaming — no UI degradation
• Switched via ALOOK_CLAUDE_BACKEND=pty env var; default (pipe) behavior unchanged

New files

• src/cli/daemon/agent/claude-pty.ts — ClaudePTYBackend using Bun.Terminal API + session JSONL polling
• src/cli/daemon/agent/ansi-scanner.ts — Responds to Ink's DA1/DA2/DSR/XTVERSION terminal queries to prevent TUI hang

Key design decisions

• PTY output is NOT parsed for content — only used for lifecycle control (readiness detection + /exit)
• Business data comes entirely from ~/.claude/projects/<encoded-cwd>/sessions/<id>.jsonl
• Incremental byte-offset JSONL reading for performance on long sessions
• Zero new npm dependencies (uses Bun native PTY API)

Test plan

• Set ALOOK_CLAUDE_BACKEND=pty and run a full task: prompt → thinking → tool-use → tool-result → final reply
• Verify AgentMessage stream matches pipe mode output
• Verify timeout handling (PTY startup timeout, execution timeout, JSONL inactivity)
• Verify no zombie processes after task completion
• Verify ALOOK_CLAUDE_BACKEND=pipe (default) still works unchanged
• Verify on Max subscription that usage is not billed to Agent SDK pool

🤖 Generated with Claude Code

[gusye1234]

Review: Request Changes

The PTY-based Claude backend concept is sound, but there are several blocking issues:

Blocking:

1. JSONL path encoding is likely incorrect — The code assumes Claude Code encodes cwd by "replacing / with -", but the actual encoding is more complex (percent-encoding or base64 depending on version). If wrong, the JSONL file will never be found and the backend silently produces zero messages. This needs verification against the actual Claude Code implementation.

2. Premature /exit on first end_turn — If Claude produces multiple assistant messages (e.g., after tool use), the first end_turn stop reason triggers /exit, killing the session before tool results are processed and the final response is generated. Multi-turn tool-use flows will break.

3. Hardcoded --permission-mode bypassPermissions — This is extremely permissive. Claude can execute any tool without confirmation. This should at minimum be configurable or inherited from the caller's options, and documented as a security trade-off.

4. Monolithic execute() method (~280 lines) with deeply nested closures. The JSONL parsing, PTY lifecycle, and message queueing should be separate methods for testability and maintainability.

Non-blocking:

• ❯ readiness detection is fragile — will break if Claude Code changes its prompt character.
• No input sanitization on prompt before writing to PTY stdin. Control characters or /exit in the prompt could break the session.
• stripAnsi regex is incomplete for all CSI sequences. Consider a well-tested pattern.
• lastOutput naming is misleading — it only stores the last text block, not all output.

Recommend addressing items 1-4 before merging.