Epic: Migrate ADF Claude agents to Agent SDK

[AlexMikhalev]

Context

ADF spawns Claude agents as claude -p subprocesses with stdout/stderr capture. This causes SIGPIPE bugs, fragile output parsing, cold-start waste, and no native cost tracking. The Claude Agent SDK (renamed from Claude Code SDK) provides query() async iterator with typed hooks, native session resume/fork, and in-process cost tracking.

Proposal

Replace subprocess spawning for Claude-based agents with Agent SDK query() calls. Codex agents remain as subprocesses (no SDK equivalent).

Migration Strategy

1. Create thin Python wrapper per agent that calls Agent SDK query() and forwards typed JSONL events to stdout using the RPC envelope (Implement JSONL RPC envelope for agent-orchestrator communication #688)
2. Migrate one agent at a time: product-development first (lowest risk), then security-sentinel
3. Preserve orchestrator reconciliation loop, Safety/Core/Growth layers, NightwatchMonitor

What This Eliminates

• SIGPIPE bug (async iterator, no pipes)
• Stdout/stderr parsing fragility (typed message stream)
• Cold-start waste (native resume=session_id)
• Cost tracking complexity (hooks on every API call vs output parsing)
• Session management complexity (ClaudeSDKClient handles state)

Open Question: Bridge Architecture

• Option A: Rust orchestrator -> thin Python wrapper (subprocess) -> Agent SDK
• Option B: Rust orchestrator -> local HTTP server running Agent SDK -> typed responses
• Option C: Implement Agent SDK protocol directly in Rust (high effort, fragile)
• Spike needed to determine which approach best preserves Rust orchestrator while gaining SDK benefits

Sub-issues (to be created during design phase)

• Spike: validate Agent SDK wrapper bridge pattern
• Migrate product-development agent to Agent SDK
• Implement cost tracking via Agent SDK hooks (supersedes Phase 1: Cost tracking with budget enforcement for ADF agents #638 approach)
• Implement session resume via Agent SDK (supersedes Phase 2: Session persistence for Claude Code agents in ADF #639 approach)
• Migrate security-sentinel agent to Agent SDK

References

• Agent SDK reference: cto-executive-system/knowledge/claude-agent-sdk-reference.md
• Current ADF subprocess model: crates/terraphim_spawner/src/lib.rs
• Comparison table: Agent SDK reference, "Comparison with Current ADF Setup" section
• Plan: cto-executive-system/plans/adf-architecture-improvements.md (item 2.1)
• Depends on: Implement JSONL RPC envelope for agent-orchestrator communication #688 (RPC envelope)
• Related: Epic: Leverage Paperclip features into AI Dark Factory #637 (Paperclip epic), Phase 1: Cost tracking with budget enforcement for ADF agents #638, Phase 2: Session persistence for Claude Code agents in ADF #639