Phase 2: Session persistence for Claude Code agents in ADF

[AlexMikhalev]

Parent Epic

#637

Summary

Store Claude Code session_id from run results and use --resume on Safety agent restarts to avoid cold-starting.

Changes Required

terraphim_spawner

• Extract session_id from Claude stream JSON init and result events (already parsed in Phase 1)
• Expose session_id via AgentHandle or CostEvent

terraphim_orchestrator

• Add session_state.json file: {agent_name: {session_id, last_run_at, total_tokens}}
• On Safety agent restart, if session_id exists, add --resume <session_id> to CLI args
• Add --resume arg injection to spawn_agent() when session state available
• Clear session state on max-restart-exceeded or critical drift (fresh start)

terraphim_spawner/config.rs

• Add with_resume(session_id: &str) method to AgentConfig
• For claude CLI: --resume <session_id> before task prompt

Spike Required

• Verify claude -p --resume <session_id> works in non-interactive mode
• Verify session survives process exit and can be resumed

Reference

• Paperclip session state: packages/db/src/schema/agent_runtime_state.ts
• Claude CLI: --resume flag (verified in 2.1.70+)

Acceptance Criteria

• Session ID captured from Claude Code output and persisted to session_state.json
• Safety agent restarts use --resume when session exists
• Session cleared on critical drift or max restarts
• Codex agents unaffected (no session support)
• Atomic file writes (write to .tmp then rename)

[AlexMikhalev]

Storage update

Use terraphim_persistence::Persistable trait for SessionState instead of agent-state.json. See #637 comment for rationale. Key pattern: sessions/{agent_name}. Load from fastest backend on startup, save to all backends on session update.