feat: subagent tool, undo system, history compaction, TUI shortcuts

[OrionArch]

Summary

• Subagent tool: New subagent_run tool for spawning stateless sub-agents that execute tasks in parallel within their own context window
• Undo system: UndoHistoryManager with /undo command for multi-turn file edit rollback
• History compaction: CompactActiveSession with /compact command to summarize older rounds and reclaim context window space
• CompactHistory hook: Wired across all LLM adapters (Anthropic, OpenAI, Genkit)
• TUI prefix shortcuts: ! (shell run), # (memory save), & (background run)
• /skill command: List and inspect registered skills from ~/.iroha/skills/
• Tests: Integration tests for compaction, unit tests for subagent handler and undo manager

Changed files (14 files, +833 / -20)

Core

• pkg/agent/runner.go — undo manager, compact history hook, session_id context injection
• pkg/agent/session_store.go — CompactActiveSession implementation
• pkg/agent/pool.go — refactored ExecuteMessageInDir, dynamic model getter hooks
• pkg/agent/tools.go — registered subagent_run tool
• pkg/agent/auto_review.go — added subagent_run to low-risk tool classification

New files

• pkg/agent/tools_subagent.go — subagent spawn and execute handler
• pkg/agent/tools_subagent_test.go — unit tests for subagent
• pkg/agent/undo_test.go — unit tests for undo manager
• pkg/agent/compaction_integration_test.go — integration tests for compaction

LLM adapters

• pkg/llm/adapter.go — added CompactHistory to AdapterHooks interface
• pkg/llm/anthropic.go, openai.go, genkit_adapter.go — call CompactHistory before message conversion

TUI

• pkg/tui/update_keys.go — prefix shortcuts (!, #, &) and slash commands (/compact, /undo, /skill)

Test plan

• go test ./pkg/agent/... — all existing and new tests pass
• go test ./pkg/llm/... — retry and adapter tests pass
• Manual TUI testing of !echo hello, #prefer dark mode, &npm run build, /compact, /undo, /skill
• Verify subagent spawns correctly and returns summary

🤖 Generated with Claude Code