Agent Handover

# Called before each LLM call
def check_handover_needed(session):
    used_tokens = estimate_token_count(session.context_window)
    max_tokens = session.config.window_size
    margin = session.config.handover_margin  # e.g., 4000 tokens

    if used_tokens + margin > max_tokens:
        handover_doc = generate_handover(session)
        start_fresh_session(handover_doc)
        return True

    return False

# Pseudocode: how handover is loaded into a new session

def start_session(handover_doc=None):
    system_prompt = load_system_prompt()

    if handover_doc:
        # Insert the handover as structured context at the top
        # of the new session's system prompt
        user_prompt = """User prompt {original_last_user_prompt} ## Handover Doc ... """

    # Start the prompt loop. The LLM sees the handover
    session = start_prompt_loop(system_prompt, user_prompt)
    return session

## User's Most Recent Request

"check for resumability bugs and edge cases . the agent should be able to resume its work at any point if it was working no matter what tool it was calling, what subagent it was working on. it should also be able to preserve main agent-subagents hierarchy when resuming"

## Current State of the Code

### Completed fixes (all tested, 480 tests passing):

1. Partial multi-tool-call resume . Added _get_pending_toolcalls() method to Agent class.
   resume_dangling_tool_work now detects and executes remaining tool calls when only some
   were executed before a crash.

2. Vision tool outputs saved to DB . _execute_tool_call now calls save_message() for
   vision messages (role "user" with structured content + tool_call_id). Fixed save_message/
   load_messages in chat_history_db.py to JSON-serialize list/dict content and deserialize on load.

3. Handover preserves subagent metadata . _perform_handover now calls get_session_info()
   to copy parent_session_id, toolcall_id, and depth into the new session, preserving the hierarchy.

4. Redirect chain resolution . Added resolve_session_id(), get_session_info().
   is_session_finished now follows redirects. dispatch_subagent.handle resolves child sessions
   through the redirect chain.

5. Vision messages trigger resume . _has_dangling_tool_work now detects role == "user"
   with tool_call_id as a tool response. resume_dangling_tool_work uses tool_call_id presence
   instead of role == "tool".

6. Agent init resolves redirects . Agent.__init__ calls resolve_session_id(session_id)
   so stale redirected sessions load from the active session.

7. Changes/files across redirect chain . Added get_all_changes_by_session_chain() and
   get_all_session_files_chain(). Updated dispatch_subagent.py and todo_list.py to use them.

### Files modified:

  src/Agent/agent.py . _has_dangling_tool_work, _get_pending_toolcalls, _execute_tool_call,
    resume_dangling_tool_work, _perform_handover, __init__
  src/Agent/chat_history_db.py . save_message, load_messages, is_session_finished,
    + resolve_session_id, get_session_info, get_all_changes_by_session_chain,
    get_all_session_files_chain
  src/Tools/dispatch_subagent.py . redirect resolution in handle, chain functions
  src/Tools/todo_list.py . get_all_session_files_chain in handle_execute_next_task
  tests/test_resumability_bugs.py . 30 new tests covering all fixes

## Important Files and Symbols

  src/Agent/agent.py
    Agent.resume_dangling_tool_work, Agent._has_dangling_tool_work,
    Agent._get_pending_toolcalls, Agent._execute_tool_call,
    Agent._perform_handover, Agent.__init__

  src/Agent/chat_history_db.py
    resolve_session_id, get_session_info, is_session_finished,
    save_message, load_messages, get_all_changes_by_session_chain,
    get_all_session_files_chain

  src/Tools/dispatch_subagent.py
    handle, _collect_subagent_output

  src/Tools/todo_list.py
    handle_execute_next_task

  src/interactive.py
    run_interactive (calls resume_dangling_tool_work at start)

  tests/test_resumability_bugs.py . all 30 tests
  tests/test_dangling_tool_resume.py . existing tests (still passing)
  tests/test_subagent_session.py . existing tests (still passing)

## Errors, Blockers, or Failed Attempts

  No current errors. All 480 tests pass.
  Initial code_search tool failed in the first session . worked around by reading files directly.

## Exact Next Step

  Look for additional edge cases in the todo_list.py crash recovery path . specifically
  handle_execute_next_task when a task's toolcall_id doesn't match any child session
  (e.g., if the child session was deleted or the toolcall_id was never set). Also consider
  testing the full end-to-end resume flow with a mock LLM to verify the agent correctly
  continues after partial tool execution + handover.

$ raggie code /path/to/project

Raggie Agent (code) v0.1.1
Indexing codebase...
  Index up to date (3 files changed since last index)

[INFO] Found incomplete todo list from previous session:
  Task 2/5: "Update API routes to use new validation" (PENDING)
  Resume this todo list? [Y/n]: Y

Resuming task 2/5...
    GetFileCodeSemantics(file_path="src/api/routes.py")