Add TODOS checklist and DELEGATE context-isolated subtask features

[BillJr99]

Summary

This PR introduces two new agent coordination primitives to the CLK orchestration framework:

1. TODOS: A lightweight, mutable per-turn checklist that sits between the append-only progress log and heavyweight charter/plan. Agents maintain a short list of items (todo/doing/done) that overwrites on each emission (last-write-wins).

2. DELEGATE: A context-isolation mechanism for bounded subtasks. An agent can hand off a self-contained task to a child agent that runs in a fresh session without access to the caller's blackboard, but may perform real work (file writes/commits). The child's distilled result returns as a single post.

Key Changes

TODOS Checklist System

• New module clk_harness/orchestration/todos.py: Complete TODOS implementation with parsing, persistence, and rendering

  • parse_todos_blocks(): Extracts checklist items from agent responses (last complete block wins)
  • apply_todos_blocks(): Persists per-author checklists to .clk/state/todos.json (mutable, last-write-wins)
  • render_todos(): Formats checklist for prompt injection via $todos placeholder
  • Thread-safe per-author storage with metadata (timestamp, stage_id, workflow)

• Integration points:

  • TranscriptMixin._apply_todos(): Parses and persists TODOS blocks from agent responses (skipped in meta phases)
  • PromptsMixin._collect_context(): Injects agent's current checklist into next prompt
  • runner.py: Added _todos_lock for thread-safe concurrent access
  • Template updates: Added _TODOS_PROTOCOL_BLOCK documentation to prompts

• Tests: Comprehensive unit tests in tests/test_todos.py covering parsing, persistence, and rendering

DELEGATE Context-Isolation System

• New parser parse_delegate_proposals() in casting/director.py: Extracts DELEGATE blocks with target, context, and task

• Core implementation TranscriptMixin._apply_delegate():

  • Spawns context-isolated child agents with fresh session (no blackboard inheritance)
  • Enforces depth cap (max_delegate_depth, default 1) and per-turn limit (max_delegates_per_turn, default 2)
  • Guards against unknown targets, self-delegation, and cycles
  • Child's POST blocks suppressed; distilled result returned as single delegate_result post
  • Child may perform real work (ACTION blocks execute under child's name)

• Configuration: New robustness settings in clk.config.json:

  • max_delegate_depth: Nesting limit (default 1)
  • max_delegates_per_turn: Per-turn dispatch cap (default 2)
  • delegate_result_max_chars: Result truncation cap (default 2000)

• Tests: Integration tests in tests/test_robustness_integration.py and unit tests in tests/test_delegate_parser.py

PI Extension (TypeScript)

• New tool clk_todos: Allows agents to update their checklist with status marks ([ ]/[~]/[x])
• New tool clk_delegate: Dispatches bounded subtasks to context-isolated children
• State management: setTodos() and getTodos() in state.ts for checklist persistence
• Consensus module: runDelegate() function that constructs isolation preamble and distillation instructions
• Tests: Unit tests for both tools and the delegate runner

Quality & Response Validation

• response_quality.py: Added malformed TODOS block detection (balanced TODOS:/END_TODOS check)
• Prompt templates: Added protocol documentation blocks for both TODOS and DELEGATE

Configuration & Documentation

• Updated .env.example with new DELEGATE settings
• Updated CONFIGURATION.md with robustness tuning guidance
• Updated .gitignore to exclude scratch/ (context-offload working directory)

Implementation Details

https://claude.ai/code/session_01KUsgsFFA4ovessfnpsesTW