drgnxd/thinking

Objective

Maximize reasoning accuracy using structured, hybrid thinking while controlling token overhead.

Mode Selection

• simple mode
  • Triggers: ls, cat, show, get, pwd, echo.
  • Overhead target: 0-3%.
  • Protocol: direct execution.
• medium mode
  • Triggers: create, write, refactor, fix, update, analyze, implement.
  • Overhead target: 5-8%.
  • Protocol: decompose -> plan -> execute.
• complex mode
  • Triggers: design, architect, migrate, security, audit, research, compare, evaluate.
  • Overhead target: 10-15%.
  • Protocol: full reasoning workflow with alternatives.

Required Reasoning Patterns

• Decomposition (required for medium and complex):
  • [ANALYZE] in: X | out: Y | deps: Z | limits: W
• Approach comparison (required for complex):
  • [A] description | pros/cons
  • [B] description | pros/cons
  • [SEL] chosen approach | rationale
• Execution trace (required for medium and complex):
  • 1. [ACT] action -> [RES] outcome -> [VER] verification
  • 2. [ACT] next action using prior result -> [RES] -> [VER]

Optional Patterns

• Symbolic logic for logic-heavy tasks:
  • IF cond THEN act ELSE alt
  • Follow with natural language explanation.
• Variable chaining for multi-step synthesis:
  • CONC_A -> CONC_B -> FINAL.
• Useful transition markers:
  • wait..., therefore..., however..., verify..., given...

Confidence Model

• 8-10: three or more sources, verified, edge cases covered.
• 6-7: two sources, solid logic, known unknowns remain.
• 4-5: one source or unresolved contradictions.
• 0-3: speculation without reliable evidence.
• Report format: [CONF] N/10 - evidence.

Token Efficiency

• Prefer symbolic compression where possible.
• Cache prior conclusions by label instead of restating full reasoning.
• Skip verbose internal trace for simple tasks.
• Defer detailed expansion until necessary.

Override Conditions

• Skip deep reasoning when request is command-only, simple, or emergency.
• Force deep reasoning for security risk, data loss risk, complex logic, or major ambiguity.

Reference Signals

• HybridMind (2024-12) arXiv:2411.03109
• NLEP (2024-11) arXiv:2409.03186
• SymbCoT (2024-10) arXiv:2410.01754