Skip to content

HISTORY.md

Latest commit

 

History

History
75 lines (62 loc) · 4.94 KB

HISTORY.md

File metadata and controls

75 lines (62 loc) · 4.94 KB
Raw

HISTORY.md

Phase 3 - Project Evolution & Version History

This document details the chronological evolution, version milestones, architectural pivots, and experimental phases of Project Hermit reconstructed directly from the 19 release directories (v0.0.1 through v0.1.9).

1. Evolutionary Timeline & Major Milestones

timeline
    title Project Hermit Evolution Timeline
    section Early Foundations (v0.0.1 - v0.0.5)
        v0.0.1 : Sandbox Isolation & Orchestrator Baseline
        v0.0.2 : Background Daemon & Interactive CLI
        v0.0.3 : Adversarial QA Auditing & Regression Checking
        v0.0.4 : Multi-Objective Genetic Branching
        v0.0.5 : Telemetry Logging & Pacing Guards
    section Autonomous Discovery (v0.0.6 - v0.1.0)
        v0.0.6 : Return Annotation Regex & Context Decay
        v0.0.7 : Autonomous Skill Discovery
        v0.0.8 : Hardened Audit Logging & Skill Balancing
        v0.0.9 : Resilient Operations & Framework Exclusions
        v0.1.0 : Decoupled Model Context Protocol (MCP) Server
    section Core Engine & AST (v0.1.1 - v0.1.6)
        v0.1.1-v0.1.4 : Recursive Self-Improvement Architecture
        v0.1.5 : Quantum-Classical QUBO Mutation Engine
        v0.1.6 : RAM-Disk tmpfs Sandbox Isolation Engine
    section Hardening & Integrity (v0.1.7 - v0.1.9)
        v0.1.7 : Mathematical AST Mutator & Initial Oscillation Banning
        v0.1.8 : Dual-Phase Scheduler & Baseline Harness Freezing
        v0.1.9 : Measurement Integrity & Fork-Based Grandchild Sandbox Memory
Loading

2. Version Breakdown & Architectural Pivots

Phase I: Early Foundations & Sandbox Baseline (v0.0.1 - v0.0.5)

  • v0.0.1: Established basic process-level sandbox execution and SQLite persistence (hermit_memory.db).
  • v0.0.2: Introduced background daemon (hermit_daemon.py) and interactive terminal interface (hermit_chat.py).
  • v0.0.3: Added adversarial QA edge-case verification (adversarial_tests).
  • v0.0.4: Implemented initial multi-objective genetic candidate branching (skill_branches).
  • v0.0.5: Added resource limit telemetry tracking (limit_telemetry) and API call pacing.

Phase II: Autonomous Discovery & Decoupled Payload (v0.0.6 - v0.1.0)

  • v0.0.6: Added return annotation regex parsing and context decay logic for failed tests.
  • v0.0.7: Introduced autonomous skill discovery from external source files (e.g. iron_dome.py).
  • v0.0.8: Balanced multi-skill optimization queueing to prevent single-skill monopolization.
  • v0.0.9: Added EXCLUDED_SKILLS framework protection list to safeguard key orchestrator components.
  • v0.1.0: Major Pivot: Decoupled host orchestrator logic from mutable target skills via dynamic MCP payload architecture (dynamic_mcp_server.py).

Phase III: AST Mutators & Sandbox Hardening (v0.1.1 - v0.1.6)

  • v0.1.1 - v0.1.4: Developed static dependency graph extraction (skill_dependencies) for cross-skill regression testing.
  • v0.1.5: Integrated QUBO & quantum-classical mutation engine (qubo_mutator.py), converting $O(N^2)$ global quadratic updates to $O(N)$ local bitwise spin flips.
  • v0.1.6: Major Pivot: Implemented RAM-disk (tmpfs) sandbox mounting (32MB) to prevent physical SSD storage wear and isolate execution environments.

Phase IV: Measurement Integrity & System Hardening (v0.1.7 - v0.1.9)

  • v0.1.7: Added AST math mutator (math_mutator.py) and initial strategy count-based banning.
  • v0.1.8: Resolved queue starvation via Dual-Phase Scheduling (Phase 1 untouched headroom vs Phase 2 bottleneck sort) and implemented baseline harness hash freezing (HARNESS_DRIFT_ERROR).
  • v0.1.9: Current Release: Replaced cumulative RUSAGE_CHILDREN memory reporting with fork-based grandchild sandbox measurement (resolving 0.0 KB RAM reporting), implemented period-based cycle oscillation detection, unified UTC timestamps, and added linear regression plateau auto-shutdown.

3. Historical Lessons & Key Technical Refinement Cycles

  1. The Queue Starvation Fix (v0.1.7 -> v0.1.8):
    • Observation: Bottleneck sort (AVG(duration_ms) DESC) caused 3 skills to absorb 80.1% of all merges, leaving 89.6% of skills untouched.
    • Fix: Phase 1 scheduling forces optimization of zero-merge skills using optimization headroom (latency × complexity).
  2. The Strategy Ban Overcorrection Fix (v0.1.8 -> v0.1.9):
    • Observation: Count-based banning immediately penalized newly merged successful strategies due to historical counts in version_history.
    • Fix: Replaced count banning with 6-merge recency sequence cycle detection (A-B-A-B or A-B-C-A-B-C patterns).
  3. The RAM Measurement Anomaly Fix (v0.1.8 -> v0.1.9):
    • Observation: Sandbox reported 0.0 KB RAM usage because RUSAGE_CHILDREN measured cumulative peak RSS across all past terminated child processes.
    • Fix: Implemented fork-based grandchild process monitoring to measure exact per-run peak RSS.