v0.15.1 — parallel_tasks + 4-Layer Code Review + Autonomous Ops

Highlights

parallel_tasks MCP tool (NEW)

• Execute multiple tasks simultaneously with automatic model routing
• 2-axis automatic model selection: task type × input complexity → optimal Gemma4 variant (e2b / e4b / 26b / 31b)
• asyncio.gather-based parallel execution for light tasks (e2b/e4b), sequential for heavy tasks (31b+) to avoid GPU contention
• Benchmark: 5 parallel tasks on 501-line input — e2b+e4b mixed 51 s / 10 GB VRAM (vs 31b single 130 s / 20 GB)

4-Layer Code Review Pipeline (code_review MCP tool)

• Layer 2: gemma4 ReAct review with web_search + RAG ($0)
• Layer 3: Sonnet 4.6 verification + cross-file analysis (~¥10)
• Layer 4: Opus 4.6 meta-review — summary-only, no source code (~¥5)
• Codex consultant for P1 issues (on-demand)
• Empirical: gemma4+RAG ($0) outperforms Codex GPT-5.3 (~¥50) in finding uniqueness
• skip_sonnet=True for daily $0 reviews, codex_consult=True for release gates

Codex reasoning effort control

• VALID_CODEX_EFFORTS = {"none","minimal","low","medium","high","xhigh"}
• codex_effort parameter on code_review, think(provider="codex"), agent_task(provider="codex")
• Default: high
• Auto-escalation: when the 4-layer pipeline detects ≥3 P1 issues, Codex is invoked with xhigh automatically

Department RAG (dept_*)

• Per-department Qdrant collections: dept_hr / dept_research / dept_design / dept_build / dept_qa + mem0_shared
• Exposed via dept_search / dept_store MCP tools

Autonomous operations harness (scripts/)

• system_auditor.py — periodic integrity and drift audit
• anomaly_dispatcher.py — routes detected anomalies to the right department
• env_self_heal.py — auto-repairs common environment regressions
• critical_files_guard.py — SHA-256 snapshot protection for CLAUDE.md, settings.json, etc. (30 generations)
• helix_overview.py — single-command 9-domain overview
• supervisor.py — watches 9 resident daemons, auto-restart on failure
• Audit → dispatch → heal chain under Windows Task Scheduler

Stats

• MCP tools: 27 (was 23)
• Tests: 347 passing
• 7 OSS MCP servers in the tsunamayo7 ecosystem
• Runs on RTX 5070 Ti 16GB + RTX PRO 6000 96GB dual-GPU (auto-tiering based on VRAM)

Install

git clone https://github.com/tsunamayo7/helix-agent.git
cd helix-agent
uv sync
uv run python server.py

Full changelog: https://github.com/tsunamayo7/helix-agent/blob/master/CHANGELOG.md

🤖 Generated with Claude Code

v0.15.0 — 4-Layer Code Review Pipeline + Qwen3-VL Vision

🔍 4-Layer Automated Code Review Pipeline

A multi-LLM code review system that achieves 100% issue coverage at ~¥30 ($0.20):

Layer	Model	Role	Cost
L2	gemma4:31b (local)	ReAct review with web_search + RAG	$0
L3	Sonnet 4.6	Cross-file verification	~¥10
L4	Opus 4.6	Meta-review (no code reading)	~¥5
Codex	GPT-5.3	Consultant (P1 only)	on-demand

Empirical Results (5-Model Comparison)

gemma4 + RAG ($0) outperforms Codex GPT-5.3 (~¥50) in code review findings. The combined 4-layer pipeline catches every issue found by any individual model.

🌐 gemma4 Context Expansion

• web_search tool: Qdrant RAG + SearXNG — gemma4 now accesses external knowledge
• Auto-categorization: 9 categories for memory organization
• 12 tools total: gemma4 operates as a fully autonomous ReAct agent

📷 Qwen3-VL 32B Vision/OCR

• 95%+ OCR accuracy on Japanese text (vs gemma4's 60-70%)
• Phone numbers, postal codes, addresses — all correct
• Free-form vision analysis with mode="auto"

🛡️ Security & Operations

• ReAct injection defense (5 rules for untrusted search results)
• 2-hour periodic health checks (Task Scheduler)
• Enhanced daemon monitoring

Stats

• 347 tests passing (17 new for code review pipeline)
• 22 files changed, 2,508 lines added
• $0 daily review cost with gemma4 alone

v0.4.0: ReAct Agent Loop — Local LLM as Autonomous Agent

What's New

helix-agent evolves from a reasoning proxy to an autonomous agent. Local Ollama models can now reason step-by-step, use tools, and iterate to solve complex tasks.

ReAct Agent Loop (Phase 1)

• agent_task MCP tool: multi-step reasoning with tool use
• Prompt-based ReAct: works with ALL Ollama models (not limited to function-calling models)
• JSON structured output: reliable action parsing with automatic retry on parse errors
• Configurable: max_steps, model selection, tool filtering

Built-in Tools (Phase 2)

Tool	Description
read_file	Read file contents (PathGuard protected)
write_file	Write to files (PathGuard protected)
list_files	List directory contents
search_in_file	Regex search within files
run_command	Execute shell commands (allowlist: git, python, uv, ollama)
calculate	Safe math expression evaluation
search_memory	Qdrant semantic search (placeholder)

PathGuard Security

• Allowlist directories: only C:\Development and C:\Users\tomot\Documents by default
• Sensitive file blocking: .env, credentials, SSH keys, certificates
• Path traversal prevention: resolves symlinks and blocks ../ attacks

Progress Notifications (Phase 3)

• MCP ctx.report_progress() for each agent step
• Claude Code can track agent progress in real-time

Live Test Result

Task: "Read pyproject.toml and tell me the version and dependencies"
Step 1: thought="I need to read the file" → action=read_file → observation=(file contents)
Step 2: thought="I have the answer" → action=finish → answer="v0.3.0, deps: fastmcp>=2.0, httpx>=0.27"
✅ Correct, 2 steps

Stats

• 144 tests passing (+25 from v0.3.1)
• 7 built-in agent tools
• New files: src/react_loop.py, src/tools.py, src/builtin_tools.py, src/pathguard.py

Full Changelog: v0.3.0...v0.4.0

v0.3.0: Local Benchmark + Model Override

What's New

Local Benchmark Engine

• 8 automated tests: code generation (FizzBuzz, string manipulation), reasoning (logic, math), instruction following (JSON output, list format), Japanese (translation, summarization), and speed (tokens/sec)
• Auto-scoring: regex + pattern matching validators
• Persistent cache: results saved to ~/.helix-agent/benchmarks.json
• New model detection: automatically identifies unbenchmarked models

Benchmark-Aware Routing

• Benchmark scores directly influence model selection priority
• Category-specific scoring (code, reasoning, speed) for optimal matching
• models list now includes benchmark scores

Model Override

• models(action="use", model_name="...") — lock all tasks to a specific model
• models(action="use_auto") — switch back to auto-selection

Stats

• 82 tests passing
• 9 files changed, 1112 insertions

Full Changelog: v0.2.0...v0.3.0