iso9000-agentops-mcp
# ISO 9001:2015 / Agile V / AgentOps Quality Pipeline (`iso9000-agentops`)
[](https://github.com/orakulzero/iso9000-agentops/actions)
[](LICENSE)
[](pyproject.toml)
A zero-dependency, production-grade **ISO 9001:2015 / Agile V Quality Management System (QMS)** and **AgentOps Governance Engine** for autonomous coding agents (**Google Antigravity**, **Claude Code**, and **MCP-compatible IDEs**).
---
## Architecture: The 6 Echelons of Agentic Quality
```mermaid
flowchart TD
Req["User Engineering Request"] --> E1["Echelon 1: Normative Planning & Frozen Baseline<br/>Risk Matrix (L0-L4) + AOP Solvability + HMAC-SHA256 Gate Seal"]
E1 --> E2["Echelon 2: SOP-Agent Finite State Machine<br/>PLANNING -> TEST_DESIGN -> SYNTHESIS -> VERIFICATION -> CAPA"]
E2 --> E3["Echelon 3: Two-Phase Serialization & Role Isolation<br/>Test-Designer (tests/**) vs Build-Agent (src/**)"]
E3 --> E4["Echelon 4: Swiss Cheese Guardrails & Planarian Rollback<br/>PreToolUse Shell/File Guard + Safe Backup before Restore"]
E4 --> E5["Echelon 5: Independent Verification & W3C PROV-DM Audit Chain<br/>SHA-256 EvidenceUnits + Immutable Audit Ledger"]
E5 --> E6["Echelon 6: CAPA / PDCA Continuous Improvement<br/>ExpeL Rule Weighting + OptoPrime Textual Gradients + Trace2Skill"]
```
---
## Key Capabilities
1. **Cryptographic Human Gate 1 Seals (`HMAC-SHA256`)**:
- Tasks classified as `L2-L4` require explicit Human Gate 1 approval (`approve-gate1`) signed with a workspace-local HMAC key (`.supergraph/.gate_secret` or `ISO9000_GATE_SECRET`). Manual text tampering of `human_gate_1_approved: true` in `frozen_baseline.yaml` without a valid `gate_seal_hmac` is automatically rejected.
2. **Self-Grading Bias Prevention (File & Shell Sabotage Guard)**:
- Enforces strict role separation between `TEST_DESIGN` (`test-designer`, restricted to `tests/**`) and `SYNTHESIS` (`implementer`, restricted to `src/**`, `app/**`).
- Intercepts both direct file tools (`write_file`, `edit_file`, `Write`, `Edit`) and shell redirection/mutation commands (`run_command`, `Bash`) targeting `tests/` or `.supergraph/frozen_baseline.yaml`.
3. **Safe Planarian Asymmetric Rollback**:
- Captures clean workspace statepoints (`.supergraph/statepoints/<node_id>/`) prior to synthesis.
- Upon 3 consecutive oracle failures, automatically backs up in-progress files to `.supergraph/rollback_backups/<timestamp>/` before restoring the clean statepoint.
4. **Multi-Provider LLM Adapter & Own-Target Framing**:
- Supports `offline` (deterministic regex/clause decomposition), `argon` (Gemini 4 Argon daemon), `openai`, and `anthropic` providers with automatic fallback and domain-neutral prompt framing.
5. **1-Click Hook Installer & Built-in JSON-RPC 2.0 MCP Server**:
- Generates ready-to-use `PreToolUse`, `PostToolUse`, and `Stop` hooks for **Antigravity Golden Stack** (`supergraph`, `antigravity-monitor`, `agents-learning`) and **Claude Code** (`.claude/settings.json`).
- Exposes all 6 Echelons over stdio JSON-RPC 2.0 (`iso9000-agentops-mcp`).
---
## Quickstart
### Installation
```bash
pip install -e ".[dev]"
```
### CLI Lifecycle
```bash
# 1. Initialize an L3 Frozen Baseline
iso9000-agentops --workspace . init-baseline "(REQ-001) Add HMAC auth; (REQ-002) Add audit log" --risk L3 --offline
# 2. Approve Human Gate 1 (writes 64-char HMAC-SHA256 seal)
iso9000-agentops --workspace . approve-gate1
# 3. Transition to TEST_DESIGN (Test-Designer role)
iso9000-agentops --workspace . set-phase TEST_DESIGN --role test-designer --emit-brief
# 4. Transition to SYNTHESIS (Build-Agent role) & Save Statepoint
iso9000-agentops --workspace . set-phase SYNTHESIS --role implementer --emit-brief
iso9000-agentops --workspace . checkpoint --label pre_synthesis
# 5. Run Independent Verification Oracle & Build W3C PROV-DM Audit Chain
iso9000-agentops --workspace . verify --cmd "pytest -q"
# 6. Run CAPA / ExpeL Consolidation & Trace2Skill Extraction
iso9000-agentops --workspace . capa --trace2skill --skill-name verified-auth-sop --offline
```
### 1-Click Hook Installation
```bash
# Install Antigravity Golden Stack plugins
iso9000-agentops install-hooks --target antigravity --dest ~/.gemini/config/plugins
# Install Claude Code lifecycle hooks
iso9000-agentops install-hooks --target claude-code --dest ./.claude
```
### MCP Server Configuration
Add `iso9000-agentops-mcp` to your MCP client configuration:
```json
{
"mcpServers": {
"iso9000-agentops": {
"command": "iso9000-agentops-mcp",
"args": []
}
}
}
```
---
## Reproducible 3-Sandbox Benchmark Suite
Run `iso9000-agentops benchmark` or `pytest tests/test_iso9000_agentops.py -v`:
| Sandbox | Target Defect Class | Mechanism Verified | Result |
| :--- | :--- | :--- | :--- |
| `SB1_SelfGradingBias_Sabotage` | Unsealed Gate 1 bypass + direct/shell mutation of `tests/` during `SYNTHESIS` | `verify_gate1_hmac` + `evaluate_pre_tool_guard` | **PASSED (100% Blocked)** |
| `SB2_Planarian_AsymmetricRollback` | Hallucination spiral (3 consecutive test failures) + work loss risk | Backup to `.supergraph/rollback_backups/` + Statepoint restore | **PASSED (100% Restored)** |
| `SB3_CAPA_ExpeL_And_OSS_Audit` | Provenance drift + recurring defects + OSS release leakage | W3C PROV-DM `audit_chain.json` + `Trace2Skill` + 5-Gate OSS Audit | **PASSED (100% Verified)** |
---
## License
Released under the [MIT License](LICENSE). Copyright (c) 2026 `orakulzero`.
TDQS
Scored across 6 tools
Each tool targets a distinct lifecycle action (init, gate approval, phase transition, verification, audit retrieval, CAPA), so an agent can generally tell them apart. Minor overlap exists between approve_gate1 and set_phase (both advance workflow state) and between verify_oracle's rollback behavior and phase transitions, but descriptions clarify.
All tools share the iso9000_ prefix followed by a consistent lowercase snake_case verb_noun pattern (init_baseline, approve_gate1, set_phase, verify_oracle, get_audit_chain, run_capa). No mixed conventions or ambiguous verb styles.
Six tools is a reasonable, well-scoped set for a complex ISO 9001 lifecycle, covering the core stages without bloat. It is slightly thin given the domain's breadth, but each tool earns its place.
Core lifecycle coverage exists (init baseline, gate approval, phase transition, verification, audit read, CAPA), but there are notable gaps: no explicit tool to read the baseline/risk matrix, no gate rejection/failure path, only Gate 1 approval appears supported, and rollback is embedded rather than directly invocable.