Enhancements vs original (file1)

• PASS_THRESHOLD — configurable via env var (default 0.7). Reviewer verdict is overridden to FAIL if confidence < threshold, even if the model says "pass".
• Per-issue severity — each reviewer issue now has severity: low|medium|high. Any high-severity issue overrides the verdict to FAIL.
• Override logging — when the rule overrides a model "pass" verdict, the summary field explains why.
• Version bumped to 2.1.0

Stack B — Python Reviewer Package (port 8000)

What it does: Standalone reviewer Python package — Router + PrimaryAgent + ReviewerAgent + HTTP server. Runs fully offline with ScriptedLLM (no API key needed). Supports OpenAI-compatible endpoints for live mode.

Run

cd reviewer

# Scripted demo — all 4 scenarios, no API key:
python demo.py
python demo.py escalate         # just the escalation scenario

# HTTP server for trace_viz.html frontend:
python -m reviewer.server       # → http://localhost:8000  POST /run

# Live LLM (any OpenAI-compatible endpoint):
export OPENAI_API_KEY=sk-...    # OPENAI_BASE_URL / LLM_MODEL optional
python demo.py --live "Reconcile INV-88 $12,400 vs PO $9,900" --exposure high

API Reference

Use as a library

from reviewer import HumanAgentOS, ReviewerAgent, LiveLLM, ScriptedLLM

# Just the reviewer:
critique = ReviewerAgent(LiveLLM()).review(task_text, agent_draft)
if not critique.passed:
    ...  # reroute with critique.issues, or escalate

# Full loop:
result = HumanAgentOS(LiveLLM(), max_retries=1, pass_threshold=0.7).run(
    {"text": "...", "risk": "medium", "reversibility": "low", "exposure": "high"})
print(result.to_dict())

Frontend

Both UIs are in frontend/. Open them directly — no build step.

The Dispatch Console has a 🔍 Trace Viewer → button in the top-right that opens trace_viz.html in a new tab.

To point index.html at a deployed backend:

<script>window.HAOS_API_BASE = "https://your-backend.com";</script>

Demo script (3 scenarios, ~2 minutes)

Use the four preset buttons in index.html:

Then open trace_viz.html to show the Escalate scenario — the structured critique with per-dimension scores and fix hints, the reroute step, and the final escalation.

Architecture

User submits task
       │
       ▼
  ┌─────────┐
  │  ROUTER  │  ← Risk score, category, route
  └─────────┘
       │
  ┌────┴──────────────────────┐
  │                           │
human_only               agent_only / mixed
  │                           │
  ▼                           ▼
Queue for human          ┌─────────┐
(no AI touches)          │  AGENT  │ ← Draft output
                         └─────────┘
                              │
                    ┌─────────┴─────────┐  (PARALLEL)
                    │                   │
             ┌──────┴──┐         ┌──────┴──────────┐
             │ Quality  │         │   Compliance    │
             │ Reviewer │         │   Reviewer      │
             └──────┬──┘         └──────┬──────────┘
                    └─────────┬─────────┘
                              │
                   Both pass? (conf≥threshold AND no high-severity)
                              │
                    YES ──────┤────── NO (retry with critique)
                              │
                    agent_only: done          Retries exhausted?
                    mixed: human sign-off           │
                                             YES → ESCALATION → Human

Reviewer pass rule (legible, not a black box)

PASS iff model says pass AND confidence ≥ threshold (default 0.7) AND no high-severity issue.

On any API failure, reviewers fail safe — they do not pass, so the task escalates for human review.

Built with

Merge notes