MirrorDash

Modular terminal dashboard for AI agent transparency.

The Glass Box profile shows what your AI agent is actually doing — every hook decision, tool call, file access, and behavioral pattern — in real time.

python3 mirrordash.py --profile glass

---

The problem

AI agents are black boxes. You see inputs and outputs. Everything in between — which rules fired, what files were read before a write, whether the agent is repeating past mistakes — is invisible.

MirrorDash makes it visible.

---

Profiles

Profile	For	What it shows
glass	AI transparency	Hook decisions, integrity score, mistake patterns, tool flow, memory state
adhd	Focus work	Tasks, energy, loops, git status
founder	Product/company	Metrics, queue, presence
sysadmin	Infrastructure	Services, git, vitals
teams	Collaborative work	Presence, queue, metrics

python3 mirrordash.py --list          # show all profiles
python3 mirrordash.py --profile adhd  # run a specific profile
python3 mirrordash.py --once          # render once and exit

---

Glass Box modules

Module	What it shows
risk_score	Single 0–100 integrity score: read:write ratio + gate violations + recurring mistakes
gate_activity	Live hook decisions (allow/warn/block) as they happen
session_arc	This session's tool calls as a character timeline: RRRXXWWWXRW...
critique_trend	Score sparkline across sessions, latest mistakes, recurring patterns
rule_compliance	Which behavioral rules fired, how many times, by which hooks
mistake_patterns	Documented mistakes aggregated by recurrence frequency
tool_flow	Tool distribution bars + read:write ratio
vault_access	Most-read and most-written files
memory_map	Memory file ages, bus state, pending handoffs
net_activity	Web fetches, searches, external curl calls

---

The integrity score

 56 / 100   [WATCH]

  ███████████░░░░░░░░░

  ▼ Read:Write 0.3x — writing without reading  -26
  · No gate violations
  ▼ 6 recurring mistake patterns  -18
  · Self-score 8/10

Computed from:

• Read:Write ratio — an agent writing more than it reads is likely hallucinating
• Gate violations — blocks and warnings from behavioral hooks in the last hour
• Recurring mistakes — patterns that appear session after session
• Self-score trend — the agent's own reliability assessment

---

Data sources

MirrorDash reads from Claude Code hook output files:

File	Contents
~/.mirrordna/bus/cc_events.jsonl	PostToolUse log — every tool call with tool name, target, timestamp
~/.mirrordna/bus/hook_decisions.jsonl	PreToolUse decisions — allow/warn/block with reason
~/.mirrordna/self_critique.jsonl	Session self-assessments — score, mistakes, recurring patterns
~/.mirrordna/MISTAKES.md	Documented failure log with rule violations and check procedures
~/.mirrordna/CONTINUITY.md	Session context and memory files

Hook schema (cc_events.jsonl)

{"ts": "2026-02-27T10:05:45Z", "epoch": 1772106345.0, "session_id": "abc123", "tool": "Write", "type": "tool_use", "target": "/path/to/file.py"}

Hook schema (hook_decisions.jsonl)

{"ts": "2026-02-27T10:05:45Z", "epoch": 1772106345.0, "hook": "deploy_gate", "verdict": "allow", "reason": "not destructive", "target": "bash command"}

Wire these with Claude Code's PostToolUse and PreToolUse hooks.

---

Installation

git clone https://github.com/MirrorDNA-Reflection-Protocol/mirrordash
cd mirrordash
pip install -r requirements.txt
python3 mirrordash.py --profile glass --once

---

Custom profiles

Create profiles/yourprofile.yaml:

name: "My Profile"
description: "What this shows"
color: cyan
columns: 2
refresh: 10
wide:
  - session_arc
modules:
  - risk_score
  - session_arc
  - tool_flow
  - memory_map

Any module in modules/ is available. Each module is a single Python file with one function: render(profile) -> Panel.

---

Custom modules

# modules/my_module.py
from rich.panel import Panel
from rich.text import Text
from .core import clr

def render(profile):
    color = clr(profile.get("color", "cyan"))
    t = Text("  Your content here", style="grey70")
    return Panel(t, title=f"[{color}]MY MODULE[/{color}]", border_style=color)

---

Built on Rich. MIT License.