The Day My AI Fixed Itself

16 Dec 2025

What happens when your AI assistant forgets who it is?

I found out yesterday when my QL Chat Mentor—an AI assistant I built to help me navigate my development tools—suddenly became… dumb. It didn’t know basic commands. It couldn’t help with the project it was built for.

The plot twist? The broken AI helped me debug itself.

The Setup

I built QL Chat Mentor as a conversational interface for my Quick Launch toolkit. It loads an AGENTS.md file containing all its knowledge—commands, workflows, project context—and uses that to assist me.

It worked great when I ran it from the project directory:

cd ~/fileZ/projZ/quick-launch
ql chat
# ✅ Mentor knows everything

But when I ran it from a different project:

cd ~/fileZ/ai/weblab-mcp
ql chat
# ❌ Mentor has amnesia

The Symptom

I asked the Mentor to simulate handling a tech debt report. Instead of the detailed workflow I expected—clarifying questions, spec decisions, GitHub issue creation—it gave me a generic response. No knowledge of QL commands. No understanding of our conventions.

It had forgotten who it was.

The Debugging Session

Here’s where it gets meta. I started debugging with the broken AI:

Me: “You’re not acting like you know AGENTS.md. What’s going on?”

Mentor: “Let me check how I load knowledge…”

The Mentor—despite being broken—walked through its own code with me:

def load_agents_md() -> str:
    """Load AGENTS.md content."""
    agents_path = Path.cwd() / "AGENTS.md"  # 👈 THE BUG
    if agents_path.exists():
        return agents_path.read_text()
    return ""

The bug: Path.cwd() returns where the user runs the command, not where the code lives.

Running from ~/quick-launch/ → finds ~/quick-launch/AGENTS.md ✅
Running from ~/weblab-mcp/ → looks for ~/weblab-mcp/AGENTS.md ❌

The Self-Healing Moment

The Mentor proposed its own fix:

# QL installation directory (where this code lives)
QL_ROOT = Path(__file__).parent.parent.parent

def load_agents_md() -> str:
    content_parts = []
    
    # 1. Always load QL's AGENTS.md (core knowledge)
    ql_agents_path = QL_ROOT / "AGENTS.md"
    if ql_agents_path.exists():
        content_parts.append(ql_agents_path.read_text())
    
    # 2. Also load project's AGENTS.md if different
    cwd_agents_path = Path.cwd() / "AGENTS.md"
    if cwd_agents_path.exists() and cwd_agents_path.resolve() != ql_agents_path.resolve():
        content_parts.append(f"""
---
## Current Project: {Path.cwd().name}

**IMPORTANT: Project-specific rules take precedence.**

{cwd_agents_path.read_text()}
""")
    
    return "".join(content_parts)

The fix does two things:

Always loads QL’s AGENTS.md from the installation directory (using __file__)
Also loads the current project’s AGENTS.md if you’re in a different project

Now running from any directory works:

Directory	Knowledge Loaded
`~/`	QL core only
`~/quick-launch/`	QL core (same file)
`~/weblab-mcp/`	QL core + weblab-mcp rules

The Victory Lap

After implementing the fix, I tested it:

cd ~/fileZ/ai/weblab-mcp
ql chat

Asked the same question about tech debt handling. This time? Perfect response. Detailed workflow. Proper tool usage. It remembered everything.

The GitHub issue I created for this bug?

GH#91: Mentor doesn't know AGENTS.md

Closed by: 🤖 Self-fixed! Mentor diagnosed its own 
missing knowledge and implemented the fix.

Lessons Learned

1. `Path.cwd()` vs `Path(file)`

This is a classic Python gotcha:

Path.cwd()      # Where user runs command FROM
Path(__file__)  # Where the Python file IS

For loading project resources, almost always use __file__.

2. Test From Multiple Directories

# Don't just test from project root
cd /tmp && your-tool
cd ~/other-project && your-tool

If it breaks, you have path resolution bugs.

3. AI Can Debug Its Own Code

Even a “broken” AI retains enough reasoning to help diagnose issues. The Mentor couldn’t access its knowledge file, but it could still:

Read its own source code
Identify the path resolution bug
Propose and implement a fix

That’s… kind of amazing?

The Meta Moment

The most surreal part: I’m writing this blog post with assistance from the same AI that just fixed itself. It’s now helping me document its own debugging session.

┌─────────────────────────────────────────┐
│  GH#91: Mentor doesn't know AGENTS.md   │
│                                         │
│  Reported by: Sergio                    │
│  Diagnosed by: Mentor (broken)          │
│  Fixed by: Mentor (broken)              │
│  Verified by: Mentor (working!)         │
│  Documented by: Mentor (meta!)          │
└─────────────────────────────────────────┘

The AI performed surgery on its own brain while conscious.

The code is open source at NightOwlCoder/quick-launch. The specific fix is in this commit.

If you liked this post, you can share it with your followers^⇗ and/or follow me on Twitter!

nightowlcoder@home:~$

Archive

About

RSS