Inspector Panel

What the Inspector is

The Inspector is the right-side panel in the AI Partner UI. It gives you a live window into what the agent knows, what it's thinking, and how it's reasoning — at any point during a conversation or goal execution.

Toggle it with the Inspector button in the top-right header, or use the keyboard shortcut Ctrl+I.

---

Inspector sections

CORE MEMORY

Shows what the agent has loaded as its active context for this session:

No user preferences found — means your USER.md is empty or hasn't been filled in yet. Click Edit Memory to open the workspace editor.

Once configured, Core Memory shows a summary of your USER.md content: name, role, current focus, and communication style.

CONTEXT USAGE

A token counter showing how much of the model's context window is currently in use:

Context usage: ~614 / 8,000 tokens
████░░░░░░░░░░░░░░░░  7.7%

The limit shown is the effective context window for the currently selected model. As a conversation gets longer:

• At 30 messages: ConversationCompressor summarizes older history, freeing space
• At ~70% usage: the Inspector shows a yellow warning
• At ~90% usage: the Inspector shows a red warning and compression is forced

This counter helps you understand why a very long conversation might be less coherent — the model has less context to work with.

RECENT EVENTS

A live feed of the last few agent reasoning phases. Each line is one step in the OODA loop:

Phase label	What it means
REFLECT	Agent is reviewing the outcome of the last action
ORIENT	Agent is updating its world model with new information
DECODE	Agent is parsing tool output (structured data extraction)
DECIDE	Agent has chosen the next action
ACT	Agent is executing a tool call
ASSESS	Agent is evaluating whether the goal is complete

Example from the screenshot:

REFLECT  Chat completed after 1 iteration(s)
DECODE   Iteration 1/25 – Analyzing next action…
ORIENT   Using model: MiniMax-M2.7

This tells you: the last chat completed in 1 step, the current goal is on iteration 1, and the model being used.

PERSONA

Shows the active agent persona — the system-level identity the agent is operating under for this session. This is derived from:

• workspace/SOUL.md (behavior rules, tone)
• workspace/USER.md (your identity, injected as context)
• The active agent profile (if you routed to @fin-analyst or similar)

Example:

You are a senior strategic co-pilot — part management consultant,
part market analyst, part execution engine. You operate at the
intersection of AI, technology strategy, markets, and high-stakes
decision-making. You think in frameworks, communicate in...

RECENT MEMORY

The last few memory entries the agent retrieved during this session — what it looked up from episodic, biographic, or vector memory to inform its responses:

• User told me to which functionalities you have access t...  Steps: 1  Tools: 0
  00:47:02

• User: thats gud
  00:43:18

• User: image kamasutra positions
  00:40:14

• User: create the json of this workflow which can be exec...  Steps: 1  Tools: 0
  00:44:32

• Goal: COMPLETED {"goal_summary":"Create an executable JSON workflow file for Viral Video Reverse Engineering process","outcome":"success","iterations_used":2,"tools_used":[...]...
  00:16:18

Click any entry to expand the full memory content.

ACTIVE TASKS

Shows any in-progress goals or scheduled tasks:

No scheduled tasks

When a goal is running, each sub-task appears here with its status (pending / running / completed / failed).

---

The OODA loop in the Inspector

AI Partner uses an OODA loop (Observe → Orient → Decide → Act) for reasoning. The Inspector's Recent Events feed shows each phase in real time.

Phase	What the agent does
Observe	Gathers new information (tool result, user message, memory retrieval)
Orient	Updates its model of the situation — "what does this new info mean?"
Decide	Chooses the next action based on the updated model
Act	Executes the chosen tool or generates a response

Watching this in real time helps you understand why the agent made a particular choice — especially useful when debugging unexpected behavior.

---

Browser live preview

When the agent is using computer use (browser automation), the Inspector shows a live screenshot of the browser state:

• The current page is streamed as a JPEG thumbnail, updating every 2 seconds
• If a CAPTCHA is detected, the screenshot expands and a "Take Control" button appears
• After you solve the CAPTCHA, click Resume to hand control back to the agent

---

Debug mode

Toggle Debug OFF → Debug ON in the Inspector header to see:

• Raw JSON of every tool call and result
• Full system prompt (not summarized)
• Memory search queries and raw results
• Confidence scores on skill matches

Debug mode is useful for developers extending the system or diagnosing why a goal isn't behaving as expected.

---

Inspector vs. Goal Progress panel

	Inspector	Goal Progress panel
Shows	Agent's internal state and reasoning	Goal execution status and sub-tasks
Best for	Understanding why the agent did something	Tracking what is happening
Updates	Every reasoning step	Every iteration
Location	Right side (always available)	Right side (only during goal execution)