Celestial Observer

Access to Celestial Observer is restricted to operators authorized by the tenant administrator. The console supports authentication via Firebase (email and password) or corporate OIDC providers configured in the client's infrastructure. Every session is cryptographically bound to the operator's identity claims.

Authentication Flow. Enter the corporate email and password provided by the tenant administrator. On first access, the system will require a forced password change. After authentication, UI permissions are dynamically assembled from the user's Firebase custom claims: the interface will only display tabs and actions you are authorized for.

The Command Center is the task-oriented landing page of the dashboard. It provides a real-time overview of your AI platform operations and surfaces proactive action cards for tasks that require your attention.

The interface is divided into two main areas. The Status Grid shows five top-level indicators: Connection status, Agent health, Telemetry freshness, Wiki health, and Inspection readiness. The Action Grid provides quick links to common workflows like investigating traces, reviewing wiki changes, drafting governed changes, syncing the wiki, and checking logs.

The System Health tab provides real-time operational metrics and traffic analysis for the AI agent. It is the primary view for monitoring platform performance and diagnosing elevated latency or error rates.

Key Metrics. The top stat cards display Total Requests, Token Volume, Avg Latency, and Success Rate for the selected time range. The large Traffic Analysis area chart visualizes this data over time, allowing you to identify peaks and anomalies.

Service Pulse. The right panel monitors the health of individual subsystems (API Gateway, LLM Inference, RAG Service, Vector DB, etc.). A green pulsing dot indicates the service is healthy, amber indicates elevated latency or timeouts, and red indicates a failure.

The User Analytics tab provides an audit trail of end-user interactions with the AI platform, enabling operators to monitor adoption, identify high-volume users, and investigate user-specific issues.

The User Interaction Audit table lists users by identity (email or user ID), showing their total request volume, token consumption, and last active timestamp. Clicking any row opens a detail panel with that user's full session history. The table can be sorted by any column and filtered by email or user ID.

For deployments utilizing the Meta Cloud API integration, this tab provides dedicated omnichannel telemetry. It tracks inbound and outbound volume, delivery status, token consumption specific to the channel, and estimated costs.

The Connection Card at the top left shows the current provider, Phone Number ID, WABA, mode, daily limit, billing model, and reply window. The Message Flow area chart visualizes inbound and outbound volume over time. The Delivery Status panel breaks down messages by status: Delivered, Read, Failed, and Pending. The Message Events table at the bottom allows you to click Open to launch the Flight Recorder for any specific WhatsApp message.

The Wiki Workspace is the center of knowledge governance for the agent. Every business rule, FAQ, policy, and tool configuration that the agent uses in production is managed here, with full Git versioning and authorship traceability.

Workspace Header. The top strip displays the current state of the repository. You will see badges indicating if the workspace is Ready, Unsaved, or Review Pending. The Sync Wiki button pushes approved changes to the runtime agent. The Refresh button re-fetches the latest state from the remote repository.

TinaCMS Editor. The main editing interface uses an embedded TinaCMS instance. It allows you to edit Markdown files with visual frontmatter controls. The left sidebar shows the repository file tree (e.g., content/policies, prompts, tools). Ensure YAML frontmatter is valid before saving.

Wiki Health & Integrity. The right panel continuously monitors the health of the Wiki connection. Git Sync verifies if the local TinaCMS session is in sync with the remote Git repository. Tina Session confirms the authentication session is active. Document Count displays the total number of indexed documents across all collections.

The Wiki Copilot is an AI assistant designed to help operators author, validate, and safely stage changes to the agent's knowledge base. It enforces governance rules and ensures changes follow the correct schema before they can be merged.

All changes to the Wiki, whether proposed by operators via Copilot or generated by the agent as Memory Candidates, must pass through the Review Queue before becoming official institutional memory. This is the governance firewall that prevents AI-generated errors from corrupting the knowledge base.

Review Pending. Displays active Pull Requests. You can review the branch name, document path, and head commit hash. Actions include Approve, Merge & Sync, and Close. Approving a PR does not automatically sync it to the live agent: you must also click Merge & Sync.

Memory Candidates. These are potential long-term memory items identified automatically by the agent during runtime interactions. Operators with the wiki.approver claim must explicitly Approve or Reject each candidate. Approved candidates are committed to the wiki as governed knowledge. This flow breaks the negative feedback loop described in the Institutional Memory Governance spec.

The Aura Flight Recorder v2 is the deep-trace instrument of Celestial Observer. It captures a complete, immutable orbit log of exactly how the AI agent processed a specific transaction, from initial intent signal to final output, across every tool invocation and LLM cognitive sequence.

KPI Bar. The modal header displays the Transaction ID and overall trace status. The KPI bar provides quick metrics: LLM CORE, the model used (e.g. gpt-4o); RUNTIME, total execution time; TOKEN_LOAD, total tokens consumed; OBJECTS_GEN, structured objects generated; NETWORK_OPS, external tool and API calls.

Flow (Visual) Tab. This tab visualizes the agent's execution pipeline chronologically. It starts with the User Intent Input and any associated feedback. Below, the Agent Execution Pipeline displays a sequence of tool invocations. Each step card shows the tool name, duration, exact prompt arguments sent to the tool, and the operational result returned. This is the primary view for understanding why the agent took a specific action.

Brain (LLM) Tab. This tab exposes the raw cognitive process of the LLM. The left column lists the Cognitive Sequence Index: every individual inference call made during the transaction, each with its timestamp and token count. Selecting a step reveals the exact Cognitive Injection (the full system prompt, context, and user data sent to the model) and the Model Response Output. This is the definitive view for debugging hallucinations, prompt injection issues, or unexpected model behavior.

The Audit Trail provides an immutable log of security events, administrative actions, and data access across the platform. It is the compliance and forensics layer of Celestial Observer.

Events are categorized by severity (INFO, WARN, ERROR) and type (AUTH, ADMIN, DATA_ACCESS). Each entry includes the user ID, masked IP address, and specific action details. The Audit Distribution chart on the right shows the breakdown of event types over the selected period. The Sign-In Configuration panel shows the active authentication provider and its connection status.

The Settings modal is the configuration hub for the dashboard. It is accessible from the sidebar and provides access to all platform-level configuration options.