# ContextPane

- **Event:** [Meta Global AI Developer Hackathon](https://cerebralvalley.ai/e/meta-ai-developer-hackathon)
- **When:** Oct 26, 12:00 PM – Nov 4, 9:00 PM (EST)
- **Where:** Online
- **Team:** [Timur Dudhasch](https://cerebralvalley.ai/u/dudhasch)
- **GitHub:** https://github.com/dudhasch/contextpane
- **Demo video:** https://www.youtube.com/watch?v=4c1xI3tcBdg
- **Gallery:** https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery
- **Page:** https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/85

ContextPane is a human-in-the-loop cloud remediation workbench for resolving any failed Azure or M365 incidents problems when e.g. a support ticket does not name the resource or the engineer needs to correlate data about the incident. Instead of jumping across eight separate admin portals (Entra ID, Intune, Activity Logs, Azure, Azure IAM, licensing, ticket queues etc), the engineer opens the ticket, sees the correlated identity and incident evidence on one screen, steers and reviews a sandbox-validated fix, and approves it — with a one-click revert.

HOW CONTEXTPANE WORKS 
ContextPane pairs a left-hand Ticket & Steering Pane, a center Diagnostic Canvas, and a right-hand Code Rail:

1. Cross-Portal Context Gathering & Hydration (Muse Spark 1.3):
When an incident opens or an engineer highlights error text, the orchestrator queries three Model Context Protocol (MCP) servers (Microsoft Graph, Azure IaC, and ERP/Ticket Queue) in parallel to pull correlated identity, device, activity-log, and policy records. Muse Spark 1.3 hydrates the diagnostic layout and checks whether our catalog already has a view for that failure mode.

2. Constrained 5-Block Diagnostic View Assembly (Muse Wonder 1.0):
When no saved view matches an unfamiliar policy failure, Muse Wonder 1.0 assembles a validated view specification from five building blocks — quota-gauges, policy-chain, data-table, config-diff, and event-stream — rendered by our local React interpreter (SynthesizedDynamicCard.tsx). No generated code is dynamically executed. The spec is saved to dynamic-catalog.json so future tickets of that class open as an instant catalog hit with zero model calls.

3. PII Vault, ptr:// Data Plane & Open-Weights Hover Diagnostics (Muse Glimmer 30B):
Prompts pass through a local PII vault that replaces UPNs, names, and IPs with session tokens, while correlation datasets travel as opaque ptr:// pointers resolved locally by the orchestrator. Muse Glimmer 30B open weights (GGUF Q4_K_M) run under llama-server on our Modal L40S container (glimmer_open_weights.py) with strict GBNF JSON output to power inline hover diagnostics, cached locally in hover-cache.json.

4. Human-Steered Code Rail, Isolated Modal Sandbox & One-Click Revert:
Nothing mutates a tenant without an explicit engineer approval click. Every PowerShell 7 and HCL bundle is checked in an isolated, credential-free Modal sandbox (contextpane_sandbox.py) running PowerShell AST parsing and Terraform 1.9.8 validate (azurerm ~> 4.0). Validation errors are fed back to Muse Wonder 1.0 for up to 2 repair attempts before the engineer reviews the diff. Every execution records a pre-change state snapshot for one-click rollback.

## More from Meta Global AI Developer Hackathon

- [Hearth](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/81)
- [Hokfam](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/82)
- [BARRED](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/83)
- [Volgat](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/86)
- [VideoLab](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/87)
- [Doodytoon](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/88)

---

Markdown version of https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/85. Site index for agents: https://cerebralvalley.ai/llms.txt · full text: https://cerebralvalley.ai/llms-full.txt