# Latent Signal

- **Event:** [The IDE Reimagined: JetBrains Codex Hackathon](https://cerebralvalley.ai/e/jetbrains-x-openai-hack)
- **When:** Apr 18 at 9:00 AM – Apr 19 at 5:00 PM (PDT)
- **Where:** Shack15, San Francisco, CA
- **Placement:** Finalist
- **Team:** [Ann Catherine Jose](https://cerebralvalley.ai/u/annjose), [George Chiramattel](https://cerebralvalley.ai/u/georgeck)
- **GitHub:** https://github.com/latentsignal-org/periscope
- **Demo video:** https://youtu.be/aWSPgNUlkBA
- **Gallery:** https://cerebralvalley.ai/e/jetbrains-x-openai-hack/hackathon/gallery
- **Page:** https://cerebralvalley.ai/e/jetbrains-x-openai-hack/hackathon/gallery/12

Periscope is a local-first viewer that shows you exactly what's in your coding agent's context window, turn by turn, and tells you what to do about it.

What it does
- Visualize the context of any agentic coding session, grouped by turn, so you can see what's actually filling the window - tool outputs, file reads, stale plans
- Analyze the session to surface bloat, repetition, and rot.
- Recommend the next move - compact, branch off, start fresh, rewind so you can shape context instead of just watching it grow.
- It runs locally, so your session logs are protected
- Lives inside JetBrains as a plugin, so you can read your context next to the code you're writing.
- This could be built on ACP protocol since it works across agents - Codex, Claude Code, Copilot, Gemini, Cursor, Amp, and more - by parsing each agent's session format directly 

Why it matters
The success of an agentic coding session is determined by three things: the model, the harness, and the context. The first two get all the attention and keeps improving. Context is the one you control, but also the one with almost no visibility today. Long sessions silently degrade: context rot sets in, the agent starts contradicting itself, tool outputs crowd out the context. However, current agentic tools give do not provide enough visibility into the context to improve it. 

Our personal inspiration
The main inspiration to build this was our own observations on the context growing constantly, and quality degrading. We tried to solve it with the recommendations from Claude team (eg: https://x.com/trq212/status/2044548257058328723) to improve the context quality. We wanted to make it easy to use these techniques without constantly worrying about managing context, with a good balance of autonomy and control.

Tech stack
Local Go server + SQLite (FTS5) + embedded Svelte 5 SPA + JetBrains plugin. Open source on GitHub.

---

Markdown version of https://cerebralvalley.ai/e/jetbrains-x-openai-hack/hackathon/gallery/12. Site index for agents: https://cerebralvalley.ai/llms.txt · full text: https://cerebralvalley.ai/llms-full.txt