Sereno

The problem

Sereno started with the problem of following several agent sessions at once. A live process does not tell you whether it is writing, inside a command, waiting for a reply or stopped.

The five CLIs record their activity in different formats. Sereno reads those local histories and returns comparable states without copying real prompts or depending on a remote service.

Sereno How it works

1. Histories stay where the tools wrote them

   Sereno reads the local transcript queues of Claude Code, Codex, Gemini, Kilo and Antigravity. Each reader maps its format to the same states without uploading the content elsewhere.

2. The last useful event determines the state

   A tool_use without its tool_result means the session is in a command. The stop_reason and later events separate writing, waiting and stopped.

3. The cache follows mtime

   A transcript is only parsed again when its modification time changes. Refreshing stays cheap even with many sessions open.

4. No networking is enforced by a test

   The application opens no connections. An AST test rejects imports from networking modules, so that guarantee does not rely on someone remembering a rule while editing the file.

5. The demo never touches real sessions

   Demo mode uses synthetic histories in an isolated environment. The public capture shows the states without client names, prompts or machine data.

Sereno Where the bar sits

• 1 file Python 3.8+
• 0 dependencies outside the standard library
• 5 CLIs with comparable states