Agent Standup (GitHub repository)

The repo behind Agent Standup: a zero-dependency Node.js tool that turns Claude Code's local session files into a live "team call" visualization of the agent tree. One 607-line server.js (plain http, no npm dependencies) reads the JSONL transcripts Claude Code already writes under ~/.claude/projects (honouring CLAUDE_CONFIG_DIR), parses the main session file plus one file per sub-agent, and merges them into a single timestamped event timeline. The same server ships four ways: a VS Code extension (extension.js), a CLI (bin/agent-standup.js, including a statusline mode that reads Claude Code's status-line JSON on stdin), an Electron desktop app with a tray indicator, and npx for one-shot use.

The core abstraction is the event timeline: each transcript line becomes one of a handful of event kinds (user, say, tool, order, report, finish, usage, results), where order/report events reconstruct who delegated to whom by matching the Agent/Task tool-use id to the sub-agent whose .meta.json records it. Consecutive tool calls by one agent merge into a single activity "beat". The reader is incremental — each file is parsed once and only appended bytes are read, in ≤16 MB chunks with careful handling of half-written last lines, and a 3-session LRU cache with per-file stamps.

Deliberate economics: observation costs zero tokens. The only model calls are opt-in (French translation and "quips"), made via the local claude CLI with warm persistent workers, or the API directly if ANTHROPIC_API_KEY is set. Board state for many parallel sessions is derived from a 192 KB tail read per file, classifying sessions into working / waiting / blocked / turn / idle heuristically from the last pending tool call. Security is unusually careful for a local server: localhost-only host/origin checks, Sec-Fetch-Site filtering on side-effect endpoints, path validation, and lstat to refuse symlinked character files. The whole frontend is one 211 KB public/index.html — no framework, no build step.