Imported from Shivansh2703/agent-radar (AGENTS.md). Install upstream with npx skills add Shivansh2703/agent-radar. Copyright stays with the author.

Agent Radar — project instructions

TypeScript VS Code/Cursor extension that monitors Claude Code sessions and subagents. Live on the marketplace as Shivansh2703.agent-radar. All code lives in extension/.

Laws

• PUBLIC repo: commits, PRs, and code comments carry a single named author only — no co-author trailers, no "Generated with" or tool-attribution lines. Ever.
• Marketplace publish is owner-only. Never run vsce publish — stage the exact command and hand it to the owner.
• Test gate: cd extension && npm test green before any merge. The completion-detection/liveness precedence chain is the load-bearing subtle part — never change its timeouts or ordering without cited measurements (past regressions came exactly from unmeasured timeout guesses).
• Path-staged commits only (never git add -A); one logical change per commit.

Check

cd extension && npm test — compiles with tsc then runs the suite with node --test (no VS Code Extension Host, no display needed). ~6s, 302 tests: 299 run, 3 are the liveFire.test.ts live-fire cases, gated behind LIVE_FIRE=1 because they need a real running agent process to observe — the suite passes with them skipped by design, not by omission. CI runs the same command on every push and PR (.github/workflows/ci.yml), on macos-latest — this extension's real target is a developer's Mac (it inspects local processes and real shells), not Linux, and public repos get macOS runners free.

Trap: two tests in liveness.test.ts (~line 284, ~line 331) symlink a real /bin/zsh binary to stand in for a fake claude process. That only works on a host that ships zsh — on a Linux runner both fail with spawn .../bin/claude ENOENT, which names the fake claude path and gives no hint that the real problem is a missing shell. If CI ever moves back to Linux, this is why those two go red first.

Red means a real regression in the classify/collector/liveness logic, most likely the completion-detection precedence chain — re-read the rung comments in deriveSessionState before touching timeouts or ordering; past regressions came from unmeasured timeout guesses.

Not covered: anything that only breaks inside a real VS Code window — command registration, the dashboard webview rendering, status-bar rendering — since these tests never boot the Extension Host. Marketplace packaging (npm run package) and publishing are separate, manual steps this gate does not touch.

• Create and modify files with your editor tools (write/edit), never shell heredocs/redirection — those trip permission walls here.