Node Inspect Debugger

SkillMonitoring & ops

See runtime state a console.log cannot reach — set real breakpoints, step, and dump the scope chain of a paused Node/TypeScript process. Use on "set a breakpoint", "inspect runtime state", "console.log isn''t enough", "step through this", "what''s in this closure at runtime", "attach a debugger". Not a logging or observability-setup workflow.

Available today. Use it from your connected AI after setup.

Connect ahel once, and every AI you use reads what you have installed.

Then ask your AI: use the Node Inspect Debugger skill

What this skill tells your AI

The instructions your AI receives, as published by blackbelttechnology/pi-agent-dashboard in packages/eng-disciplines/.pi/skills/node-inspect-debugger/SKILL.md and read by ahel’s review.

Overview

console.log is a guess with a print statement attached — you have to already suspect where to look and what to print. A breakpoint inverts that: you stop the world at a line and read every local and closure variable in the frame, walk the call stack, and evaluate expressions in the paused context. When the bug lives in state a log can't cheaply reach — a closure captured three calls ago, a paused async frame, the Electron main process, the internals of a long-lived WebSocket server — the inspector is the tool.

This skill is the TOOL half of a pair; the METHOD is systematic-debugging (its Phase 1 "gather evidence" and Phase 3 "test one variable" are exactly when you reach for a breakpoint).

When to Use

Reach for the inspector when the state you need is expensive or impossible to log:

  • The jiti server (packages/server, TypeScript run directly via jiti) — request handlers whose failure depends on accumulated in-memory state.
  • The restart orchestrator / PTY workers (restart-helper.ts) — a detached process whose closure state you can't console.log from the parent.
  • Dual WebSocket server closure state — connection maps and buffers held in closures across the bridge and browser servers.
  • The Electron main process — lifecycle/bootstrap state that never reaches a browser console.
  • The bridge extension — code running inside every pi session, where adding logs means reloading every session.

If a single well-placed console.log would answer the question, use the log. The inspector earns its setup cost when the state is deep, closure-bound, or in a process you can't easily instrument.

Two tiers

TierUse whenInterface
REPL (node inspect)interactive, one-off, you're drivingthe node inspect command REPL
Programmatic CDPscripted, repeatable, or attaching from another processscripts/cdp-inspect.ts (this skill)

Tier 1 — the node inspect REPL cheat-sheet

node inspect opens a REPL against the inspector. Core commands:

CommandDoes
sb('file.ts', N)set breakpoint at line N of file.ts (see jiti note below — .ts works directly)
ccontinue to next breakpoint
n / s / ostep next / step-in / step-out
btbacktrace (call stack of the paused frame)
list(5)show 5 source lines around the pause point
repldrop into a REPL evaluated in the paused frame's scope — inspect any local
watch('expr')re-evaluate expr at every pause
exec exprevaluate one expression in the paused frame
restart / killrestart / kill the inspected process

In repl mode you can read any in-scope variable by name — this is the fastest way to answer "what is x right now?".

Tier 2 — pi-dashboard jiti launch (spike-verified)

This repo runs TypeScript directly through jiti (no dist/*.js build). Launch the target with the inspector and jiti's register hook:

node --inspect-brk=<port> --enable-source-maps --import <jiti-register-hook-url> cli.ts
  • --inspect-brk=<port> halts at the very first line so you can set breakpoints before anything runs.
  • --enable-source-maps is not required for line-level breakpoints (jiti is line-preserving, see pitfall below) but keep it: it gives exact column precision and correct Error.stack line numbers.

Locating the jiti register hook (do not hard-code the path)

The register hook path differs across local / standalone / global installs. Resolve it the same way this repo's launcher (packages/server/bin/pi-dashboard.mjs) does — via createRequire, trying each supported jiti package:

import { createRequire } from "node:module";
import { dirname, join } from "node:path";
import { pathToFileURL } from "node:url";

const req = createRequire(import.meta.url);       // or createRequire(process.argv[1])
const JITI_PACKAGES = ["jiti", "@mariozechner/jiti"];
let hook;
for (const pkg of JITI_PACKAGES) {
  try {
    const pkgJson = req.resolve(`${pkg}/package.json`);
    hook = pathToFileURL(join(dirname(pkgJson), "lib", "jiti-register.mjs")).href;
    break;
  } catch { /* try next */ }
}
// pass `hook` to `node --import <hook> cli.ts`

Attaching to the live server (not a fresh launch)

To debug the already-running server instead of a cold --inspect-brk launch, start it with the inspector enabled and discover the port:

NODE_OPTIONS="--inspect=0" pi-dashboard start        # 0 = pick a free port
curl -s http://127.0.0.1:<port>/json/list            # find the webSocketDebuggerUrl / port

--inspect=0 avoids a port collision when something already holds the default 9229.

Pitfalls

The upstream "emitted JS" pitfall does NOT apply to jiti — corrected

A common guide warns: "breakpoints hit the emitted JS, not your .ts, and node inspect doesn't follow sourcemaps." This is false for this repo. jiti transpiles line-preserving and registers the compiled JS under the .ts URL (1:1 line alignment). There is no separate dist/*.js file and no sourcemap indirection to fight.

Consequences you can rely on:

  • sb('cli.ts', 42) in the plain node inspect REPL binds directly to line 42 of the .ts source and hits.
  • Debugger.setBreakpointByUrl('cli.ts', line) over CDP binds to the .ts line.
  • CDP may report sourceMapURL: "" (jiti attaches none) — this does not mean the breakpoint won't bind. It binds and hits anyway.

Do not port the emitted-JS workaround; it's solving a problem jiti doesn't create.

Pending breakpoints return empty locations but still hit

A breakpoint set before the target script has parsed (e.g. right after the --inspect-brk halt) returns locations: [] at set-time. This is not a failure — the breakpoint is deferred and resolves and hits once the script parses. Setting a breakpoint after the script has parsed returns populated locations. Do not treat an empty locations array at set-time as "the breakpoint didn't take."

Programmatic CDP — scripts/cdp-inspect.ts

For scripted or repeatable inspection, this skill ships a dependency-free TypeScript helper. It uses Node 24's global WebSocket — no chrome-remote-interface dependency. It attaches to a paused target, sets a .ts breakpoint, resumes past the entry halt, and on the hit prints the paused frame plus every local and closure variable:

# 1. launch the target halted (see jiti recipe above), noting <port>
# 2. attach + break at a line inside a function:
npx tsx packages/eng-disciplines/.pi/skills/node-inspect-debugger/scripts/cdp-inspect.ts <port> <ts-url> <line>
# prints:  PAUSED at <file>:<line> fn=<name>
#          local    cfg = Object
#          local    doubled = 42
#          closure  label = "HELLO"

Use it as the ready-made "dump the frame" step whenever the REPL's interactivity isn't worth it.

Verification

  • The launch recipe used jiti's register hook resolved via createRequire, not a hard-coded path
  • .ts breakpoints were set directly (no emitted-JS workaround)
  • An empty locations at set-time was treated as deferred, not failed
  • The paused frame's locals were read (via repl, exec, or cdp-inspect.ts) — the actual state, not a guess

Signals

GitHub stars
283
Forks
41
Last commit
Sep 2026
Advanced
Catalog kind
skill
Gateway key
node-inspect-debugger
Source
github.com/blackbelttechnology/pi-agent-dashboard