Debugging failing tests

This workspace uses typed ports to fire the agent automatically when a process crashes. No copy-pasting errors into chat.

Nodes on the canvas

Node	Role
Process Manager with script npm test --watch	Keeps tests running in a loop.
Code	Opens the test file and the suspect handler.
Terminal	For manual checks the agent suggests.
Agent Terminal	Receives the automatic prompt on crash.
Text (named "Failure log")	Logs go straight here.
Git	To see what changed recently that might have broken.

Connections

From → To	Kind	What it does
Process Manager → Text	port ports:logs->appendText	Every line from the test runner becomes text.
Process Manager → Agent Terminal	port ports:crashes->prompt	A crash fires an automatic prompt.
Agent Terminal → Code	permission code-read	Agent opens and reads files.
Agent Terminal → Terminal	permission execute	Agent runs a single test to confirm a hypothesis.
Agent Terminal → Git	permission git-log	Agent looks at recent commits.
Agent Terminal → Text	permission summary	Agent writes down what it found.

How it works

1. The Process Manager brings npm test --watch up when the workspace opens (field autoStart: true).
2. Every code change reruns the test runner. Logs appear in real time in the Text via ports:logs->appendText.
3. When the test runner exits with a non-zero code, ports:crashes->prompt fires — the Agent Terminal receives an automatic prompt containing the stack trace.
4. The agent investigates: reads the test file, reads the related handler, checks the last 5 commits via Git.
5. Proposes a fix, writes the diagnosis into the "Failure log" text.

Prompt the agent receives (template)

The default prompt for the crashes->prompt port is editable on the Process Manager. Suggested template:

A process managed by Riggr crashed. Here's the stack trace:

{{logs}}

Investigate: read the test file mentioned in the error, read the
handler it covers, look at the last 5 commits via Git to see if
anything recent broke. Propose a fix (just explain, don't apply).
Write the diagnosis into the "Failure log" text node.

{{logs}} is replaced with the most recent lines that crossed the port.

Why this setup is powerful

• You don't have to be watching the terminal. You're in another tab, another tool — when you come back, the diagnosis is waiting.
• The agent works read-only on the code (code-read). It proposes, you review. Nothing gets applied without you.
• The Text is durable — it becomes a log of "this broke on day X because of Y". Useful at retro time.

Variations

• If you trust it more, swap code-read for code-full-access and add to the prompt: "if the fix is trivial, apply it and tell me what changed".
• For a whole team, save the workspace as a template and check it into the repo. Each person clones and starts contributing.