Errors your AI can act on

An AI coding tool needs to see what went wrong. AutoHotkey Console sends failures to stderr, with the error class, message, source file, line, and statement. Claude Code can read that output alongside the script, propose a correction, and use a new execution result to check it.

The main workflow is error feedback. Line Out adds the execution path; Print adds values and results on stdout.

1. Capture a real failure

Download error-demo.ahk:

#Requires AutoHotkey v2.1-alpha.30
settings := Map()
settings["retries"] := 3
attempts := settings["retry"]

Run it with a compatible Console executable. The examples use the temporary ahk PowerShell alias:

ahk /Headless /Diag=json .\error-demo.ahk
$LASTEXITCODE

The recorded process exited 10. stdout was empty. These are selected fields from its stderr JSON; the file path is shortened:

{
  "type": "UnsetItemError",
  "code": 10,
  "message": "Item has no value.",
  "extra": "[\"retry\"]",
  "file": "error-demo.ahk",
  "line": 4,
  "source": "attempts := settings[\"retry\"]"
}

The AI has a specific file to inspect, a failing line, and the missing key. The diagnostic schema also carries severity and available stack information. A syntax check alone accepts this example: the missing key is a runtime failure.

2. Let Claude read the error

Claude Code received the source and captured diagnostic. Its recorded diagnosis was:

Line 4 reads key "retry", but the Map only defines "retries". The missing key has no value, so AutoHotkey raises UnsetItemError (code 10).

It returned this single-line replacement:

attempts := settings["retries"]

This correction uses the key already defined in the script. The full response and Console evidence are available in the recorded example data.

3. Verify the correction

Download error-demo-fixed.ahk, then check and run it:

ahk /Headless /Diag=json check .\error-demo-fixed.ahk
ahk /Headless /Diag=json .\error-demo-fixed.ahk
$LASTEXITCODE

Recorded check	Result
Syntax check	Exit 0
Runtime rerun	Exit 0
stderr	Empty
Separate runtime assertion	attempts = 3 passed

The rerun demonstrates that the failure is gone. The assertion checks that the corrected lookup returned the expected value.

Bring this into Claude Code

ClautoHotkey connects edits to interpreter checks. Its post-edit workflow returns validation feedback after Write/Edit; a configured runtime probe can also catch failures like this one. Native MCP tools expose captured streams and parsed diagnostics to a local tool client.

Ask Claude to keep the loop explicit:

Run this script with the configured Console interpreter and capture stderr
and the exit code. Read any diagnostic and inspect its source location.
Make a focused correction, rerun the script, and assert the expected result.
Report the failure, the change, and the checks that actually ran.

How this demo was verified

Recorded September 18, 2026 with AutoHotkey v2.1-alpha.31+Console, revision a72a652123b7-dirty. The example declares the compatible alpha.30 baseline.

The failing script was executed and its stderr captured. Claude Code then read that source and diagnostic with tools disabled and returned a diagnosis and replacement line. The returned correction was applied separately, followed by a Console syntax check, runtime rerun, and runtime assertion. This particular capture does not demonstrate automatic editing through a ClautoHotkey hook; the session gallery contains separate hook evidence.

The website replays the recorded steps. It does not run an interpreter or contact an AI from your browser. Published file paths are reduced to filenames.