debuglisted
Install: claude install-skill Myra-Agents/skills
# Debug — Interactive Runtime-Evidence Debugging (any language)
Fix bugs with **runtime evidence**, not guesses. You stay in the loop:
```
Don't guess → Hypothesize → Instrument → [ask you to Reproduce] → Analyze → Fix → [ask you: fixed?] → Verify
```
This skill runs on the **main thread**, so it can ask you questions directly (via
`AskUserQuestion`) at the two human-in-the-loop gates: **reproduce** and **is it
fixed?**. The mechanism is language-agnostic: a localhost HTTP log server receives
NDJSON from instrumented code — browser, Go/Rust/Python/Node service, CLI, shell
via `curl` — and the logs are read directly from a file.
## Hard rules
1. **Never fix without runtime evidence.** Get values from a real run first.
2. **Never guess.** Unsure ⇒ add more instrumentation and reproduce again.
3. **Never ask the user to copy-paste console output.** Capture server-side, `cat` the log file.
4. **Never remove instrumentation before the user confirms the fix.** Keep `#region debug` blocks until verified.
5. **Tag every log with the hypothesis it tests.**
## Locate the bundled scripts
```bash
# Searches common install locations across agents, then the cwd as fallback.
# -L follows symlinks (e.g. a global ~/.claude/skills/debug → repo checkout).
SCRIPTS="$(dirname "$(find -L \
~/.claude/skills/debug .claude/skills/debug \
~/.config/opencode/skill/debug ~/.codex/skills/debug ~/.cursor/skills/debug . \
-path '*debug/scripts/debug_server.js' 2>/dev/null | head -1)")"
```
##