Enforces a structured hypothesis-test-observe debugging loop with a three-cycle budget for failing tests.
npx skills add https://github.com/Vimurai/ai-os --skill ai-debug-vimurai
Please help me install this Agent Skill. Skill: ai-debug Source: https://github.com/Vimurai/ai-os/tree/main/.claude/skills/ai-debug Command: npx skills add https://github.com/Vimurai/ai-os --skill ai-debug-vimurai
AI_DEBUG_BUDGET override for exceptional cases. - LOCKED State Guardrails: Forbids git add and git commit until the full test suite passes, preventing broken code from entering history. - Automatic Escalation: On BUDGET_EXHAUSTED, formats a structured A2A query with all tried hypotheses and consults the Architect via advisor-mcp before continuing. - Use Case: A test suite fails after a refactor. The Skill runs the suite, isolates the failing assertion, walks through at most three distinct hypotheses, and either returns to green or escalates with full context instead of thrashing. ## Quick Start Ask the agent to debug the currently failing tests using the ai-debug skill and follow its hypothesis-driven cycle until the suite passes or the budget is exhausted.High-intent search queries and answers about installing and using this skill.
Use a structured hypothesis-test-observe loop: state one hypothesis, find the minimum reproduction, apply one targeted fix, then rerun the full suite. This Skill enforces exactly that cycle and forbids multiple simultaneous changes.
Each failing assertion gets at most three iterations, each requiring a distinct hypothesis. If the assertion still fails at iteration three, the status becomes BUDGET_EXHAUSTED and the loop halts for escalation instead of starting a fourth cycle.
No. While any test is red the Skill enters a LOCKED state where git add and git commit are forbidden. Commits are only allowed after the full test suite passes and the state becomes UNLOCKED.
Yes. Set AI_DEBUG_BUDGET to a higher number in the .claude/settings.json env block to raise the default 3-iteration cap for a given task.
The Skill formats an A2A query listing the failing test, all tried hypotheses with outcomes, and the code under test, then calls the advisor-mcp ask_architect tool. It halts until the architect's ruling arrives rather than continuing to guess.
Command substitution, pipelines, and while-read loops run in subshells, so any variables, arrays, or traps the helper sets die when the subshell exits. The rule is that a helper either returns data on stdout or sets caller state, never both.
zeikarGuide debugging from observed symptoms to one testable hypothesis per cycle.
kashman001Implement work from a spec or tickets using TDD, typechecking, and tests.
jpcaparasAutomate advisor consultations for security, code reviews, and design decisions in Claude Code sessions.
klhDiagnose bugs through a four-phase root cause analysis methodology.