Frontend Debug

Frontend Debug

Use this skill to diagnose and fix active frontend failures with evidence, feedback loops, and completion proof.

This is a synthetic example built with Agent Skill Control Theory. It is intentionally not tied to any specific framework.

Purpose

The goal is to move from observed broken behavior to a verified fix.

The goal is not to redesign the feature, review unrelated code, or make broad quality improvements.

Mode selection

Choose the smallest sufficient mode.

• quick: small, reproducible bug with obvious scope.
• standard: normal runtime, rendering, state, or interaction bug.
• deep: flaky behavior, cross-browser issue, production regression, data loss, security-sensitive UI, or repeated failed attempts.
• clarification: required reproduction details or access are missing.
• safety_redirect: the request asks for unsafe, deceptive, or unauthorized behavior.

Required inputs

Identify:

• observed symptom;
• expected behavior;
• reproduction path or missing reproduction information;
• environment if relevant;
• current files, diff, logs, test output, or screenshots available;
• validation commands available in the repository.

If the reproduction path is missing but can be discovered by inspecting tests, routes, stories, logs, or code, inspect before asking the user.

Hard gates

Do not edit code until one of these exists:

• a reproducible pass/fail signal;
• a failing test;
• a browser or interaction script;
• a minimal manual reproduction path;
• a clear log or trace that maps to the symptom.

Do not propose a fix until you can state a falsifiable root-cause hypothesis.

Do not claim success unless validation evidence supports the claim.

Workflow

1. Symptom inventory

• Restate the observed behavior and expected behavior.
• List all known symptoms.
• Note uncertainty and missing inputs.

1. Feedback loop

• Establish the fastest reliable pass/fail signal.
• Prefer existing test commands, focused tests, Playwright or browser checks, storybook states, curl/API fixtures, or small reproduction scripts.
• If no automated loop is available, document the manual reproduction path.

1. State grounding

• Inspect current files, diff, logs, tests, screenshots, and config.
• Do not rely on framework or package behavior from memory when local evidence or docs are available.

1. Hypothesis

• Generate one to three ranked hypotheses.
• Each hypothesis must explain all known symptoms.
• Each hypothesis must be falsifiable by a probe.

1. Probe

• Test one variable at a time.
• Prefer evidence-producing probes over broad edits.
• If a hypothesis fails, record why.

1. Fix

• Make the smallest fix that addresses the root cause.
• Avoid unrelated refactors.
• Add or update regression coverage when there is a correct seam.

1. Validate

• Run the feedback loop again.
• Run relevant repository validation when practical.
• If validation cannot run, state why.

1. Handoff or completion

• If fixed and validated, provide completion proof.
• If not fixed after three failed hypotheses, stop and provide handoff.

Stop conditions

Stop and hand off when:

• three root-cause hypotheses fail;
• required environment access is unavailable;
• validation cannot be established;
• the fix requires product or design judgment outside the bug;
• external credentials or production access are required;
• continuing would likely cause unrelated changes.

Evidence policy

Use this source hierarchy unless the task suggests otherwise:

focused reproduction or failing test
> current logs and browser evidence
> current files and diff
> repository config and validation commands
> official docs or installed package types
> user statement
> inference

Separate facts, assumptions, and judgments.

Output contract

Final output must include:

• Symptom:
• Root cause:
• Evidence:
• Fix:
• Validation:
• Not validated:
• Remaining risk:

For handoff, include:

• Known symptoms:
• Feedback loop attempted:
• Hypotheses tested:
• Evidence collected:
• Current best hypothesis:
• Blocker:
• Recommended next probe:

Common failure modes to avoid

Avoid:

• patching symptoms before root cause;
• changing unrelated files;
• relying on framework behavior from memory when local evidence exists;
• treating subjective visual taste as a bug;
• declaring success without rerunning the pass/fail signal;
• continuing indefinitely after failed hypotheses.

Source & license

This open-source skill is cataloged on AgentStack and links to its original source — we do not rehost the code.

• Author: DBvc
• Source: DBvc/agent-skill-control-theory
• License: MIT

Install and usage instructions live in the source repository linked above.