Quality

Lesson Learned analyzes recent code changes via git history and extracts software engineering lessons. Use when the user asks 'what is the lesson here?', 'what can I learn from this?', or wants to reflect on recent work.

Captures screenshot and element snapshot on browser errors during e2e testing, saving to gitignored .ai/scratch/e2e/. Invoke when 'Something went wrong', auth errors, blank pages, or unexpected redirects appear. A building block that captures and returns, leaving further action to the caller.

Browser Feature Testing drives the running Carbon app through a specific user flow in a real browser to verify a feature or fix works end to end. It builds a plan from the diff or an issue, tests with agent-browser, and caches successful playbooks for reuse.

Orchestrates the bug-fix lifecycle from root cause analysis to implementing and proving the fix with a mandatory red→green regression test. Use when fixing a bug, test failure, or unexpected behavior.

Runs pre-commit verification gates (types, format, typecheck, tests, i18n) in order, automatically fixes linting and formatting errors, and commits only when all checks pass. Use after completing a task or making changes to ensure code quality before committing.

A systematic debugging approach: add temporary JSONL logging to the real code path, reproduce the bug, and analyze runtime events to find the exact point of divergence before fixing. Use when static analysis gives low confidence or the bug involves runtime state, ordering, concurrency, or manual reproduction.

Captures screenshot and element snapshot of current browser page when an error appears during e2e testing. Saves to gitignored .ai/scratch/e2e/. Use when agent-browser snapshot shows failure indicators. Building block: captures and returns, caller decides to continue.

An end-to-end bug-fix pipeline that diagnoses root causes, implements minimal fixes with regression tests, and optionally verifies and commits. Use for fixing bugs, test failures, or unexpected behavior.

Interviews the user relentlessly, one question at a time, to resolve every open decision in a plan, spec, or design idea. Cross-checks answers against the codebase and domain glossary, records resolutions immediately, and surfaces contradictions. Supervised use only.

A supervised interview skill that resolves every open decision in a plan or spec by asking one question at a time, cross-checking answers against the codebase, and recording resolutions immediately. Use when you need to thoroughly stress-test a design before implementation.

Audits the Carbon codebase as a senior advisor to surface improvement opportunities (bugs, security, performance, test coverage, tech debt, migrations, DX) and produces self-contained implementation plans for other AI agents to execute. Strictly read-only; never modifies code.

Generates a self-contained HTML page that explains a pull request's purpose, changes, and verification, helping reviewers quickly understand complex changes without reading the full diff.