kstack

Validate the branch diff against the task's acceptance criteria before marking it done. Use when implementation looks complete, before a status update to done/awaiting_review, or when the user asks "did we cover everything?"

Get a PR to merge-ready state by triaging comments, resolving conflicts, and fixing CI. Includes Graphite detection, migration conflict handling, and conflict resolution. Use when the user asks to babysit a PR, get a PR merge-ready, or fix CI on a PR.

Audit implementation plans for the current user, verify status against the codebase, and archive completed plans. Use when the user asks to clean up plans, audit plan status, or archive completed plans.

One-pass AI code review of a PR or branch diff (tech-lead style). Use when the user asks for a code review, PR review, AI review, or wants code quality, architecture, bugs, and patterns reviewed.

Simplify changed code in a PR or branch for readability while preserving behavior. Use when the user asks to simplify code, reduce complexity, clean up a PR, or improve readability of recent changes.

Capture new project-specific learnings into context files. Runs during reviews or after debugging to evolve the project's skill knowledge base. Use when the thread reveals patterns, anti-patterns, conventions, or codebase gotchas that should be remembered.

Debug pipeline: investigate (debug-prod or debugging) → fix-issues → scoped polish (deslop + code-simplify) → pattern-check → kestral-sync retroactive capture. Only invoked by the kstack subagent.

Debug production issues using cloud logs, CLI tools, error tracking, and database access. Use when debugging production errors, investigating service or job failures, or diagnosing masked errors in production.

Common bugs and debugging strategies. Use when debugging errors, investigating GraphQL/database/TypeScript issues, diagnosing connection pool exhaustion, resolving ESM circular dependency TDZ errors, or when the user asks to debug a local issue.

Capture the decision when a spike, prototype, or investigation concludes. Use when a prototype or spike wraps up, the user makes a go/no-go call after de-risking, or an investigation answers its question — prompts "what did you decide?" and writes the decision + rationale to the task tracker.

Remove AI-generated slop from the branch diff. Strips redundant comments, unnecessary casts, extra defensive code, and style inconsistencies. Use when cleaning up AI-generated code, removing slop, or polishing a branch before review.

Update project docs to match code changes. Use when adding docs, checking what a change requires, or syncing docs/skills/commands with code.

Agent eval development patterns including model config sharing, eval file structure, and running evals. Use when writing or modifying agent evals, creating eval fixtures or scorers, or configuring eval model settings.

Validate reported issues against the codebase and fix confirmed ones with minimal, safe changes. Use when the user provides a list of issues, bugs, or review findings to triage and fix.

Detailed GraphQL resolver patterns, error handling deep dive, cache management, and pagination standards. Use when writing resolvers, implementing server-side error handling, handling client GraphQL errors, managing cache updates, or adding paginated queries.

Create a follow-up implementation task with decision context pre-filled after a spike or prototype concludes with "proceed". Use when a validated approach needs to become trackable implementation work without the engineer re-explaining the context from scratch.

Execute an implementation plan in sequential phases with review checkpoints. Use when the user asks to implement a plan or execute a plan step-by-step.

Implementation pipeline with optional pickup and ship phases around the core build: task-pickup → implement-plan → review-pipeline polish pass → acceptance-check + merge readiness. Only invoked by the kstack subagent.

Sync session progress to your task tracker. Use after push, phase completion, PR creation, or review conclusion. Delegates to the task tracker's sync workflow for full instructions.

Interactive section-by-section PR review with pauses for discussion. Optionally syncs review findings to a linked task at the end, including draft PR creation and linking when no PR exists. Use when the user asks for a manual review, interactive review, walkthrough review, or wants to be guided through a PR change-by-change.

Generate a focused manual test plan from the current branch diff. Use when the user asks for a test plan, QA checklist, or manual testing steps for their changes. Includes API/integration host testing when the diff touches relevant endpoints.

Search existing tasks and incidents for similar bugs or review misses, and link related issues. Use after creating a retroactive bugfix task, when a review finds a familiar-looking issue, or when the user asks "have we seen this before?"

Interactive plan review and requirements validation against the codebase. Use when the user asks for a plan review, wants to validate an implementation plan, or provides a requirements doc, spec, PRD, acceptance criteria, or gap analysis to verify against the codebase.

Planning pipeline: write-plan, then hand off to review-plan-pipeline for the full review pass (plan-review, product-review). Spike mode skips the review pass and ends with decision capture. Only invoked by the kstack subagent.

Product consistency and completeness review. Use when the user asks for a product review, wants to check surface area completeness, multi-user collaboration scenarios, or AI agent interaction points across a feature.

Compress agent instruction/prompt files for minimal token cost while preserving behavioral coverage. Also detects evalmaxxing — prompt changes overfitted to eval fixtures or to one specific reported bug rather than genuine behavior improvements. Use when the user asks to refactor, compress, shorten, or optimize agent prompts, check for evalmaxxing, or review a branch that modifies prompts, eval fixtures, or both.

Fixed seven-step review pipeline: code-review, fix-issues, code-simplify, deslop, context-evolve, documentation-update, task-sync. Orchestrates the full polish pass for a branch. Only invoked by the kstack subagent.

Fixed four-step plan review pipeline: plan-review (requirements validation mode), plan-review (interactive mode), product-review, review summary to Kestral. Orchestrates the full review pass for an implementation plan. Only invoked by the kstack subagent.

Turn a review finding or incident into a durable prevention: a new or updated .cursor/rules/*.mdc rule, skill check, or checklist item. Use after a review surfaces "we should always check X", after a production incident reveals a preventable class of bug, or when the same miss keeps recurring.

Run E2E tests from a plan file using Playwright. Generates test scripts from structured test sections, executes them headless with screenshots, and reports results. Use when the user asks to run E2E tests, verify a phase, automate QA checks, or says /run-e2e.

Create a single self-contained HTML file that mocks a feature using inline React and CSS. Use when the user asks to prototype a UI flow, build a mockup, create a shareable demo, or iterate on a feature's look and feel before implementation.

Claim a task from your task tracker and start working on it. Use when the user says "I'm working on [task]", picks up a task from the board, or wants to start an existing task — handles lookup, conflict check, claim, status, branch naming, and context pull in one flow.

Vitest mocking reference for constructor mocks, global mocks, and env var mocks. Use when writing or fixing test mocks, when a test throws TypeError about constructors, when global mocks leak between tests, or when env var mocking causes undefined values.

Scaffold and implement database migrations using team conventions. Use when the user asks to write a migration, create a migration, or change the database schema.

Author structured implementation plan documents. Use when the user asks to write a plan, create an implementation plan, design a feature spec, or says /plan.