contextqa-skills
indivatools/contextqa-skills
ContextQA Agent Skills — install via skills.sh to drive the ContextQA MCP from Claude Code, Codex, Cursor, and other agents.
Instalação
npx convez-ai indivatools/contextqa-skills1 comando no terminal: detecta o formato do repositório (marketplace ou plugin avulso) e instala via Claude Code. Requer Node 18+ e o Claude Code instalado.
Métricas
README
A tradução para o português deste README ainda está na fila.
ContextQA Agent Skills
Six agent skills that drive the ContextQA MCP from any compatible coding agent — Claude Code, Codex CLI, Cursor, Antigravity, Claude Desktop, and the rest.
What's in the box
| Skill | Use when |
|---|---|
cqa-init | First-time setup or a connection check — verifies the ContextQA MCP is wired in your agent and reports a ready/not-ready verdict per agent. Run this first. |
cqa-impact | A Jira/Linear/GitHub ticket, a PR, or a branch needs impact analysis — which existing tests are affected, which gaps need new coverage. Layers Claude's own analysis on top of analyze_impact. |
cqa-author | Author tests from a source — ticket, Swagger/OpenAPI, Figma, video, Excel, n8n workflow, code diff, or free-form requirements. |
cqa-debug | A test case is failing or a bug ticket needs to be reproduced and fixed. Investigates telemetry in parallel, hypothesises a root cause, fixes via subagent, verifies, reports. Cap of 3 attempts. |
cqa-regression | Run a test plan as a regression suite, poll to completion, cluster failures, dispatch parallel read-only triage subagents, hand confirmed bugs to cqa-debug. |
cqa-bug-hunter | Find as many bugs as possible in a deployed UI — recon via a ContextQA scout test case, generate dozens of adversarial / negative / edge-case test cases (idempotency, concurrency, validation, auth boundaries…), execute, triage, report. Built for breadth. |
Prerequisites
These skills call ContextQA tools — they don't ship the tools themselves. You need:
- A ContextQA account — sign up at contextqa.com.
- The ContextQA MCP configured in your agent. The hosted server is at
https://mcp.contextqa.com/mcp— point your agent's MCP config at it. Setup details:mcp.contextqa.com/docs. - A first-run sign-in — the first MCP tool call opens a browser for OAuth. Keep the redirect tab open until it completes.
cqa-init handles every step of this in-conversation if anything is misconfigured.
Install
These skills are published in the open Agent Skills format (skills.sh). Install with the skills CLI:
# All six skills, into your default agent's skills directory
npx skills add indivatools/contextqa-skills
# Just one
npx skills add indivatools/contextqa-skills --skill cqa-init
npx skills add indivatools/contextqa-skills --skill cqa-bug-hunter
# Globally for all your agents
npx skills add indivatools/contextqa-skills -g
# To a specific agent
npx skills add indivatools/contextqa-skills -a claude-code
npx skills add indivatools/contextqa-skills -a codex
npx skills add indivatools/contextqa-skills -a cursor
# List before installing
npx skills add indivatools/contextqa-skills --list
Supported agents include Claude Code, Codex, Cursor, Antigravity, Claude Desktop, OpenCode, Goose, Kilo Code, and ~40 others — see the skills CLI README for the full list.
How to invoke
Each skill auto-triggers on natural-language prompts that match its description, or you can name it explicitly:
> /cqa-init
> /cqa-impact PR #142 in our repo
> /cqa-author from this Linear ticket: <paste body>
> /cqa-debug result_id 1406
> /cqa-regression run the smoke plan
> /cqa-bug-hunter https://staging.myapp.com
For /cqa-impact and /cqa-debug, plain pasted text is a first-class input — no Linear/Jira MCP required. The skills opportunistically use Linear MCP, Jira MCP, GitLab MCP, gh CLI, or glab CLI when present, and ask for a paste when absent.
Design principles
- Investigate → plan → confirm gate → execute. Every skill that mutates ContextQA state stops at a confirmation gate before authoring or running anything.
- Parallel subagents for breadth, sequential for depth. Investigation phases dispatch parallel agents (one Agent message, multiple invocations); execution phases use one subagent per unit of work in batches of ≤5.
- Model agency on top of the MCP.
cqa-impactruns its own case-step matcher and caller-graph analysis on top ofanalyze_impact— surfacing what the MCP misses with explicitMODEL-FOUNDlabels. - Lean by default. Each skill is 500–1200 words. Rules sections capture must-not-violate constraints; everything else is the phased flow.
- Tenant feature gating is handled gracefully. When the
/requirements/uploadpipeline isn't enabled (gatescqa-author's swagger / figma / video / excel / requirements / code-diff paths), the skill falls back to manual case authoring instead of failing silently.
Trigger reference
| Skill | Sample triggers |
|---|---|
cqa-init | "set up contextqa", "is the contextqa mcp working", "verify my mcp", "first time with contextqa" |
cqa-impact | "impact analysis", "what tests does this PR affect", "which contextqa tests should I rerun" |
cqa-author | "create tests from this ticket", "generate tests from swagger/figma/video", "build a test suite for X" |
cqa-debug | "debug this failure", "fix this failing test", "why did test result fail", "reproduce and fix this bug" |
cqa-regression | "run the regression", "execute test plan X", "run my smoke suite and triage", "kick off nightly tests" |
cqa-bug-hunter | "find bugs in this app", "hunt for bugs at ", "build adversarial coverage", "stress-test this deployed UI" |
Issues, contributions, ideas
- Issues, feature requests:
indivatools/contextqa-skills/issues - Source MCP server:
indivatools/cqa-mcp - ContextQA platform: contextqa.com
License
MIT. See LICENSE.