jarvis-harness

Plugin jarvis-harness from marketplace jarvis-254 in Ignvvcio254/Jarvis-254-Agent. Commands read from the repo's manifest — 26 days ago.

Skills (11)

• agent-uplift

  System that makes cheap/free models deliver frontier-class TASK RESULTS through a harness (compiled specs, generate→verify→repair loop, decomposition, evidence-based routing and per-model anti-patterns). Use when raising the quality of workers (other CLIs, subagents, gateway models), running an upli

• bucle-agentico

  Agentic loop: execute complex features in phases, mapping the real context BEFORE each phase. The key idea: DON'T generate every subtask up front — map context just in time and generate subtasks from the system's actual state. Activate when the task touches multiple coordinated files, needs DB + cod

• field-lessons

  Field-tested engineering laws distilled from real production incidents of an AI agent working 24/7 (verification, silent failures, deploys, migrations, hooks, security, diagnosis). Load BEFORE claiming something is done/verified/deployed, when debugging a "works but doesn't" symptom, writing guards/

• jarvis-evolve

  Permanent self-improvement loop — audits signals (project logs, anti-patterns, memory, session failures, local skills, eval bench), diagnoses gaps, proposes improvements to the user, applies the approved ones and measures. Includes the "frontier-level reasoning through context" doctrine to lift any

• long-run-harness

  Method for LONG work that spans several sessions (a module, a migration, a big feature): an initialiser session creates a feature list with pass flags, a progress file and a bootstrap script; every later session resumes, does ONE feature, verifies it with evidence and commits. Use when the work does

• multi-cli

  Orchestrate external agent CLIs (Antigravity/AGY, OpenCode, Codex…) in visible terminals with a human supervising. File-based task dispatch, compact .done signals, transcript monitoring, collision-free parallel work. Use when the user asks to delegate work to other CLIs or to "dispatch" multi-CLI ta

• prp

  Plan a complex feature before implementing it. Generates a PRP (Product Requirements Proposal) with goal, expected behaviour, data model and implementation phases. Activate ALWAYS before bucle-agentico, when the task involves multiple coordinated files/phases, or when the user says: plan this, give

• scratchpad

  The agent's working memory — its own notes during long tasks + detection of user corrections + propagation to the right memory layer. Fixes fragmentation (knowledge that lives in only one place). Activate on multi-phase tasks, when the user corrects you, or when closing a significant session.

• subagent-orchestration

  Orchestrate subagents (Agent/Task) with a context firewall, the fan-out→synthesize pattern and a rubric grader (Performance Outcomes). Use when a task is massively parallel, long, or benefits from isolated context windows — multi-source research, multi-file refactors, multi-perspective review, bench

• token-router

  Decide the MINIMUM sufficient model for each task/subtask before executing or delegating. Token savings by difficulty — never default to the most powerful model. Trivial→Haiku/free model, standard→Sonnet, complex→Opus. Activate when planning work, delegating to workers (other CLIs, subagents), or on

• verify-repair

  OBJECTIVE verdict gate before reporting any delivery. A task is not done because the model says so, but because an executable signal confirms it (exit code, test, lint, grep, screenshot). Includes a repair loop with a critic separate from the generator. Activate when closing any code task, fix, buil

Slash commands (4)

• /jarvis-harness:consumo

  Reporte de consumo estimado de tokens en la sesión actual. Propone modo Caveman si el usuario lo aprueba.

• /jarvis-harness:doctor

  Diagnostica el estado de instalación portable de Jarvis y detecta faltantes críticos.

• /jarvis-harness:eval-jarvis

  Measure the agent against objective tests of its own rules (prompt injection, destructive actions, secrets, verify-before-report, no hallucination, concision)

• /jarvis-harness:memoria

  Opera el wiki interno de Jarvis — Sistema Global. Subcomandos: ingest (default), query, lint.

Subagents (9)

• architect

  Decisiones de arquitectura y diseño de sistemas — elegir entre enfoques, definir límites entre módulos, evaluar trade-offs de escalabilidad, seguridad y mantenimiento. Úsalo cuando haya una decisión estructural que sería cara de revertir. No escribe código.

• build-error-resolver

  Diagnostica y corrige errores de build, TypeScript, lint o tests que fallan. Úsalo cuando `npm run build`, `tsc`, `eslint` o la suite de tests no pasan. Arregla de a un error y verifica cada vez.

• code-reviewer

  Revisa código recién escrito o modificado (git diff) buscando bugs, calidad y mantenibilidad. Úsalo PROACTIVAMENTE después de escribir o editar código y antes de cualquier commit. Solo lee y reporta; nunca edita.

• doc-updater

  Actualiza documentación (README, docs, comentarios de API, CHANGELOG) para que refleje el código actual tras un cambio. Úsalo después de implementar una feature o cambiar un comportamiento público.

• planner

  Convierte una feature o refactor complejo en un plan de implementación por fases, con riesgos y criterio de verificación por fase. Úsalo PROACTIVAMENTE al recibir una petición grande que toca varios archivos. No escribe código.

• security-reviewer

  Analiza seguridad de código que toca autenticación, autorización, entrada de usuario, SQL, archivos, APIs externas, criptografía o pagos. Úsalo ANTES de commitear ese tipo de cambios. Solo lee y reporta.

• silent-failure-hunter

  Caza fallos silenciosos en código recién escrito o modificado — errores tragados, valores por defecto que ocultan una falla, respuestas "ok" con datos vacíos cuando el origen falló, streams cortados presentados como completos. Úsalo PROACTIVAMENTE después de escribir manejo de errores, llamadas a re

• tdd-guide

  Guía el desarrollo dirigido por pruebas (RED → GREEN → REFACTOR) para features nuevas y corrección de bugs. Úsalo PROACTIVAMENTE al empezar una feature o un fix: primero la prueba que falla, después el código mínimo.

• worker-low

  Ejecutor de tareas acotadas y mecánicas que llegan con spec completo — editar archivos, correr comandos de verificación, crear rutas siguiendo una plantilla existente. Sonnet con effort bajo, para que una tarea barata cueste barato. NO usarlo para diseño, arquitectura, investigación abierta ni decis

Hooks (6)

• SessionStart

  sh "${CLAUDE_PLUGIN_ROOT}/hooks/run-py.sh" "${CLAUDE_PLUGIN_ROOT}/hooks/session_boot.py"

• UserPromptSubmit

  sh "${CLAUDE_PLUGIN_ROOT}/hooks/run-py.sh" "${CLAUDE_PLUGIN_ROOT}/hooks/archetype_router.py"

• UserPromptSubmit

  sh "${CLAUDE_PLUGIN_ROOT}/hooks/run-py.sh" "${CLAUDE_PLUGIN_ROOT}/hooks/correction_detector.py"

• PreToolUse

  sh "${CLAUDE_PLUGIN_ROOT}/hooks/run-py.sh" "${CLAUDE_PLUGIN_ROOT}/hooks/destructive_guard.py"

• PreCompact

  sh "${CLAUDE_PLUGIN_ROOT}/hooks/run-py.sh" "${CLAUDE_PLUGIN_ROOT}/hooks/precompact_checkpoint.py"

• Stop

  sh "${CLAUDE_PLUGIN_ROOT}/hooks/run-py.sh" "${CLAUDE_PLUGIN_ROOT}/hooks/verdict_guard.py"

Similar plugins