gabogg/semillero-special-hotel
1
0
Fork 0
Code Issues Pull requests Projects Releases Packages Wiki Activity Actions
semillero-special-hotel/AGENTS.md
gabogg b2fdf9575d docs: add Hermes Agent guide to AI_DEVELOPER_WORKSTATION.md 
- Add Section 3 documenting Hermes model routing (deepseek-chat / deepseek-v4-pro)
- Add Section 3.2 with 5 project-specific skills (hotel-project-map,
  hotel-remuneration-orchestrator, hotel-seeding-testing, hotel-rls-verification,
  hotel-n8n-deployment)
- Add Section 3.3-3.6 covering AGENTS.md, MCP servers, cron test runner,
  and .hermes/ plans directory
- Add Section 3.7 quick-reference table for all common commands
- Update AGENTS.md to reference Hermes skills instead of inline
  orchestration, add worktree mode guidance
- Bump AGENTS.md changelog for migration date
2026-06-15 01:23:21 +00:00

4.7 KiB
Raw Blame History

Project-Specific Agent Instructions

These instructions extend the baseline global AGENTS.md rules. When executing tasks inside this repository directory, apply the following patterns.

1. Operating Mode & Token Optimization

1.1. Caveman Mode (Lite)

  • Enforce caveman skill on lite level.
  • Keep responses professional, precise, and tight.
  • Eliminate all pleasantries ("sure", "certainly", "of course"), hedging, and filler words.
  • Write full sentences and preserve articles/grammar structure to prevent MCP tool calling confusion, but maintain strict brevity.

1.2. Rust Token Killer (RTK) Command Proxying

  • To conserve input tokens, all shell command executions must be proxied through rtk if it is installed on the host.
  • Examples:
    • Run rtk npm run test instead of npm run test
    • Run rtk git status instead of git status
    • Run rtk prisma migrate dev instead of prisma migrate dev
  • You must only use the normal version of the program only after verifying the rtk version fails.

2. MCP-Specific Guidelines

2.1. n8n MCP Integration (n8n-mcp)

  • Silent Execution: Execute n8n tools without step-by-step commentary. Trigger all searches or node detail fetches silently, and only output user-facing results after all tool steps complete.
  • Parallel Execution: Execute independent node searches or workflow executions in parallel to optimize token roundtrips.

2.2. Forgejo MCP Integration (forgejo-mcp / gitea-mcp)

  • PR Integrity: Always query the pull request details and code reviews using forgejo-mcp tools before running local Git merge commands or updating staging branches.
  • Traceable Commits: Align commit descriptions with issue references queried from the Forgejo instance.

2.3. Prisma & PostgreSQL MCP Integration (@prisma/mcp)

  • Schema Validation: Introspect database structures and validate table states via MCP query tools prior to executing Next.js prisma schema updates.
  • Type-Safety Checks: Run dry-run checks on schemas after any migrations are applied.

3. Hermes Skills (Preferred Method)

This project has dedicated Hermes skills for domain tasks. Load these with skill_view(name) instead of describing the task inline — they contain exact commands, pitfalls, and edge cases. Skills auto-load when you work from this directory via a cron job or new Hermes session.

Available Skills

Skill When to Load
hotel-project-map Start of any session — structural context (schema, routes, pages, deps)
hotel-remuneration-orchestrator Compensation calculations, settlement audits, formula planning
hotel-seeding-testing Database seeding, running tests, debugging failures
hotel-rls-verification Row-level security testing, role isolation, auth policies
hotel-n8n-deployment n8n workflow bootstrap, verification, troubleshooting

Loading Order

For most tasks:

  1. skill_view(name="hotel-project-map") for context
  2. skill_view(name="hotel-<relevant-skill>") for the domain task

For remuneration orchestration (complex multi-agent):

  1. skill_view(name="hotel-project-map")
  2. skill_view(name="hotel-remuneration-orchestrator")
  3. Optionally load hotel-rls-verification or hotel-seeding-testing if needed

4. Git Worktree Mode (Parallel Agents)

When running multiple Hermes sessions editing this project simultaneously, use worktree mode to avoid git conflicts:

hermes -w    # Start in isolated worktree

Each worktree agent edits its own copy. Merge manually or via PR when done.

To check active worktrees:

rtk git worktree list

5. Harness: remuneration-plugin

  • Goal: Orchestrates the multi-agent team (planner, calculator, qa-auditor) for the Variable Remuneration, Compensation, and Commissions system of Hoteles Estelar.
  • Trigger: When requests relate to Variable Remuneration, Compensation, or Commission calculation or verification, utilize the hotel-remuneration-orchestrator skill to delegate tasks to subagents.
  • Changelog:
    Date Change Target Reason
    2026-06-11 Initial scaffolding All files Initial team setup
    2026-06-12 Implement i18n & bilingual AI audit notes n8n, Prisma, E2E tests, agents config Support multilingual UI rendering and LLM translations
    2026-06-12 Synchronize harness skills remuneration-plugin skills & orchestrator Align skills with sequential orchestrator template, i18n bilingual JSON requirements, zero-trust hostnames, webhook branching, and RLS validation checks
    2026-06-14 Migrated to Hermes skills AGENTS.md, skills/ Replaced inline orchestration with hotel-* Hermes skills for portability and token efficiency