hermes-agent/plugin-catalog/model-router.yaml

name: model-router
title: "Model Router"
repo: https://github.com/metamindedu/hermes-model-router
sha: e207667489ac8a796d53f875126d6ee227201112
description: "Explicit per-profile readiness and native missing-profile preparation; Python installation remains per profile, never automatic on refresh. Profile model settings, default-on Heavy/Light task routing, fallback controls, runtime limits and Fast inheritance. Reads Hermes profile configuration; explicit saves update selected profiles via the native CLI. Routed jobs launch background Hermes children using the user's configured providers/MCP and store private prompts/results locally. No separate credential store, telemetry, self-updater or approval bypass. Disclosure — after you trust and enable the plugin, routing defaults ON for every installed profile without a saved choice, including profiles that installed an earlier version and never saved one (the update turns it on); existing saved OFF and explicit skill opt-outs remain OFF. On enabled profiles, the model-callable `coding_route` tool starts detached background Hermes agents. They get file, terminal and code-execution tools in any existing folder the model names, run under that profile's credentials and approval policy (single-query mode: dangerous commands are refused unless `approvals.single_query_mode: approve`), and their per-job time limit can be set to none. Prompts and results are stored in `plugin-data/model-router/jobs/`. Saving settings writes model, delegation, auxiliary, fallback and `skills.disabled` keys into the config.yaml of the profiles listed in the preview (backups are taken first) and copies a skill into those profiles' `skills/` folders. The Desktop \"prepare profiles\" action, after you select profiles and confirm, installs and enables this same catalog pin in up to 32 other local profiles through Hermes's plugin manager, where routing then defaults ON. From install, native subagents inherit the parent session's Fast (priority/ultrafast) tier on first-party OpenAI/Codex routes, which can raise costs."
maintainer: metamindedu
tier: community
category: models
version: "0.4.8"
docs_url: https://github.com/metamindedu/hermes-model-router#readme
readme: true
platforms:
  - windows
screenshots:
  - https://raw.githubusercontent.com/metamindedu/hermes-model-router/e207667489ac8a796d53f875126d6ee227201112/docs/fast-inheritance.png
known_issues:
  - "v0.4.8 adds explicit local missing-profile preparation (maximum 32). Selection and known PyYAML dependency/activation consent are required, and the installed source record must match the currently resolved catalog pin. Existing, disabled, partial or custom code is manual-only; no global/future-profile auto-install or self-updater."
  - "Native writes require an exact fresh target descriptor on the same local source; missing, duplicate, remote or aliased routes fail manual-only. Catalog lookup and profile-name binding are not atomic: do not rename/delete/recreate profiles or edit packages while preparing. Readiness observes files and effective Python selection, not existing-session registration or provider success."
  - "Cancel/scope change stops unissued work, not an already-issued native install. Unknown outcomes are never automatically retried; both settled and unknown writes retain a settings-rebase fence until explicit owner-scoped full reload. Readiness-only refresh preserves dirty drafts; full reload discards them. Existing saved routing OFF remains OFF."
  - "v0.4.7 also requests native completion wakes for validated default-local turns without a UI-session ID; configured-model quota can be used by continuation. Stop, busy, ownership and explicit notification opt-out remain native. Consume the owned terminal process result before reading the routed verdict; pre-consumption raw replay is not an exactly-once guarantee."
  - "Cold PM binding selects only a verified already-imported generation; multi-argument sealed native launch layouts still fail closed. Windows noninteractive launches suppress console windows; skipped OS/permission tests do not establish other-platform coverage."
  - "Routing is default ON after package trust/activation only for installed profiles with no saved choice. Existing OFF/skill opt-outs remain OFF; install does not rewrite models/fallback consent/budgets or launch jobs. The advanced UI retains opt-out."
  - "Heavy/Light may use the same model in independent purpose-bound jobs. Same-model context separation does not provide different-model review diversity."
  - "Native YAML and served-runtime bootstrap compatibility are verified on Windows. Existing private core imports remain a host-version compatibility risk and are reported by plugin validation."
  - "Small bounded changes stay with Main. Ordinary implementation delegation requires two substantive independent workstreams with net parallel benefit; explicit delegation and required Heavy/independent review remain exceptions. This is guidance, not an API two-job minimum."
  - "Preferences now live under Settings > Plugins. The main-sidebar settings row is removed; legacy /model-router bookmarks remain supported. A host supporting the public plugin-settings area is needed; no private-core or sidebar fallback is added."
  - "The Desktop UI is Korean-first. Windows runtime/SDK/Desktop behavior has been tested; full Linux/macOS runtime support is not claimed."
  - "Requires a Hermes build with unified Python/Desktop packages and request/tool middleware; new Python tools may require a new session."
  - "Fast and fallback can increase API charges or subscription-credit use. Transports bypassing the request middleware, including codex_app_server, are outside Fast inheritance coverage."
  - "The native scanner can report CAUTION for the manual source-repair helper; routing-disabled capability probes can warn that coding_route was not registered."
  - "Default-profile mid-run activation was not confirmed on the developer's shared backend; package installation is distinct from existing-session activation."
capabilities:
  provides_tools:
    - coding_route
  provides_hooks:
    - pre_api_request
    - post_api_request
    - api_request_error
    - pre_tool_call
    - post_tool_call
    - subagent_start
    - subagent_stop
  provides_middleware:
    - tool_execution
    - llm_request
  requires_env: []
requires_hermes: ">=0.21.6"