Skip to content

Latest commit

 

History

History
186 lines (150 loc) · 18.7 KB

File metadata and controls

186 lines (150 loc) · 18.7 KB

Provider Runtime Layouts

This document records the installed runtime layout for the provider lines used by Orchestrarium today, including provider source trees that already exist in the monorepo. It is an install/runtime reference, not a source-layout reference.

Production auto-routing in the root integration contract is limited to Codex plus Claude Code. Gemini CLI and Qwen are documented here as explicit example integrations only; both are classified in this repository as WEAK MODEL / NOT RECOMMENDED.

Do not confuse these runtime surfaces with the monorepo authoring trees such as src.codex/, src.claude/, src.gemini/, or src.qwen/.

Architecture note: on the Codex line, the installed AGENTS.md is intentionally the compact universal minimum. Detailed installed role contracts and runtime guidance belong in the installed skills/<role>/SKILL.md files and the built-in .codex/agents/*.toml overrides; shared/provider reference trees are source-maintainer canon, not target-project install payload. Claude already follows the analogous pattern through a short CLAUDE.md entrypoint plus .claude/agents/*.md role files.

Read the tables with three layers in mind:

  • Official provider behavior means the provider's own documented runtime surface or configuration model.
  • Orchestrarium runtime contract means the install shape and conventions introduced by this repository.
  • Observed installed behavior means the result verified in an installed target.

Do not collapse those layers into one claim. When a row is Orchestrarium-owned rather than provider-native, the notes call that out explicitly.

Scope legend

Scope Meaning
Global User-level installed runtime surface, usually under the provider's home directory
Local Project-level installed runtime surface inside the current repository or target project

Codex

Global

Item Path or shape Notes
Installed pack root ~/.codex/ Global Codex pack install target
Governance entrypoint ~/.codex/AGENTS.md Installed Codex runtime entrypoint; intentionally the compact universal minimum rather than the full role/runtime manual
Skill tree ~/.codex/skills/<role>/SKILL.md Orchestrarium Codex runtime organizes each role as a skill directory
Built-in subagent overrides ~/.codex/agents/default.toml, worker.toml, explorer.toml Orchestrarium-installed custom agents that override Codex built-ins and pin them to gpt-5.5 + xhigh; stale Orchestrarium-owned templates refresh on reinstall, while real user-customized prompt or structure is preserved
Validation script ~/.codex/skills/lead/scripts/validate-skill-pack.sh Same lead script tree as the repo source
Publication-safety scan ~/.codex/skills/lead/scripts/check-publication-safety.sh PowerShell wrapper exists alongside the shell script
Global operator overlay ~/.codex/.agents-mode.yaml Orchestrarium-owned default operator file seeded on first global install and preserved on reinstall; legacy sibling ~/.codex/.agents-mode is compatibility input only

Local

Item Path or shape Notes
Installed pack root <project>/.agents/skills/ Role skills are copied here
Governance entrypoint <project>/AGENTS.md Codex pack section is merged into the project-root AGENTS.md; the installed Codex section stays intentionally compact and defers detailed installed role/runtime guidance to the skill tree and .codex/agents/ overrides
Skill tree <project>/.agents/skills/<role>/SKILL.md Mirrors the global skills/ structure
Built-in subagent overrides <project>/.codex/agents/default.toml, worker.toml, explorer.toml Project-level custom agents that override Codex built-ins and pin them to gpt-5.5 + xhigh; seeded on first install, refreshed on reinstall when still an Orchestrarium-owned template, and preserved when genuinely customized
Local config <project>/.agents/.agents-mode.yaml Canonical Orchestrarium local state file; local install seeds the default and $init-project reviews or updates it, while legacy sibling <project>/.agents/.agents-mode remains compatibility input only. Decision-driving reads use this local scope first, then fall back to the global Codex overlay when the local scope is absent.
Validation script <project>/.agents/skills/lead/scripts/validate-skill-pack.sh Run from the target project root after install
Publication-safety scan <project>/.agents/skills/lead/scripts/check-publication-safety.sh PowerShell wrapper exists alongside the shell script

Claude Code

Global

Item Path or shape Notes
Global context file ~/.claude/CLAUDE.md Official user-level Claude Code instruction file; intentionally short while detailed role behavior lives under ~/.claude/agents/
Global personal skills ~/.claude/skills/<skill-name>/SKILL.md Official preferred user-level extension surface
Global personal subagents ~/.claude/agents/*.md Official user-level custom subagent surface
Global legacy commands ~/.claude/commands/*.md Still supported, but Claude docs now recommend skills as the preferred model
Global operator overlay ~/.claude/.agents-mode.yaml Orchestrarium-owned default operator file seeded on first global install and preserved on reinstall; not a Claude-native file from official docs, and legacy sibling ~/.claude/.agents-mode is compatibility input only

Local

Item Path or shape Notes
Project context file <project>/.claude/CLAUDE.md or <project>/CLAUDE.md Official project-level Claude instruction entrypoints; keep the entrypoint short and the detailed role files under .claude/agents/
Local personal override <project>/CLAUDE.local.md Official personal, uncommitted project override layer
Project skills <project>/.claude/skills/<skill-name>/SKILL.md Official preferred project-level extension surface
Project subagents <project>/.claude/agents/*.md Official project-level custom subagent surface
Legacy commands <project>/.claude/commands/*.md Still work, but lose precedence to a skill with the same name
Orchestrarium shared governance copy <project>/.claude/AGENTS.md Repo-local overlay copied by Orchestrarium install scripts; not a Claude-native runtime requirement
Orchestrarium local config <project>/.claude/.agents-mode.yaml Canonical Orchestrarium local state file; local install seeds the default and /agents-init-project reviews or updates it, while legacy sibling <project>/.claude/.agents-mode remains compatibility input only. Decision-driving reads use this local scope first, then fall back to the global Claude overlay when the local scope is absent.
Pack memory <project>/.claude/memory/ Repo-local installed memory payload for the Orchestrarium Claude pack

Gemini CLI (Example Integration)

Global

Item Path or shape Notes
Global context file ~/.gemini/GEMINI.md Official user-level Gemini context file
Global shared-governance import ~/.gemini/AGENTS.md Orchestrarium-installed markdown module imported by GEMINI.md; not a Gemini-native required filename
Global user skills ~/.gemini/skills/ Official user-level skill location; Orchestrarium leaves this tier for user overrides instead of mirroring the pack here
Global user skills alias ~/.agents/skills/ Official alias; within the user tier, the alias takes precedence over ~/.gemini/skills/
Global user subagents ~/.gemini/agents/ Official preview user-level subagent location; Orchestrarium leaves this tier for deliberate user overrides
Global custom commands ~/.gemini/commands/ Official user-level Gemini custom commands; Orchestrarium does not mirror its own commands here
Global settings ~/.gemini/settings.json Official CLI configuration, including optional context.fileName overrides
Global operator overlay ~/.gemini/.agents-mode.yaml Orchestrarium-owned shared-routing overlay seeded on first global install and preserved on reinstall; not a Gemini-native settings surface, and legacy sibling ~/.gemini/.agents-mode is compatibility input only
Global extensions ~/.gemini/extensions/<extension>/ Official runtime location for installed or linked extensions; Orchestrarium materializes orchestrarium-gemini here on global Gemini install
Extension manifest gemini-extension.json inside an extension Official extension manifest; extensions can bundle skills, commands, context, and MCP servers

Local

Item Path or shape Notes
Project context file <project>/GEMINI.md Official default project-level Gemini context file; built-in /init generates or tailors this file
Project shared-governance import <project>/AGENTS.md Orchestrarium-installed markdown module imported by GEMINI.md; not a Gemini-native required filename
Parent-context hierarchy <project>/../GEMINI.md up to project root Gemini walks parent directories until the .git root
Sub-directory context <project>/<subdir>/GEMINI.md Gemini also loads more specific context files below the current working directory
Workspace skills <project>/.gemini/skills/ Official workspace skill location; Orchestrarium leaves this tier for explicit repo-local overrides instead of mirroring the pack here
Workspace skills alias <project>/.agents/skills/ Official alias; within the workspace tier, the alias takes precedence over .gemini/skills/
Workspace subagents <project>/.gemini/agents/ Official preview project-level subagent location; Orchestrarium leaves this tier for deliberate repo-local overrides
Workspace custom commands <project>/.gemini/commands/ Official project-local Gemini custom commands; Orchestrarium does not mirror its own commands here
Workspace extensions <project>/.gemini/extensions/<extension>/ Official workspace extension location; Orchestrarium materializes orchestrarium-gemini here on project-local Gemini install
Workspace settings <project>/.gemini/settings.json Official project-local Gemini settings
Orchestrarium operator overlay <project>/.gemini/.agents-mode.yaml Repo-local shared routing overlay for consultant, delegation, MCP, external-provider preferences, named priority profiles, and opinion counts; local install seeds the default, Gemini /init still owns GEMINI.md, the Orchestrarium Gemini init helper reviews or updates the overlay when project-specific choices are needed, and legacy sibling <project>/.gemini/.agents-mode remains compatibility input only. Decision-driving reads use this local scope first, then fall back to the global Gemini overlay when the local scope is absent.
Optional context filename override context.fileName in settings AGENTS.md is not a default Gemini entrypoint; Orchestrarium uses GEMINI.md imports instead of taking over this settings-owned surface
Extension-provided skills installed extension content Official third discovery tier after workspace and user skills
Important overlap note workspace .agents/skills/ If a repository already uses .agents/skills/ for Codex, Gemini will also discover those skills because this alias is official Gemini behavior
Conflict-avoidance note user/workspace tiers vs extension Gemini resolves precedence as workspace > user > extension, so Orchestrarium's Gemini installer keeps the pack in the extension tier and cleans legacy mirrored duplicates from top-level user/workspace tiers on reinstall

Qwen (Example Integration)

Qwen is maintained in this monorepo as a native explicit example integration classified as WEAK MODEL / NOT RECOMMENDED, not as a production auto-routing line. The tables below record the Orchestrarium-owned Qwen example runtime contract verified from this checkout and installer dry-runs; they do not promote Qwen into production auto.

Global

Item Path or shape Notes
Global context file ~/.qwen/QWEN.md Orchestrarium Qwen example entrypoint installed by the Qwen installer
Global shared-governance import ~/.qwen/AGENTS.md Orchestrarium-installed markdown module imported by QWEN.md; not a production routing surface
Global user skills ~/.qwen/skills/ Left for deliberate user overrides instead of mirroring the pack here
Global user subagents ~/.qwen/agents/ Left for deliberate user overrides instead of mirroring the pack here
Global custom commands ~/.qwen/commands/ Left for deliberate user overrides instead of mirroring the pack here
Global settings ~/.qwen/settings.json Qwen-native runtime configuration surface; Orchestrarium does not own this file
Global operator overlay ~/.qwen/.agents-mode.yaml Orchestrarium-owned shared-routing overlay seeded on first global install and preserved on reinstall; legacy sibling ~/.qwen/.agents-mode is compatibility input only
Global extensions ~/.qwen/extensions/<extension>/ Runtime location where Orchestrarium materializes orchestrarium-qwen on global Qwen install
Extension manifest qwen-extension.json inside an extension Orchestrarium Qwen extension manifest source; carries context file, skills, agents, and commands fields for the example payload

Local

Item Path or shape Notes
Project context file <project>/QWEN.md Qwen-line project entrypoint managed by the Orchestrarium installer
Project shared-governance import <project>/AGENTS.md Orchestrarium-installed markdown module imported by QWEN.md; not a production routing surface
Workspace skills <project>/.qwen/skills/ Left for explicit repo-local overrides instead of mirroring the pack here
Workspace subagents <project>/.qwen/agents/ Left for deliberate repo-local overrides instead of mirroring the pack here
Workspace custom commands <project>/.qwen/commands/ Left for deliberate repo-local overrides instead of mirroring the pack here
Workspace extensions <project>/.qwen/extensions/<extension>/ Runtime location where Orchestrarium materializes orchestrarium-qwen on project-local Qwen install
Workspace settings <project>/.qwen/settings.json Qwen-native project runtime configuration surface; Orchestrarium does not own this file
Orchestrarium operator overlay <project>/.qwen/.agents-mode.yaml Repo-local shared routing overlay for consultant, delegation, MCP, external-provider preferences, named priority profiles, and opinion counts; local install seeds the default and Qwen init helpers review or update it when project-specific choices are needed. Legacy sibling <project>/.qwen/.agents-mode remains compatibility input only. Decision-driving reads use this local scope first, then fall back to the global Qwen overlay when the local scope is absent.
Extension-provided skills installed extension content Orchestrarium example payload installed under the extension root
Extension-provided agents installed extension content Orchestrarium example specialist-agent payload installed under the extension root
Extension-provided commands installed extension content Orchestrarium example custom-command payload installed under the extension root
Conflict-avoidance note user/workspace tiers vs extension The Qwen installer keeps the pack in the extension tier and leaves top-level user/workspace tiers for explicit overrides, mirroring the Gemini example-line isolation model

Quick comparison

Provider Global runtime root Local runtime root Native instruction entrypoint
Codex ~/.codex/ <project>/.agents/ plus root AGENTS.md AGENTS.md
Claude Code ~/.claude/ <project>/.claude/ and optional root CLAUDE.md CLAUDE.md
Gemini CLI ~/.gemini/ <project>/.gemini/ plus GEMINI.md hierarchy GEMINI.md
Qwen ~/.qwen/ <project>/.qwen/ plus root QWEN.md QWEN.md

Sources

Terms and Abbreviations

  • AGENTS.md: agent governance file used directly by Codex and installed as a shared-governance module for example providers.
  • agents-mode: Orchestrarium routing and operator overlay file for provider preferences and execution policy.
  • CLI: Command-Line Interface, a terminal command surface for a provider runtime.
  • Codex: OpenAI Codex runtime and production provider line.
  • Claude Code: Anthropic's Claude runtime and production provider line.
  • extension: provider-supported package directory that can bundle context, skills, agents, commands, or manifests.
  • Gemini CLI: Google's Gemini command-line runtime, documented here as an explicit example integration.
  • global: user-level install scope such as ~/.codex/, ~/.claude/, ~/.gemini/, or ~/.qwen/.
  • local: project-level install scope under a target repository.
  • MCP: Model Context Protocol; a mechanism for exposing tool and resource servers to agent runtimes.
  • Qwen: Qwen provider line, documented here as an explicit example integration.
  • runtime root: provider-facing directory where installed context, skills, agents, commands, or overlays live.
  • WEAK MODEL / NOT RECOMMENDED: repository classification for example integrations that do not belong in production auto routing.