Use the bundled Python 3.11+ installer on Linux or macOS. Native npx skills distributes this skill; it does not execute the installer or apply global prompts.

Orchestration's interactive Manager, native Worker ladder and configured context policy live in the orchestrate-harnesses skill. Window values come from the effective model configuration, not skill or role instructions. Updating that skill changes its distributed defaults, not the current Manager's interactive model/settings or the user's partial model overrides. Remove obsolete roles.manager/roles.preflight/roles.acceptance and controller-only log/prompt overrides as described in that skill before the next dispatch. Setup does not migrate those private settings or rewrite historical tasks. Private rules outside the managed block may still require a Gateway/fixed Manager or Python controller; inspect and reconcile those explicitly when migrating.

After global distribution, explicitly apply the setup:

python3 ~/.agents/skills/workbench-setup/scripts/workbench.py install
~/.local/bin/agent-workbench status

The installer creates ~/.agents/AGENTS.md, links Codex's ~/.codex/AGENTS.md, and ensures ~/.claude/CLAUDE.md has one native import resolving to the shared instructions. An existing equivalent Claude import is reused. Personal text belongs outside  and  in the shared file. Existing Codex regular-file bytes migrate there on first installation; existing Claude text stays in its file.

agent-workbench install reapplies distributed assets. agent-workbench update first updates only installed skills owned by onsager-ai/agent-workbench through the native skills lock, then runs freshly distributed setup code. Node.js 22.20+ and npx are needed for distribution and updates. Do not edit the native skills lock yourself.

When orchestrate-harnesses is installed, install (and update) also generates one Claude Code subagent per routable tier at ~/.claude/agents/<role>.md (luna, haiku, flash, terra, sonnet, sol, opus), so a Claude Manager can dispatch subagent_type: <role>. A tier is routable when its role has cli: claude (using model) or a non-null claude_model in the effective model configuration, which merges the user's partial override over the distributed models.json. Claude tiers are generated by default; Codex tiers need a gateway route, for example {"roles": {"luna": {"claude_model": "<gateway model ID>"}}}, and no gateway or provider prefix is built in. Each file carries the tier's model and effort in frontmatter and the Worker instructions as its prompt. Rerun agent-workbench install after editing the override; a tier whose route is removed loses its generated file. Without orchestrate-harnesses no agents are generated, and status says so.

Generated agent files are recorded with content hashes in the state file. Setup never overwrites a file it does not own, or an owned file you edited, and reports a conflict instead; other files in ~/.claude/agents are left alone. status lists owned agents and uninstall removes unmodified ones.