Harness Setup

One row per harness devkit routes to: what it's for, how to install, sign in, enable and check it, and what it can't do. Payload adapters (shared skills/hooks projected into Codex, OpenCode, Pi) are documented in Runtime Adapters; this page is about reaching the harness at all. content/model-runtimes.json is the pinned registry behind every row.

codex

Role	payload adapter (~/.codex) and delegate/review runtime, incl. native codex review
Install	devkit sync (brew cask on macOS, npm @openai/codex on Linux; unpinned, installed when missing)
Sign-in	codex login — ChatGPT OAuth or an OpenAI API key
Enable	devkit config codex auto|on|off (default auto: on while installed)
Hook trust	devkit config codex-hook-trust devkit|all|off — devkit sync pre-trusts hooks in ~/.codex/config.toml so Codex never asks; devkit (default) covers devkit's own hooks, all also foreign global hooks and every trusted project's .codex/hooks.json
Check	devkit runtime-route --inventory; devkit doctor
Limits	review_role: true, the only native-review runtime; GPT-6 Sol/Luna

claude

Role	the primary runtime; reverse-direction delegate target via claude -p
Install	install Claude Code from Anthropic directly; devkit does not manage it
Sign-in	Claude Pro/Max OAuth or an Anthropic Console API key
Enable	devkit config runtime claude auto|on|off; always available as primary
Check	devkit runtime-route <tier> reports RUNTIME=claude when resolved
Limits	outside panel and routes.review; same-model review fallback escalates via escalation.claude (fable), never opus

dsh

Role	explore/implement/review/architecture delegate, one-shot headless profile
Install	devkit sync --optional (pinned)
Sign-in	DEEPSEEK_API_KEY (env or $DSH_HOME); readiness stays explicitly unknown
Enable	devkit config runtime dsh auto|on|off (default auto)
Check	devkit runtime-route --inventory
Limits	implement gets a 3600 s bound; cannot commit from a linked worktree

pi

Role	explore/implement/review/architecture delegate; first implement choice for every spec below high risk since 2026-09-24
Install	devkit sync (pinned @earendil-works/pi-coding-agent, plus bubblewrap on Linux and the Claude OAuth adapter; macOS sandboxes implement with the built-in Seatbelt instead)
Sign-in	Claude Pro/Max OAuth via pi-claude-oauth-adapter (accepted account risk — https://code.claude.com/docs/en/legal-and-compliance), or an Anthropic Console API key; other providers report their own readiness
Enable	devkit config pi auto|on|off (default off)
Check	pi auth check; devkit doctor --target pi; devkit runtime-route --inventory <tier> --models --json
Limits	no review_role; extension discovery stays off outside explicit personal paths
DeepSeek	Fallback when the primary Pi model is not signed in or hits its usage limit: DeepSeek V4.1 Flash for explore/review/implement, then Codex gpt-6.1-sol for implement. Enable it with pi, /login, DeepSeek and an API key from platform.deepseek.com. DEVKIT_DELEGATE_PROVIDER/_MODEL/_EFFORT pins one tuple and turns the fallback off. Trial 2026-09-26: every seeded defect found, faster than dsh
Interactive use	shared skills ask through AskUserQuestion, which Pi lacks; pi install npm:@juicesharp/rpiv-ask-user-question adds the matching ask_user_question tool (personal config, never loaded by delegates)