Tool Configs
Each harness keeps a basic config set directly in its own files: model, effort, permissions, and MCP servers. MCP servers are single-sourced in MCP servers; models are set per harness, with no shared registry.
| Page | Owns |
|---|---|
| Config merging | harness config sources and run_onchange merge scripts |
| Claude Code | Claude Code settings and subagent profile |
| Pi coding agent | Pi packages, settings, models, extensions, APPEND_SYSTEM.md |
| Other harnesses | Codex, OpenCode, OMP, Crush, tuicr, lgtm, secrets |
Launcher defaults
No command-line, environment, or resumed-session overrides.
— means no explicit context override; numbers are context tokens.
| Launcher | What it does | Default model | Effort | Context |
|---|---|---|---|---|
claude | Claude Code | claude-opus-5-5 | medium | long |
,claude-openrouter | Claude Code through OpenRouter's Messages endpoint | xiaomi/mimo-v2.6-pro | high (preset) | long |
,claude-llama-cpp | Claude Code against the local llama.cpp server | nemotron-3.5 | client high; server auto | 262,144 |
codex / ,codex | Codex; interactive codex routes through ,codex, which adds local-model metadata | gpt-6.1-sol | ultra | native default |
,codex-openrouter | Codex through OpenRouter's Responses endpoint | xiaomi/mimo-v2.6-pro | high (preset) | long |
,codex-llama-cpp | Codex against the local llama.cpp server | nemotron-3.5 | client high; server auto | 262,144 |
pi | Pi | openai-codex/gpt-6.1-sol | xhigh | native default |
omp | Oh My Pi | anthropic/claude-opus-5-5 | medium | native default |
opencode | OpenCode | openrouter/z-ai/glm-5.3-flash@preset/glm-lanes-high | high | long |
,opencode-llama-cpp | OpenCode with the local llama.cpp provider | llama-cpp/nemotron-3.5 | server auto | 262,144 |
crush | Crush | openrouter/z-ai/glm-5.3-flash | high | long |
freebuff | Freebuff | not exposed by CLI | — | — |
,q | one-shot Pi with tools and a short system prompt | openrouter/deepseek/deepseek-v4.1-flash-1017 | high | long |
command codexbypasses the fish function that routescodexto,codex.- OpenRouter wrappers append
@preset/effort-<level>;--context shortkeeps the cheaper context tier. - Local wrappers do not pin backend effort;
server autois llama.cpp'sreasoning=auto. - Crush's runtime data JSON takes precedence over
crushrcat runtime.
Sources: home/exact_bin/ wrappers, ,q, Codex, Claude, OpenCode, Pi, OMP, llama.cpp router.