# Provider Configurations `oh-my-opencode-pi` does not hard-code one provider stack. It uses pi model strings and per-agent overrides, so you can run a single-provider setup or mix providers by specialist. ## Config file location Put provider choices in either: - `~/.pi/agent/oh-my-opencode-pi.json` or `.jsonc` - `.pi/oh-my-opencode-pi.json` or `.jsonc` Project-local config overrides global config. ## How provider selection works in this port Provider/model choice is configured per agent and per council preset. Typical places: - `agents..model` - `agents..variant` - `council.presets..master.model` - `council.presets..councillors[].model` - `fallback.agentChains` - `fallback.councilMaster` Authentication is handled by pi and the configured provider environment, not by this package itself. ## Match model IDs to your authenticated provider This is the easiest way to avoid silent council/delegate failures. Examples: - `openai-codex/gpt-5.5` → ChatGPT Plus/Pro subscription auth via pi `/login` - `openai/gpt-5.5` → OpenAI API key auth via `OPENAI_API_KEY` or `auth.json` - `anthropic/claude-sonnet-4-5` → Anthropic API key auth If the provider prefix in `agents.*.model` or `council.presets.*` does not match the credentials pi actually has, Pantheon subagents can fail before returning any answer. Note: these examples move the main OpenAI model to `gpt-5.5`, but keep mini entries on `gpt-5.4-mini` because `gpt-5.5-mini` is not available yet. --- ## Default-style OpenAI setup A straightforward starting point: ```jsonc { "extends": ["durable"], "agents": { "oracle": { "model": "openai/gpt-5.5", "variant": "high" }, "explorer": { "model": "openai/gpt-5.4-mini", "variant": "low" }, "librarian": { "model": "openai/gpt-5.4-mini", "variant": "low" }, "designer": { "model": "openai/gpt-5.4-mini", "variant": "medium" }, "fixer": { "model": "openai/gpt-5.4-mini", "variant": "low" } }, "council": { "defaultPreset": "review-board", "presets": { "review-board": { "master": { "model": "openai/gpt-5.5", "variant": "high" }, "councillors": [ { "name": "reviewer", "model": "openai/gpt-5.5" }, { "name": "architect", "model": "openai/gpt-5.4-mini", "variant": "medium" }, { "name": "skeptic", "model": "openai/gpt-5.4-mini", "variant": "medium" } ] } } } } ``` --- ## Codex subscription setup If you authenticated with ChatGPT/Codex in pi, prefer `openai-codex/...` model ids: ```jsonc { "extends": ["durable"], "agents": { "oracle": { "model": "openai-codex/gpt-5.5", "variant": "high" }, "explorer": { "model": "openai-codex/gpt-5.4-mini", "variant": "low" }, "librarian": { "model": "openai-codex/gpt-5.4-mini", "variant": "low" }, "designer": { "model": "openai-codex/gpt-5.4-mini", "variant": "medium" }, "fixer": { "model": "openai-codex/gpt-5.4-mini", "variant": "low" } }, "council": { "defaultPreset": "review-board", "presets": { "review-board": { "master": { "model": "openai-codex/gpt-5.5", "variant": "high" }, "councillors": [ { "name": "reviewer", "model": "openai-codex/gpt-5.5" }, { "name": "architect", "model": "openai-codex/gpt-5.4-mini", "variant": "medium" }, { "name": "skeptic", "model": "openai-codex/gpt-5.4-mini", "variant": "medium" } ] } } } } ``` --- ## Mixed-provider setup The pi port works best when you give different specialists different strengths. ```jsonc { "extends": ["research", "durable"], "agents": { "oracle": { "model": "anthropic/claude-sonnet-4-5", "variant": "high" }, "explorer": { "model": "openai/gpt-5.4-mini", "allowedAdapters": ["local-docs", "github-code-search", "web-search"] }, "librarian": { "model": "openai/gpt-5.4-mini", "allowedAdapters": ["local-docs", "docs-context7", "npm-registry", "github-releases", "web-search"] }, "designer": { "model": "google/gemini-2.5-pro" }, "fixer": { "model": "openai/gpt-5.5" } }, "council": { "defaultPreset": "review-board", "presets": { "review-board": { "master": { "model": "anthropic/claude-sonnet-4-5", "variant": "high", "prompt": "Prioritize correctness and simplicity." }, "councillors": [ { "name": "reviewer", "model": "openai/gpt-5.5" }, { "name": "architect", "model": "anthropic/claude-sonnet-4-5" }, { "name": "skeptic", "model": "google/gemini-2.5-pro" } ] } } }, "fallback": { "agentChains": { "fixer": ["openai/gpt-5.4-mini"], "librarian": ["openai/gpt-5.4-mini"] }, "councilMaster": ["openai/gpt-5.5"] } } ``` --- ## Using `variant` When both `model` and `variant` are set, Pantheon combines them into the effective model pattern used for agent runs. Typical use: ```jsonc { "agents": { "oracle": { "model": "anthropic/claude-sonnet-4-5", "variant": "high" } } } ``` --- ## Fast vs durable presets Built-in top-level presets help with overall behavior: - `default` — minimal extra config - `fast` — lower-overhead council/research defaults - `research` — stronger librarian/explorer bias - `durable` — more conservative retry/debug defaults Example: ```jsonc { "extends": ["research", "durable"] } ``` --- ## Fallback chains Use fallback chains when one model is ideal but you want graceful degradation. ```jsonc { "fallback": { "agentChains": { "explorer": ["openai/gpt-5.4-mini"], "librarian": ["openai/gpt-5.4-mini"] }, "councilMaster": ["openai/gpt-5.5"], "finalMessageGraceMs": 1500 } } ``` --- ## Practical recommendations - Give `oracle` your strongest reasoning model. - Give `explorer` and `librarian` something fast and cheap enough for repeated reconnaissance. - Give `designer` a model you trust for UI polish and broader synthesis. - Keep council diverse when you want real second opinions; three copies of the same model rarely add much. - If a provider is flaky, bias toward fallback chains and longer council timeouts. - If a provider prints a final answer but lingers before exiting, tune `fallback.finalMessageGraceMs` instead of raising the full attempt timeout. See also: - [configuration.md](configuration.md) - [council.md](council.md) - [authors-preset.md](authors-preset.md)