Agent Configuration defines how individual agents are provisioned with models, variants, prompts, tool permissions, and reasoning settings. This configuration is managed primarily through the agents field in the plugin configuration and is processed during the 6-phase configuration pipeline.
Agent configuration is specified in the agents object within .opencode/oh-my-openagent.jsonc (or legacy oh-my-opencode.jsonc). The system uses a Zod-validated schema to ensure type safety.
Sources: docs/reference/configuration.md79-97 docs/examples/default.jsonc7-48 packages/omo-opencode/src/config/schema.test.ts182-219
The applyAgentConfig function orchestrates the assembly of the final agent record. It merges built-in agents, project-specific agents, and user-defined agents while applying overrides and performing key remapping for UI compatibility.
Title: "Agent Configuration Loading and Composition Pipeline"
The loading process follows a strict hierarchy:
$HOME; closer configs win. docs/reference/configuration.md44-46oh-my-openagent.jsonc are applied to the built-in set. docs/reference/configuration.md149-160Sources: docs/reference/configuration.md42-59 docs/reference/features.md3-32
Sisyphus is a highly tuned system that requires specific model families to function correctly as an orchestrator.
Sisyphus is only verified on a narrow set of models. Using unlisted models like MiniMax or Qwen for the orchestrator role is explicitly discouraged as they often fail under the complex 1,100-line orchestration prompt. docs/guide/agent-model-matching.md7-32
| Agent | Primary Model Preference | Fallback Chain |
|---|---|---|
| Sisyphus | claude-opus-4-7 | kimi-k2.6 → kimi-k2.5 → gpt-5.5 → glm-5 |
| Hephaestus | gpt-5.5 | GPT-capable providers only |
| Oracle | gpt-5.5 | gemini-3.1-pro → claude-opus-4-7 → glm-5.1 |
| Prometheus | claude-opus-4-7 | gpt-5.5 → glm-5.1 → gemini-3.1-pro |
Sources: docs/guide/agent-model-matching.md7-88 docs/reference/features.md11-32
task() or call_omo_agent() (Oracle, Librarian, Explore). docs/guide/orchestration.md94-95Permissions enforce security and operational boundaries for different agent roles.
Title: "Tool Permission Enforcement Logic"
Sources: docs/reference/features.md44-54 docs/examples/coding-focused.jsonc15-20
Instead of hardcoding model names in tool calls, the system uses a Categories System. Sisyphus delegates tasks by category, and the configuration maps these categories to specific models.
| Category | Purpose | Default Model |
|---|---|---|
quick | Trivial tasks | gpt-5-nano |
deep | Autonomous implementation | gpt-5.5 |
visual-engineering | UI/UX and diagrams | gemini-3.1-pro |
ultrabrain | Strategic reasoning | gpt-5.5 (xhigh) |
writing | Documentation and prose | kimi-k2.5 |
Sources: docs/reference/configuration.md99-124 docs/examples/default.jsonc50-58 docs/guide/overview.md69
Agents can be provisioned with specific skills (pre-defined prompt sets and tools) via the skills array in the agent configuration.
The system discovers skills from four scopes: built-in, project-local (.opencode/skills), global user config, and MCP-embedded. packages/omo-codex/plugin/test/sync-skills-test-support.mjs7-28
Sources: packages/omo-opencode/src/config/schema.test.ts248-260 packages/omo-codex/plugin/test/sync-skills-test-support.mjs7-36
Refresh this wiki
This wiki was recently refreshed. Please wait 1 day to refresh again.