This page documents the specialized read-only agents designed for debugging problems and consulting on technical decisions: Oracle, Explore, and Librarian. These agents complement the execution-focused agents (Sisyphus, Hephaestus) by providing investigation, analysis, and consultation capabilities without modifying code.
For planning workflows that use Metis and Momus before execution, see 9.2 Planning Workflow For autonomous deep work execution, see 9.3 Deep Work with Hephaestus
The consultation and debugging system consists of specialized read-only agents designed for specific investigation patterns. These agents are typically invoked via the call_omo_agent tool or delegated through Sisyphus using subagent_type parameters docs/reference/features.md36-43
| Agent | Mode | Cost | Primary Use Case |
|---|---|---|---|
| Oracle | subagent | EXPENSIVE | Strategic consultation, architecture decisions, hard debugging docs/reference/features.md15 |
| Explore | subagent | CHEAP | Codebase search, pattern discovery, file location docs/reference/features.md17 |
| Librarian | subagent | CHEAP | External documentation, open-source reference, library internals docs/reference/features.md16 |
| Multimodal Looker | subagent | CHEAP | Analyzing PDFs, images, and diagrams docs/reference/features.md18 |
The following diagram illustrates the flow from orchestration agents to specialized consultation subagents.
Title: Agent Delegation and Tool Access Flow
Sources: docs/reference/features.md11-18 docs/reference/features.md36-43 docs/guide/orchestration.md66-68
Oracle is a high-reasoning strategic advisor specialized in complex analysis, architecture decisions, and hard debugging problems. It is invoked when the primary agent faces multi-system tradeoffs or has failed a fix attempt multiple times docs/reference/features.md15 Its default model is openai/gpt-5.5 docs/reference/configuration.md91
openai/gpt-5.5 docs/reference/configuration.md91subagent docs/reference/features.md36google/gemini-3.1-pro, anthropic/claude-opus-4-7, opencode-go/glm-5.1 docs/reference/features.md15Oracle follows a pragmatic minimalism philosophy docs/guide/orchestration.md49:
Quick(<1h), Short(1-4h), Medium(1-2d), or Large(3d+).Sources: docs/reference/features.md15 docs/reference/configuration.md91 docs/guide/orchestration.md49
Explore is a contextual search specialist that answers "Where is X?" and "Which file contains Y?" through parallel, multi-angle searches docs/reference/features.md17 Its default model is github-copilot/grok-code-fast-1 docs/examples/default.jsonc31
github-copilot/grok-code-fast-1 docs/examples/default.jsonc31subagent docs/reference/features.md36opencode-go/qwen3.5-plus, opencode-go/minimax-m3, opencode-go/minimax-m2.7, anthropic/claude-haiku-4-5, openai/gpt-5.4-nano docs/reference/features.md17Explore uses a variety of tools to find actionable results, including semantic search (LSP tools), structural patterns (ast_grep_search), text patterns (grep), and file patterns (glob) docs/guide/orchestration.md67
Sources: docs/reference/features.md17 docs/examples/default.jsonc31 docs/guide/orchestration.md67
Librarian specializes in multi-repository analysis, searching remote codebases, and retrieving official documentation using GitHub CLI and Web Search docs/reference/features.md16 Its default model is google/gemini-3-flash docs/examples/default.jsonc28
google/gemini-3-flash docs/examples/default.jsonc28subagent docs/reference/features.md36opencode-go/qwen3.5-plus, opencode-go/minimax-m3, opencode-go/minimax-m2.7, anthropic/vercel/claude-haiku-4-5, openai/vercel/gpt-5.4-nano docs/reference/features.md16Librarian classifies every request into four mandatory categories docs/guide/orchestration.md68:
gh clone).gh issues/prs).Sources: docs/reference/features.md16 docs/examples/default.jsonc28 docs/guide/orchestration.md68
Multimodal Looker is a specialized utility agent for interpreting files that cannot be read as plain text, such as PDFs, diagrams, and images docs/reference/features.md18 Its default model is google/gemini-3.1-pro docs/examples/default.jsonc34
google/gemini-3.1-pro docs/examples/default.jsonc34subagent docs/reference/features.md36opencode-go/kimi-k2.6, zai-coding-plan/glm-4.6v, openai/github-copilot/opencode/gpt-5-nano docs/reference/features.md18look_at ToolThe primary entry point for media analysis is the look_at tool, defined in packages/omo-opencode/src/tools/look-at/tools.ts. It extracts specific information or summaries from documents and describes visual content packages/omo-opencode/src/tools/look-at/tools.ts14-20
path to file_path packages/omo-opencode/src/tools/look-at/tools.test.ts65-71prepareLookAtInput function packages/omo-opencode/src/tools/look-at/look-at-input-preparer.ts91-203 validates inputs, handles file paths and base64 image data, and performs necessary image format conversions (e.g., to JPEG) packages/omo-opencode/src/tools/look-at/look-at-input-preparer.ts121-141 packages/omo-opencode/src/tools/look-at/look-at-input-preparer.ts161-177 It also supports reading JSON files as text parts packages/omo-opencode/src/tools/look-at/look-at-input-preparer.ts111-118Title: Multimodal Analysis Pipeline
Sources: docs/reference/features.md18 docs/examples/default.jsonc34 packages/omo-opencode/src/tools/look-at/tools.ts13-61 packages/omo-opencode/src/tools/look-at/tools.test.ts65-71 packages/omo-opencode/src/tools/look-at/look-at-input-preparer.ts91-203
Consultation agents are frequently used by Planning agents (Metis, Prometheus, and Momus) to gather data before finalizing a work plan.
Title: Planning to Consultation Bridge
Sources: docs/guide/orchestration.md56-57
Refresh this wiki