The Consensus Tool is a high-stakes decision-making mechanism within the oh-my-openagent ecosystem. It leverages a "multi-lineage" architecture to spawn parallel voters from diverse model families—including Anthropic, OpenAI, Google, and open-source providers—to synthesize a unified position on complex or critical tasks.
In high-stakes scenarios where a single model might hallucinate or exhibit bias, the Consensus Tool (invoked via the ultrabrain or deep categories and parallel subagent execution) allows the system to gather multiple perspectives. By utilizing models with different training backgrounds and reasoning strengths, the system can identify commonalities and outliers in proposed solutions.
background_task system to prevent blocking the main session docs/guide/orchestration.md59-69google/gemini-3.1-pro for visual engineering, openai/gpt-5.5 for deep reasoning) docs/reference/configuration.md112-117sisyphus or atlas) to evaluate the output of all voters and generate a final consensus docs/guide/orchestration.md71-76The consensus flow is powered by the delegate_task tool and the BackgroundManager which coordinates asynchronous execution.
When a consensus-style task is initiated, the system utilizes the subagent_type or category parameters to route work.
task() (the UI alias for delegate_task) with run_in_background: true for each voter docs/reference/features.md63-65Category System determines which model lineage to use. For example, the ultrabrain category maps to openai/gpt-5.5 with variant: xhigh docs/examples/default.jsonc57-58BackgroundManager creates independent sub-sessions. If tmux integration is enabled, each voter may receive its own pane for visual monitoring docs/reference/configuration.md25-26This diagram illustrates how a user request is transformed into a parallel multi-model debate.
"Consensus Flow: Natural Language to Code Entities"
Sources: docs/guide/overview.md54-67 docs/guide/orchestration.md34-76 docs/reference/features.md30-32
The Consensus Tool relies on predefined categories to ensure diverse model selection. Each category is tuned with specific variant and model configurations in the oh-my-openagent.jsonc file.
| Category | Primary Model | Strength | Use Case |
|---|---|---|---|
ultrabrain | openai/gpt-5.5 (xhigh) | Strategic reasoning, complex logic | Architectural debate docs/examples/default.jsonc57 |
visual-engineering | google/gemini-3.1-pro | UI/UX, multimodal analysis | Frontend consistency docs/examples/default.jsonc55 |
unspecified-high | anthropic/claude-opus-4-7 | Instruction following, refactoring | Complex code changes docs/examples/default.jsonc53 |
deep | openai/gpt-5.5 | Autonomous exploration | Root cause analysis docs/examples/default.jsonc56 |
Each voter is initialized via the Dynamic Prompt Building system. When a task is delegated to a category like writing, the model receives a prompt_append specific to that role (e.g., focusing on atomic commits or clear documentation) docs/reference/configuration.md109-123
The following diagram maps the internal components involved in executing and monitoring a consensus task.
"Entity Mapping: Execution Path"
Sources: docs/reference/configuration.md127-138 docs/reference/features.md55-65 docs/guide/agent-model-matching.md101-102
The Consensus Tool should be invoked for high-stakes decisions where "cognitive drift" or model-specific bias must be mitigated:
@oracle (GPT-5.5) vs a visual-engineering (Gemini) perspective docs/guide/overview.md118-121Metis (Gap Analyzer) and Momus (Reviewer) to critique a plan generated by Prometheus docs/guide/orchestration.md155-162anthropic/claude-opus (precision) and openai/gpt-5.5 (deep reasoning) docs/guide/agent-model-matching.md77-87explore agents in parallel to search for patterns using different speed-optimized models docs/guide/overview.md61-62To prevent rate-limiting during a large consensus debate, the system enforces a providerConcurrency limit (typically 3-5 for expensive providers like Anthropic/OpenAI and up to 10-15 for cheaper ones) docs/examples/coding-focused.jsonc74-83
Sources: docs/reference/configuration.md127-138 docs/guide/orchestration.md34-76 docs/guide/agent-model-matching.md35-57
Refresh this wiki