Skip to content

ask.md

Latest commit

 

History

History
194 lines (127 loc) · 5.64 KB

ask.md

File metadata and controls

194 lines (127 loc) · 5.64 KB

duh ask

Run a consensus query across multiple LLMs.

Synopsis

duh ask [OPTIONS] QUESTION

Description

The ask command is duh's primary command. By default it sends your question through the full consensus protocol:

  1. PROPOSE -- Strongest model answers the question
  2. CHALLENGE -- Other models challenge the proposal with forced disagreement
  3. REVISE -- Proposer revises based on genuine challenges
  4. COMMIT -- Decision extracted with confidence score and dissent

The process repeats for up to max_rounds (default: 3) or until convergence is detected.

You can also select alternative protocols:

  • Voting (--protocol voting) -- Fan out to all models in parallel, then aggregate the best answer
  • Auto (--protocol auto) -- Classify the question first (reasoning vs. judgment), then route to consensus or voting
  • Decomposition (--decompose) -- Break the question into subtasks, solve each with consensus, and synthesize

Output is displayed in real-time with Rich-styled panels:

  • Green panel: Proposal
  • Yellow panel: Challenges (with sycophancy flags)
  • Blue panel: Revision
  • Cyan panel: Votes, decomposition plan, subtask results, tool usage
  • White panel: Final decision with confidence and cost

Arguments

Argument Required Description
QUESTION Yes The question to ask. Wrap in quotes if it contains spaces.

Options

Option Type Default Description
--rounds int From config (3) Maximum consensus rounds. Overrides general.max_rounds in config.
--decompose flag false Decompose the question into subtasks before consensus.
--protocol choice From config (consensus) Protocol: consensus (default), voting, or auto (classify first).
--tools / --no-tools flag From config Enable or disable tool use (web search, code exec, file read). Overrides tools.enabled in config.
--proposer string Auto-selected Override the proposer model (e.g. anthropic:claude-opus-4-6).
--challengers string Auto-selected Override challengers (comma-separated model refs, e.g. openai:gpt-5.2,google:gemini-3-pro-preview).
--panel string All models Restrict consensus to these models only (comma-separated model refs). Overrides consensus.panel in config.

Examples

Basic question:

duh ask "What are the security implications of using JWTs for session management?"

Single-round consensus (faster, cheaper):

duh ask --rounds 1 "Quick: should I use Redis or Memcached for caching?"

Deeper debate (more rounds):

duh ask --rounds 5 "Design a microservices architecture for an e-commerce platform"

Use the voting protocol:

duh ask --protocol voting "Which Python web framework should I use?"

Auto-detect whether to use consensus or voting:

duh ask --protocol auto "Is Rust worth learning in 2026?"

Decompose a complex question into subtasks:

duh ask --decompose "Design a complete CI/CD pipeline for a monorepo"

Enable tool use (web search, code execution, file read):

duh ask --tools "What is the current LTS version of Node.js?"

Disable tool use even if enabled in config:

duh ask --no-tools "Explain the CAP theorem"

Use a specific proposer:

duh ask --proposer openai:gpt-5.2 "Compare REST and GraphQL"

Override challengers:

duh ask --challengers google:gemini-3-pro-preview,anthropic:claude-sonnet-4-6 "Best database for IoT?"

Restrict to a panel of models:

duh ask --panel anthropic:claude-opus-4-6,openai:gpt-5.2 "Design a caching strategy"

With a specific config:

duh --config ./local-config.toml ask "question"

Output sections

Round header

────────────── Round 1/3 ──────────────

Proposal (green panel)

Shows the proposer's model reference and initial answer.

Challenges (yellow panel)

Shows each challenger's model reference and challenge text. Sycophantic challenges are flagged with a yellow "sycophantic" label.

Revision (blue panel)

Shows the reviser's model reference and improved answer.

Commit line

✓ COMMIT  Confidence: 100%  (no dissent)

Round footer

Round 1/3 | 6 models | $0.0342 | 4.2s

Final decision (white panel)

The full, untruncated consensus decision with confidence and cost.

Dissent panel (yellow, if present)

Preserved genuine challenges that represent minority viewpoints.

Voting output

When using --protocol voting, the output shows:

  • Votes panel (cyan) -- Each model's independent answer
  • Decision panel (white) -- The aggregated best answer
  • Stats line -- Strategy (majority/weighted), confidence, vote count, cost

Decomposition output

When using --decompose, the output shows:

  • DECOMPOSE panel (magenta) -- The subtask DAG with labels, descriptions, and dependencies
  • Subtask results -- Each subtask's consensus decision and confidence
  • SYNTHESIS panel (white) -- The merged final answer from all subtask results
  • Stats line -- Aggregate confidence and cost

Tool usage output

When tools are used during consensus (enabled via --tools or config), a TOOLS panel (cyan) appears after the decision showing each tool call with the phase, tool name, and arguments.

Related commands

  • recall -- Search for this decision later
  • show -- Inspect the full thread details
  • feedback -- Record whether the decision worked out
  • cost -- Check cumulative spending