dissenter 1.0.4

Multi-LLM debate engine for complex questions — surface disagreement, synthesize decisions

dissenter

Run multiple LLMs through a structured debate for complex questions. Surface where they disagree. Synthesize a decision.

make ask Q="Should I use Kafka or a Postgres outbox pattern for event-driven microservices?"

---

Table of Contents

• Why this exists
• What the existing tools get wrong
• What dissenter does differently
• Architecture
• Installation
• Running
• Configuration
  • Minimal config
  • Multi-round
  • Dual-arbiter final
  • CLI auth — no API keys
  • Same model, multiple roles
  • Random role distribution
  • Per-model API key
• Roles
• Output
• Testing
• Comparison
• Academic foundations
• Roadmap

---

Why this exists

There are already tools that aggregate multiple LLMs for consensus answers. This is not that.

Every existing tool — llm-council, llm-consortium, consilium, the research implementations of Mixture of Agents — is trying to build a better oracle. They treat disagreement as noise to eliminate and convergence as success.

For architectural decisions, that's exactly backwards.

When multiple expert models disagree, that disagreement tells you where the decision is genuinely hard and context-dependent. That's not noise — it's the most useful information you can get. A tool that eliminates it to produce confident-sounding consensus is actively hiding the difficulty of your decision.

dissenter treats disagreement as the signal, not the problem.

---

What the existing tools get wrong

They use identical prompts for all models

Sending the same neutral question to five models gets you five statistically similar answers with slight variation. You're not extracting diverse perspectives — you're sampling noise from similar training distributions. The February 2025 LLM ensemble survey (arXiv 2502.18036) found this is the primary reason naive ensembles underperform.

They chase consensus

The goal of arbiter/judge patterns in llm-council, consilium, and llm-consortium is to produce a single authoritative answer. But for architectural decisions — which involve trade-offs specific to your team, stack, and constraints — false consensus is worse than acknowledged uncertainty. The models don't know your system. The arbiter doesn't know your team.

They're stateless

No tool persists your decisions. You can't ask "given we chose Kafka three months ago, how does that change this?" Every query is context-free. Architectural decisions form a causal chain; these tools treat each one as an isolated question.

They depend on OpenRouter or require specific infrastructure

llm-consortium is a plugin for Simon Willison's llm tool. consilium requires a Rust binary. MoA reference implementations need TogetherAI. None mix cloud + local models cleanly without a proxy service.

They require API keys for every model

Every tool assumes you're accessing models via API key. If you have a claude CLI or gemini CLI installed and authenticated, that credential is invisible to them — you still need a separate API key. This means you're paying twice for access you already have, and there's no path to using browser-authenticated sessions.

---

What dissenter does differently

1. Multi-round debate with context passing

Models run in parallel within each round. Each subsequent round receives all prior rounds as context. A typical pipeline:

• Round 1 (debate): Any number of models argue from adversarial roles in parallel
• Round 2 (refine): A smaller panel reviews the debate and sharpens the analysis
• Final round: 1 chairman synthesizes into a decisive ADR, or 2 arbiters (conservative + liberal) produce side-by-side recommendations

Round depth is arbitrary. Configure as many rounds as the decision warrants.

2. Role-differentiated prompting

Rather than asking all models the same neutral question, each model is assigned an adversarial role with a distinct mandate. The research backing: the "Rethinking MoA" paper (OpenReview 2025) found that diversity of framing produces better results than diversity of model. You get more useful signal from one model asked with five different stances than five models asked the same way.

3. Roles as external files

Role prompts are not hardcoded. They live in src/dissent/roles/*.toml — plain text files you can read and edit. Add a new file, get a new role. No code changes required.

4. Dual-arbiter output

The final round can use 2 models instead of 1. A conservative arbiter recommends the safest proven path; a liberal arbiter recommends the boldest high-upside path. A combine_model merges them side-by-side into a single document. Useful when the right answer genuinely depends on your team's risk tolerance.

5. Disagreement is the output, not the problem

The synthesized ADR has a dedicated Disagreements section — a structured analysis of where models converged (high-confidence signals), where they diverged, and what specific context would resolve the disagreement.

6. Two auth modes: API key or CLI session

Every model can use either an API key or the authentication from an installed CLI tool — per model, mixed freely in the same config. If you have claude and gemini CLIs installed and logged in, dissenter works with zero API key configuration. See CLI auth.

7. No OpenRouter dependency, genuine provider heterogeneity

Uses LiteLLM directly — a unified interface to 100+ providers. Cloud, local, and CLI-authenticated models all participate in the same ensemble.

---

Architecture

flowchart TD
    Q([Question]) --> CFG[Load dissenter.toml]
    CFG --> R1

    subgraph R1["Round 1: debate (parallel)"]
        M1[Model A\ndevil's advocate]
        M2[Model B\npragmatist]
        M3[Model C\nskeptic]
    end

    R1 --> CTX1[Collect outputs\n+ build context]
    CTX1 --> R2

    subgraph R2["Round 2: refine (parallel)"]
        M4[Model D\nanalyst]
        M5[Model E\ncontrarian]
    end

    R2 --> CTX2[Collect outputs\n+ build context]
    CTX2 --> FINAL

    subgraph FINAL["Final Round (1 or 2 models)"]
        direction LR
        CHAIR["1 model\nchairman → ADR"]
        OR["or"]
        CON["conservative"]
        LIB["liberal"]
        CON --> COMBINE[combine_model\nside-by-side MD]
        LIB --> COMBINE
    end

    FINAL --> OUT[decisions/decision_*.md]

---

Installation

Requires uv.

git clone <repo>
cd dissenter
make install

Choose your auth method — or mix them freely:

Option A — CLI auth (no API keys needed) If you have claude and/or gemini CLIs installed and logged in, set auth = "cli" in your config. Done.

Option B — API keys

export ANTHROPIC_API_KEY=...
export GEMINI_API_KEY=...          # or GOOGLE_API_KEY
export GROQ_API_KEY=...            # optional, free tier
export PERPLEXITY_API_KEY=...      # optional, web-search grounding

Option C — fully local, no credentials Use make ask-test — runs entirely on Ollama with ministral-3:3b. See Testing.

For Ollama models, start ollama serve before running.

---

Running

make install puts the package into a local .venv. The dissenter command is not on your PATH by default. Three options:

# Option 1 — make (recommended, always works from the project directory)
make ask Q="your question"
make show

# Option 2 — uv run (always works from the project directory)
uv run dissenter ask "your question"
uv run dissenter show

# Option 3 — install globally so bare `dissenter` works anywhere
uv tool install .
dissenter ask "your question"

Commands:

# Run a debate
make ask Q="Should I use Kafka or a Postgres outbox pattern?"

# Run with a custom config
uv run dissenter ask "..." --config ~/my-team/dissent.toml

# Run with a custom output directory
uv run dissenter ask "..." --output ./architecture/decisions

# Show configured rounds, models, and roles
make show

# Run local-only with no API keys
make ask-test Q="Should I use Kafka or a Postgres outbox pattern?"

The final decision is printed to stdout (clickable file link) and written to decisions/decision_<timestamp>.md.

---

Configuration

Edit dissenter.toml in the project directory, or ~/.config/dissenter/config.toml for a global default. Pass --config <path> to override.

Minimal config

output_dir = "decisions"

[[rounds]]
name = "debate"

[[rounds.models]]
id   = "anthropic/claude-sonnet-4-6"
role = "devil's advocate"

[[rounds.models]]
id   = "gemini/gemini-2.0-flash"
role = "pragmatist"

# Final round: must be exactly 1 or 2 enabled models
[[rounds]]
name = "final"

[[rounds.models]]
id      = "anthropic/claude-opus-4-6"
role    = "chairman"
timeout = 300

Multi-round

Rounds execute sequentially. Each round receives all prior rounds as context.

output_dir = "decisions"

[[rounds]]
name = "debate"

[[rounds.models]]
id    = "anthropic/claude-sonnet-4-6"
role  = "devil's advocate"
auth  = "cli"

[[rounds.models]]
id    = "gemini/gemini-2.0-flash"
role  = "pragmatist"
auth  = "cli"

[[rounds.models]]
id    = "ollama/mistral"
role  = "skeptic"
extra = { api_base = "http://localhost:11434" }

[[rounds]]
name = "refine"

[[rounds.models]]
id   = "gemini/gemini-2.0-flash"
role = "analyst"
auth = "cli"

[[rounds]]
name = "final"

[[rounds.models]]
id      = "anthropic/claude-opus-4-6"
role    = "chairman"
auth    = "cli"
timeout = 300

Dual-arbiter final

When the final round has exactly 2 models, set combine_model to produce a side-by-side recommendation document.

[[rounds]]
name            = "final"
combine_model   = "ollama/mistral"
combine_timeout = 60

[[rounds.models]]
id      = "anthropic/claude-opus-4-6"
role    = "conservative"
auth    = "cli"
timeout = 300

[[rounds.models]]
id      = "gemini/gemini-2.0-flash"
role    = "liberal"
auth    = "cli"
timeout = 300

CLI auth — no API keys

The default for every model is auth = "api" — litellm reads the API key from your environment. Set auth = "cli" to override on a per-model basis and use the provider's installed CLI instead. The prompt is piped to the CLI via stdin; the response is captured from stdout. Uses whatever session the CLI has — OAuth, browser login, enterprise SSO.

[[rounds.models]]
id   = "anthropic/claude-sonnet-4-6"
role = "devil's advocate"
auth = "cli"                  # uses `claude --print` via stdin

[[rounds.models]]
id   = "gemini/gemini-2.0-flash"
role = "pragmatist"
auth = "cli"                  # uses `gemini` via stdin

# Explicit CLI command (for providers not auto-detected)
[[rounds.models]]
id          = "anthropic/claude-opus-4-6"
role        = "chairman"
auth        = "cli"
cli_command = "claude"        # usually inferred automatically

Auto-detected CLI commands by provider prefix:

Provider prefix	CLI used
anthropic/	claude
gemini/ or google/	gemini
anything else	set cli_command explicitly

Same model, multiple roles

A round can list the same model ID multiple times with different roles. The dissenter-test.toml config does this to run the full pipeline with no API keys.

output_dir = "decisions/test"

[[rounds]]
name = "debate"

[[rounds.models]]
id    = "ollama/ministral-3:3b"
role  = "devil's advocate"
extra = { api_base = "http://localhost:11434" }

[[rounds.models]]
id    = "ollama/ministral-3:3b"
role  = "skeptic"
extra = { api_base = "http://localhost:11434" }

[[rounds.models]]
id    = "ollama/ministral-3:3b"
role  = "pragmatist"
extra = { api_base = "http://localhost:11434" }

[[rounds]]
name = "final"

[[rounds.models]]
id      = "ollama/ministral-3:3b"
role    = "chairman"
timeout = 180
extra   = { api_base = "http://localhost:11434" }

Random role distribution

Use [role_distribution] to randomly assign roles from a weighted distribution. Weights are relative.

[role_distribution]
"devil's advocate" = 0.3
"skeptic"          = 0.3
"pragmatist"       = 0.2
"contrarian"       = 0.2

Per-model API key

Override the environment variable with an explicit key per model.

[[rounds.models]]
id      = "anthropic/claude-sonnet-4-6"
role    = "devil's advocate"
api_key = "sk-ant-..."

---

Roles

Roles live in src/dissent/roles/*.toml. Each file defines a name, description, and prompt. Add a new .toml file to create a new role — no code changes needed.

Built-in roles

Role	Description	Typical round
devil's advocate	Argue against the obvious or popular choice	debate
pragmatist	Focus on what actually works in production at scale	debate
skeptic	Find hidden failure modes and long-term risks	debate
contrarian	Surface the minority expert view and missed nuance	debate
analyst	Rigorous balanced analysis with concrete numbers	debate / refine
researcher	Find the most current information using web access	debate
second opinion	Fresh-eyes independent review	refine
chairman	Decisive synthesis after all debate	final (1-model)
conservative	Pragmatic executor — safest proven path	final (2-model)
liberal	Ambitious visionary — boldest high-upside path	final (2-model)

Any string is a valid role — unknown roles fall back to the analyst prompt.

To add a custom role:

# src/dissent/roles/security_auditor.toml
name        = "security auditor"
description = "Identify attack surfaces and compliance risks"
prompt      = "Your role is security auditor. Identify the attack surface, likely CVEs, supply chain risks, and compliance implications of each option."

---

Output

Each run produces a decision file and a debug directory:

decisions/
  decision_20260320_143022.md          ← the ADR (commit this)
  debug_20260320_143022/
    round_1_debate/
      anthropic_claude-sonnet-4-6__devils_advocate.md
      gemini_gemini-2.0-flash__pragmatist.md
      ollama_mistral__skeptic.md
    round_2_refine/
      gemini_gemini-2.0-flash__analyst.md
    round_3_final/
      anthropic_claude-opus-4-6__chairman.md

The decision file path is printed as a clickable link at the end of each run. The ADR follows a structured format: Context, Consensus, Disagreements, Options table, Decision, Consequences, Mitigations, Open Questions.

---

Testing

make test       # runs the pytest suite

Testing without API keys — fully local:

ollama pull ministral-3:3b
ollama serve
make ask-test Q="Should I use Redis or Postgres for session storage?"

dissenter-test.toml runs ministral-3:3b with different roles across all rounds. It exercises the full multi-round pipeline with zero external API access.

ministral-3:3b is the recommended Ollama baseline. Fast, coherent under adversarial role prompting, and produces structured output reliably at 3B params.

---

Comparison

Feature	dissenter	llm-council	llm-consortium	consilium	MoA ref impl
Role-differentiated prompts	✓	✗	✗	✗	✗
Multi-round debate hierarchy	✓	✗	partial¹	partial²	partial³
Disagreement as structured output	✓	✗	✗	partial⁴	✗
Dual-arbiter output	✓	✗	✗	✗	✗
External role files	✓	✗	✗	✗	✗
Same model multiple roles	✓	✗	✗	✗	✗
CLI session auth (no API key)	✓	✗	✗	✗	✗
No OpenRouter/proxy required	✓	✗	✗	✓	✗
Local + cloud in same ensemble	✓	✗	✗	✗	✗
ADR output format	✓	✗	✗	✗	✗
Single-file config	✓	✗	partial	✗	✗
Per-model API key override	✓	✗	✗	✗	✗
Peer critique round	roadmap	partial⁵	✗	✓⁶	✗
Decision memory	roadmap	✗	✗	✗	✗
uv tool install	roadmap	✗	partial	✗	✗

¹ llm-consortium retries up to 3× when arbiter confidence < 0.8 — iteration toward convergence, not debate. ² consilium has configurable --rounds N in discuss/socratic modes. ³ MoA has configurable layers (default 3), but each layer refines toward consensus — no debate structure. ⁴ consilium uses ACH (Analysis of Competing Hypotheses) synthesis — the most honest competitor approach, but still ends in a verdict. ⁵ llm-council Stage 2 is anonymous peer ranking, not written critique of reasoning. ⁶ consilium has cross-pollination (models investigate each other's gaps) and a rotating challenger role.

---

Academic foundations

• Mixture of Agents (arXiv 2406.04692, TogetherAI, June 2024) — the canonical proposer→aggregator architecture. dissenter is a multi-layer MoA with adversarial role differentiation on the proposer layer.
• ICE: Iterative Critique and Ensemble (medrxiv, December 2024) — mutual critique between models before synthesis yields +7–45% accuracy on hard benchmarks. Basis for the planned --deep mode.
• LLM Ensemble Survey (arXiv 2502.18036, February 2025) — taxonomy of ensemble methods; identifies prompt diversity as the strongest lever.
• Rethinking MoA (OpenReview 2025) — finds diverse framing of the same question outperforms diverse models asked the same way. Direct justification for role-differentiated prompting.

---

Roadmap

Done in v0.2:

• Multi-round debate with context passing between rounds
• Role prompts as external TOML files (src/dissent/roles/*.toml)
• Dual-arbiter final round (conservative + liberal + combine_model)
• Random role distribution ([role_distribution] table)
• Per-model api_key override in [[rounds.models]]
• CLI session auth (auth = "cli") — use installed CLIs without API keys
• Same model, different roles in a single round
• dissenter show — rich tree view of configured rounds

Still to do:

• --deep flag: peer critique round (ICE paper, +7–45% accuracy on hard benchmarks)
• Disagreement classifier: factual vs. trade-off vs. context-dependent
• Persistent decision store: SQLite + embedding, surface past ADRs as context
• Confidence scoring: each model rates certainty and states what would change its answer
• Dynamic role inference: infer relevant roles from question type (security, performance, cost, maintainability)
• uv tool install distribution for global install without cloning