opencode-threadweaver

Multi-Agent Debate Plugin for OpenCode. Spawns 4-16 specialized agents that independently analyze, cross-critique, revise, and synthesize a weighted final answer.

Install

# In your OpenCode config directory
npm install opencode-threadweaver

Add to opencode.jsonc:

{
  "plugin": [
    "opencode-threadweaver"
  ]
}

How It Works

User Query
    |
    v
Complexity Router  -->  standard (4 agents) or heavy (4 core + up to 12 specialists)
    |
    v
Round 1: PARALLEL ANALYSIS    -- each agent independently analyzes the query
Round 2: CROSS-CRITIQUE       -- agents critique each other's outputs
Round 3: REVISION             -- agents revise based on critiques
    |
    v
SYNTHESIS: Captain merges all outputs with dynamic weights
    |
    v
Result: synthesis + weights + dissents + round details

Core Agents (Standard Mode)

| ID | Role | Focus | |----|------|-------| | captain | Captain | Task decomposition, mediation, final synthesis | | verifier | Verifier | Fact-checking, evidence validation, data verification | | reasoner | Reasoner | Logical analysis, step-by-step reasoning | | creator | Creator | Blind spot detection, alternative perspectives |

Specialist Agents (Heavy Mode)

Activated via sparse activation when the query touches specific domains:

• Verifier specialists: security, performance, architecture
• Reasoner specialists: math, algorithms, systems
• Creator specialists: naming, simplification, alternatives
• Captain specialists: priorities, risks, communication

Dynamic Weighting

Each agent's contribution is weighted by:

final_weight = normalize(
    selfConfidence * 0.25 +
    crossAgreement * 0.45 +
    domainRelevance * 0.30
)

Consensus (crossAgreement) carries the highest weight -- conclusions agreed upon by multiple agents are prioritized.

Keyword Triggers

The plugin auto-activates when these keywords appear in a message (outside code blocks):

debate | weave | threadweave | multi-perspective

Japanese keywords:

debate | threadweave | weave

Configuration

Create opencode-threadweaver.jsonc in your OpenCode config directory (~/.config/opencode/):

{
  // Model for all debate agents
  "model": "openai/gpt-4o",

  // Per-role model overrides ("*" = default for unspecified roles)
  "roleModels": {
    "captain": "openai/gpt-4o",
    "*": "anthropic/claude-sonnet-4"
  },

  // Number of debate rounds (1-5)
  "debateRounds": 3,

  // Complexity score threshold for heavy mode (1+)
  "heavyThreshold": 5,

  // Force complexity level: "standard" | "heavy" | null (auto)
  "forceComplexity": null,

  // Minimum domain relevance for specialist activation (0-1)
  "activationThreshold": 0.3,

  // Timeout per round (ms)
  "roundTimeoutMs": 120000,

  // Total timeout (ms)
  "totalTimeoutMs": 600000,

  // Disable specific hooks
  "disabled_hooks": [],

  // Concurrency control
  "concurrency": {
    "defaultConcurrency": 4,
    "providerConcurrency": {},
    "modelConcurrency": {}
  }
}

Project-level config can be placed at .opencode/opencode-threadweaver.jsonc (overrides user-level).

Profiles

Define multiple configurations and switch between them via activeProfile or the THREADWEAVER_PROFILE environment variable:

{
  "activeProfile": "cloud",
  "profiles": {
    "cloud": {
      "forceComplexity": "standard",
      "roleModels": { "captain": "openai/gpt-4o", "*": "anthropic/claude-sonnet-4" },
      "debateRounds": 1
    },
    "cloud-heavy": {
      "forceComplexity": "heavy",
      "roleModels": { "captain": "openai/gpt-4o", "*": "anthropic/claude-sonnet-4" },
      "debateRounds": 3
    },
    "local": {
      "forceComplexity": "standard",
      "roleModels": { "captain": "ollama/llama3:70b", "*": "ollama/mistral:7b" },
      "debateRounds": 1
    },
    "local-heavy": {
      "forceComplexity": "heavy",
      "roleModels": { "captain": "ollama/llama3:70b", "*": "ollama/mistral:7b" },
      "debateRounds": 2
    }
  }
}

Switch profiles at runtime:

# Via environment variable (takes priority)
THREADWEAVER_PROFILE=local opencode exec "debate: best caching strategy"

# Or set activeProfile in config for the default

Profile settings are merged on top of any base config values defined outside profiles.

Direct Tool Usage

threadweaver({ query: "Compare React vs Vue for a large-scale enterprise app" })

With options:

threadweaver({
  query: "Security implications of microservices vs monolith",
  complexity: "heavy",
  rounds: 3,
  json: true
})

CLI (non-interactive)

opencode run --format json "debate: React vs Vue"

Output

The tool returns:

• synthesis -- the final merged answer
• weights -- per-agent confidence, agreement, and final weight
• dissents -- topics where agents did not reach consensus
• round details -- collapsible per-round outputs

JSON mode (json: true) returns the full structured result for pipeline integration.

Architecture

src/
  index.ts                 -- Plugin entry, keyword detector, hooks
  types.ts                 -- All TypeScript type definitions
  complexity-router.ts     -- standard/heavy routing + sparse activation
  roles.ts                 -- 4 core + 12 specialist role definitions
  protocol.ts              -- Debate loop (runThreadWeaver)
  weighting.ts             -- Dynamic weights, confidence extraction, critique parsing
  prompts.ts               -- Round prompt templates
  threadweaver-tool.ts     -- Tool registration + output formatting
  config.ts                -- Config loading (user + project level) + profile resolution
  shared.ts                -- Utilities (sanitizer, concurrency pool, resilience, TtlMap)

Development

bun install
bun test        # 39 tests
bun run build   # outputs to dist/

Gotchas

• Do NOT export non-function values from the plugin entry point. OpenCode treats all exports as plugin instances and will try to call them. Only export the plugin function and type-only exports.
• The plugin is standalone -- no dependency on opencode-ultra.
• All shared utilities (sanitizer, concurrency pool, etc.) are inlined in shared.ts.

License

See LICENSE.md