@consensus-tools/universal
Single-model AI pipelines are a reliability failure. One model, one prompt, one answer — no accountability. When an autonomous agent decides to send an email, merge code, or escalate permissions, nothing checks whether that's a good idea.
@consensus-tools/universal adds a governance layer in one line. Two operating modes:
- Regex mode (default): Three rule-based reviewers (security, compliance, user-impact) review tool output using pattern matching — no LLM calls, no network requests, sub-millisecond review overhead. The tool executes first, then reviewers evaluate.
- LLM Persona mode (activated when
config.modelis provided): Multiple AI personas deliberate on each tool call BEFORE execution, with reputation-weighted voting, automatic persona respawn, and risk-tier classification.
Both modes support the same consensus.wrap() API. Block, allow, or escalate based on a configurable policy. Adapters for LangChain, Vercel AI SDK, and MCP are loaded on demand.
Install
pnpm add @consensus-tools/universalOptional peer dependencies (install only what you need):
pnpm add @consensus-tools/langchain # LangChain adapter
pnpm add @consensus-tools/ai-sdk # Vercel AI SDK adapter
pnpm add @consensus-tools/mcp # MCP adapterQuick Start
Regex Mode — any tool executor
import { consensus } from "@consensus-tools/universal";
// Your existing tool executor
async function myExecutor(toolName: string, args: Record<string, unknown>) {
// call the actual tool
}
// Wrap it — all calls now go through deliberation
const safe = consensus.wrap(myExecutor);
const result = await safe("send_email", {
to: "user@example.com",
body: "Your invoice is attached.",
});Objects with .execute, .invoke, or .call methods are also accepted:
const safe = consensus.wrap(myToolRunner); // ToolExecutor function
const safe = consensus.wrap({ execute: fn }); // .execute method
const safe = consensus.wrap({ invoke: fn }); // .invoke method
const safe = consensus.wrap({ call: fn }); // .call methodLLM Persona Mode — multi-model deliberation
Provide a model adapter to activate LLM Persona Mode. Multiple AI personas deliberate on each tool call with reputation tracking and automatic respawn.
import { consensus } from "@consensus-tools/universal";
import Anthropic from "@anthropic-ai/sdk";
const client = new Anthropic();
// Provider-agnostic model adapter — accepts messages, returns text
// Note: ModelAdapter receives { role: "system" | "user" } messages.
// The Anthropic SDK uses a separate `system` param, so extract it here.
const model = async (messages) => {
const system = messages.filter((m) => m.role === "system").map((m) => m.content).join("\n");
const userMsgs = messages.filter((m) => m.role === "user").map((m) => ({ role: "user" as const, content: m.content }));
const res = await client.messages.create({
model: "claude-sonnet-4-20250514",
max_tokens: 512,
system,
messages: userMsgs,
});
return res.content[0].text;
};
const safe = consensus.wrap(myExecutor, {
model, // activates LLM mode
policy: "weighted_reputation", // any of 9 core policies
pack: "governance", // persona pack (default: "default")
mode: "enforce", // "enforce" (default) or "shadow"
});
const result = await safe("deploy_to_prod", { service: "api", version: "2.1.0" });Shadow mode logs every decision but never blocks — useful for evaluating governance before enforcing it:
const safe = consensus.wrap(myExecutor, {
model,
mode: "shadow",
onDecision: (decision) => metrics.track("consensus", decision),
});langchain — LangChain callback handler
import { consensus } from "@consensus-tools/universal";
import { ChatOpenAI } from "@langchain/openai";
import { AgentExecutor } from "langchain/agents";
const model = new ChatOpenAI({ model: "gpt-4o" });
// Returns a ConsensusGuardCallbackHandler — attach it to your chain's callbacks
const handler = await consensus.langchain(null, {
policy: "supermajority",
failPolicy: "closed",
});
const executor = AgentExecutor.fromAgentAndTools({ agent, tools, callbacks: [handler] });Requires @consensus-tools/langchain as a peer dependency. The first argument is unused (retained for API compatibility).
aiSdk — Vercel AI SDK adapter
import { consensus } from "@consensus-tools/universal";
import { generateText } from "ai";
import { openai } from "@ai-sdk/openai";
async function generate(prompt: string) {
return generateText({ model: openai("gpt-4o"), prompt });
}
// Wrap the generate function
const safeGenerate = await consensus.aiSdk(generate, {
policy: "majority",
guards: ["security", "compliance"],
});
const result = await safeGenerate("Summarize last quarter's results.");Requires @consensus-tools/ai-sdk as a peer dependency.
mcp — MCP server adapter
import { consensus } from "@consensus-tools/universal";
// Returns a guarded MCP server instance
const server = await consensus.mcp({
policy: "unanimous",
failPolicy: "closed",
logger: (event) => console.log("[mcp]", event),
});Requires @consensus-tools/mcp as a peer dependency.
Risk Tiers (LLM Mode)
In LLM mode, tools are classified by name into risk tiers. Low-risk tools fast-path through regex only (no LLM call). High-risk tools get full persona deliberation.
| Tier | Patterns | Examples |
|---|---|---|
| High (full deliberation) | send/email/mail, delete/remove/drop/destroy, write/update/patch/put/post/create/insert, deploy/release/publish/push, merge/commit/approve, grant/revoke/escalate/permission, execute/run/eval/exec, transfer/pay/charge/refund |
send_email, delete_user, deploy_to_prod |
| Low (regex fast-path) | get/fetch/read/list/search/query/find/lookup, check/verify/validate/inspect/describe, count/sum/aggregate/stats, view/show/display/render/preview |
get_user, list_orders, search_docs |
| Unknown | No pattern match | Defaults to high (safe by default) |
High-risk patterns are checked before low-risk patterns, but all patterns are prefix-anchored (^). A tool named get_and_delete_user matches the low-risk get prefix, not the high-risk delete. Use riskTiers overrides for compound names that start with read-like prefixes but perform writes.
Override per tool:
const safe = consensus.wrap(myExecutor, {
model,
riskTiers: {
"my_safe_tool": "low", // skip LLM deliberation
"my_risky_read": "high", // force full deliberation
},
});Persona Packs (LLM Mode)
Persona packs define which AI reviewers participate in deliberation.
| Pack | Personas | Use case |
|---|---|---|
"default" |
security, compliance, operations | General-purpose agent governance |
"governance" |
5 lifecycle specialists | Structured organizational oversight |
Override with custom personas:
import type { PersonaConfig } from "@consensus-tools/personas";
const safe = consensus.wrap(myExecutor, {
model,
personas: [
{ id: "sec", name: "Security Analyst", role: "Focus on injection, data exfiltration..." },
{ id: "ops", name: "Ops Lead", role: "Focus on blast radius, rollback..." },
],
});Reputation & Respawn (LLM Mode)
Each persona's vote accuracy is tracked over time. When a persona's reputation drops below the respawn threshold (default: 0.15), it is automatically replaced with a successor that inherits the original's failure modes.
const safe = consensus.wrap(myExecutor, {
model,
reputationStore: myStorage, // persist reputation across restarts
respawnThreshold: 0.2, // trigger respawn below this score
onFeedback: (signal) => auditLog.write(signal),
});
// After a human reviews a decision, send feedback to update reputation
safe.feedback({ decisionId: "dec_abc123", type: "override_block" });Reputation events are emitted via the logger:
// { event: "persona.respawned", data: { oldPersonaId, newPersonaId, reputation, reason } }Architecture
┌─────────────────────────────────────────────────────────────────────┐
│ @consensus-tools/universal │
│ │
│ consensus.wrap() consensus.langchain() consensus.aiSdk() │
│ consensus.mcp() │
└───────────────┬──────────────────────────────┬───────────────────────┘
│ resolves Wrappable → ToolExecutor
config.model? │
┌─── no ──┴──── yes ───┐ │
▼ ▼ │
┌──────────────────┐ ┌───────────────────────┐ │
│ REGEX MODE │ │ LLM PERSONA MODE │ │
│ │ │ │ │
│ @ct/wrapper │ │ deliberate() │ │
│ + GuardTemplate │ │ + ReputationManager │ │
│ reviewers │ │ + classifyTool() │ │
│ │ │ + resolveConsensus() │ │
└────────┬─────────┘ └──────────┬────────────┘ │
│ │ │
▼ ▼ │
3 regex reviewers N persona LLM calls │
(sub-ms, sync) (per-persona timeout │
→ regex fallback, NO) │
│ │ │
└───────────┬───────────┘ │
▼ │
action: allow | block | escalate │
+ audit artifact write │
┌─────────────────────────────────────────────────────────────────────┐
│ Configurable Guard Domains │
│ send_email code_merge publish support_reply │
│ agent_action deployment permission_escalation │
│ @consensus-tools/guards (Tier 1) │
└─────────────────────────────────────────────────────────────────────┘
Tier layout (consensus-tools layered architecture):
Tier 0 schemas, secrets
Tier 1 guards, telemetry, evals, integrations, notifications, personas
Tier 2 core (job engine, ledger), policies (9 algorithms)
Tier 3 workflows, wrapper
Tier 4 universal (this package), sdk-node, mcp, openclaw, cliAPI Reference
consensus.wrap(wrappable, config?)
Synchronously wraps a tool executor with consensus governance. Returns a ToolExecutor.
function wrap(wrappable: Wrappable, config?: Partial<UniversalConfig>): ToolExecutor| Parameter | Type | Description |
|---|---|---|
wrappable |
Wrappable |
Function, or object with .execute / .invoke / .call |
config |
Partial<UniversalConfig> |
Optional configuration (see Config Reference below) |
Returns: ToolExecutor — (toolName: string, args: Record<string, unknown>) => Promise<unknown>
consensus.langchain(_chain, config?)
Returns a ConsensusGuardCallbackHandler for LangChain. Dynamically imports @consensus-tools/langchain. The first argument is unused (attach the returned handler to your chain's callbacks array).
async function langchain(_chain: unknown, config?: Partial<UniversalConfig>): Promise<unknown>Throws MissingDependencyError if @consensus-tools/langchain is not installed.
consensus.aiSdk(fn, config?)
Async adapter for Vercel AI SDK generate functions. Dynamically imports @consensus-tools/ai-sdk.
async function aiSdk(fn: unknown, config?: Partial<UniversalConfig>): Promise<unknown>Throws MissingDependencyError if @consensus-tools/ai-sdk is not installed.
consensus.mcp(config?)
Async adapter that creates a guarded MCP server. Dynamically imports @consensus-tools/mcp.
async function mcp(config?: Partial<UniversalConfig>): Promise<unknown>Throws MissingDependencyError if @consensus-tools/mcp is not installed.
Configuration Reference
All options are optional. Defaults are shown.
interface UniversalConfig {
// ── Shared (both modes) ─────────────────────────────────────────────
policy?: string; // "majority"
guards?: string[]; // ["agent_action"]
failPolicy?: "closed" | "open"; // "closed"
storage?: "memory" | IStorage; // "memory"
logger?: boolean | ((event: LogEvent) => void); // true
onDecision?: (decision: DecisionResult<unknown> | LlmDecisionResult) => void;
onError?: (err: Error, action: unknown) => void;
// ── LLM Persona Mode (activated when `model` is provided) ──────────
model?: ModelAdapter; // undefined (regex mode)
pack?: string; // "default"
personas?: PersonaConfig[]; // overrides pack
mode?: ExecutionMode; // "enforce"
riskTiers?: RiskTierMap; // per-tool overrides
onFeedback?: (signal: FeedbackSignal) => void;
reputationStore?: IStorage; // persist reputation
respawnThreshold?: number; // 0.15
personaTimeout?: number; // 3000 (ms)
}policy
Maps to an aggregation strategy for reviewers.
| Value | Mode | Strategy |
|---|---|---|
"majority" (default) |
Both | More than half of reviewers approve. Regex: threshold strategy. LLM: net positive vote score. |
"supermajority" |
Both | Regex: threshold >= 0.67. LLM: APPROVAL_VOTE with minScore requiring >= 67% approval. |
"unanimous" |
Both | Regex: all must approve. LLM: APPROVAL_VOTE with minScore requiring all YES votes. |
"threshold:X" |
Both | Weighted average score >= X (0–1) |
"weighted_reputation" |
LLM only | Votes weighted by persona reputation score (alias for WEIGHTED_REPUTATION) |
"first_wins" |
LLM only | First persona's vote wins (alias for FIRST_SUBMISSION_WINS) |
"highest_confidence" |
LLM only | Highest confidence vote wins (alias for HIGHEST_CONFIDENCE_SINGLE) |
"top_k" |
LLM only | Top-K scoring split (alias for TOP_K_SPLIT) |
"owner_pick" |
LLM only | Designated owner persona decides (alias for OWNER_PICK) |
"arbiter" |
LLM only | Trusted arbiter persona decides (alias for TRUSTED_ARBITER) |
"MAJORITY_VOTE" |
LLM only | Direct majority vote count |
"WEIGHTED_REPUTATION" |
LLM only | Reputation-weighted vote scoring |
"WEIGHTED_VOTE_SIMPLE" |
LLM only | Explicit vote weight scoring |
"APPROVAL_VOTE" |
LLM only | Approval threshold-based voting |
"FIRST_SUBMISSION_WINS" |
LLM only | First persona's vote wins |
"HIGHEST_CONFIDENCE_SINGLE" |
LLM only | Highest confidence vote wins |
"OWNER_PICK" |
LLM only | Designated owner persona decides |
"TOP_K_SPLIT" |
LLM only | Top-K scoring split |
"TRUSTED_ARBITER" |
LLM only | Trusted arbiter persona decides |
Using an LLM-only policy without providing model emits a warning and falls back to majority.
guards
Guard domain names to use as reviewers. Defaults to ["agent_action"]. When the default is used, the actual reviewers are security, compliance, and user-impact (the three domains with implemented rule sets). Custom domain names are accepted but fall back to a permissive placeholder.
failPolicy
| Value | Behavior on deliberation error or block |
|---|---|
"closed" (default) |
Throw ConsensusBlockedError — the tool call does not execute |
"open" |
Allow the tool call to proceed despite the deliberation result |
failPolicy: "open" disables enforcement. Governance still runs (decisions are made, onDecision fires), but blocked actions execute anyway. It exists for development and testing only. In production, use "closed" (the default) and handle ConsensusBlockedError in your error path.
If you want to observe governance decisions without blocking in production, use onDecision instead:
consensus.wrap(executor, {
failPolicy: "closed",
onDecision: (result) => {
// Shadow mode: log every decision, including blocks, for analysis
metrics.track("consensus.decision", { action: result.action, score: result.aggregateScore });
},
});Production warning: failPolicy: "open" and storage: "memory" both emit console.warn when NODE_ENV=production.
storage
| Value | Description |
|---|---|
"memory" (default) |
In-process storage — decisions are not persisted across restarts |
IStorage |
Any storage backend implementing the IStorage interface (e.g., @consensus-tools/storage) |
logger
| Value | Description |
|---|---|
true (default) |
Emit structured log events via console.debug |
false |
Disable all logging |
(event: LogEvent) => void |
Custom log handler |
// Custom logger example
consensus.wrap(executor, {
logger: (event) => myObservability.track(event.event, event.data),
});onDecision
Called after every consensus deliberation, including allowed actions.
consensus.wrap(executor, {
onDecision: (result) => {
// Regex mode: result has `scores` array
// LLM mode: result has `votes` array and `decisionId`
auditLog.write({
action: result.action,
score: result.aggregateScore,
});
},
});onError
Called when an unexpected error occurs during deliberation, audit writes, or callbacks. In LLM mode, the context object may include a phase field: "onDecision" (callback error) or "audit_write" (storage error). General deliberation errors have no phase field. In regex mode, context contains { toolName, args }.
consensus.wrap(executor, {
onError: (err, context) => {
Sentry.captureException(err, { extra: context });
},
});model (LLM mode)
Provider-agnostic model adapter. Accepts ModelMessage[], returns raw text. When provided, activates LLM Persona Mode.
type ModelMessage = { role: "system" | "user"; content: string };
type ModelAdapter = (messages: ModelMessage[]) => Promise<string>;// OpenAI
const model = async (msgs) => {
const res = await openai.chat.completions.create({ model: "gpt-4o", messages: msgs });
return res.choices[0].message.content;
};
// Anthropic (system messages go in separate param)
const model = async (msgs) => {
const system = msgs.filter((m) => m.role === "system").map((m) => m.content).join("\n");
const messages = msgs.filter((m) => m.role === "user").map((m) => ({ role: "user" as const, content: m.content }));
const res = await anthropic.messages.create({ model: "claude-sonnet-4-20250514", max_tokens: 512, system, messages });
return res.content[0].text;
};mode (LLM mode)
| Value | Behavior |
|---|---|
"enforce" (default) |
Blocks tool execution when consensus rejects |
"shadow" |
Logs every decision but never blocks — always executes the tool |
Shadow mode is useful for evaluating governance accuracy before enabling enforcement. onDecision still fires in shadow mode.
personaTimeout (LLM mode)
Per-persona LLM call timeout in milliseconds. Default: 3000. When a persona times out, it falls back to regex evaluation (fail-closed — regex fallback votes NO by default).
onFeedback (LLM mode)
Notification hook fired after .feedback() processes a signal. Use it to log or react to reputation updates.
.feedback(signal) (LLM mode)
In LLM mode, consensus.wrap() returns an augmented executor with a .feedback() method for sending human feedback to the reputation system.
const safe = consensus.wrap(myExecutor, { model, logger: false });
// After observing a bad block, send feedback
safe.feedback({ decisionId: "dec_abc123", type: "override_block" });
// After observing a missed risk, send feedback
safe.feedback({ decisionId: "dec_xyz789", type: "flag_miss" });Wire onFeedback for logging:
const safe = consensus.wrap(myExecutor, {
model,
onFeedback: (signal) => {
console.log("Reputation updated for decision:", signal.decisionId);
},
});The .feedback() method is only available in LLM mode. In regex mode, the returned executor is a plain ToolExecutor.
Guard Domains
Three guard domains have implemented rule sets with regex pattern matching:
| Domain | Description |
|---|---|
security |
Flags destructive operations (delete, rm -rf), secret exposure, injection risks |
compliance |
Flags SSN patterns, PII (email addresses), regulated data |
user-impact |
Flags mass operations (broadcast, mass_delete), irreversible actions |
These are the default reviewers in regex mode. The guards config accepts any string, but unrecognized domains fall back to a permissive placeholder (always votes YES, risk 0.1).
The broader consensus-tools monorepo defines seven guard categories (agent_action, send_email, code_merge, publish, support_reply, deployment, permission_escalation) as standalone guard packages. These are separate from the universal package's built-in regex reviewers.
Error Reference
| Error class | When thrown |
|---|---|
ConsensusBlockedError |
Deliberation blocked the action and failPolicy is "closed" |
MissingDependencyError |
Optional peer dependency not installed (langchain / ai-sdk / mcp) |
ConfigError |
Invalid policy string passed to policyToStrategy |
import { ConsensusBlockedError, MissingDependencyError } from "@consensus-tools/universal";
try {
const result = await safe("permission_escalation", { scope: "admin" });
} catch (err) {
if (err instanceof ConsensusBlockedError) {
// Deliberation blocked this call — log and surface to operator
console.error("Blocked:", err.message);
}
}OWASP Agentic Top 10 Mapping
The table below maps the consensus-tools monorepo's guard domains to OWASP Agentic AI categories. The universal package's built-in regex reviewers (security, compliance, user-impact) cover categories 1, 3, 4, 6, 8, and 9. The full set of 7 domains is available via the standalone guard packages.
| # | OWASP Category | Addressed by guard domain(s) |
|---|---|---|
| 1 | Excessive Agency | agent_action, permission_escalation |
| 2 | Tool Misuse | agent_action, code_merge, deployment |
| 3 | Prompt Injection | agent_action, support_reply |
| 4 | Insecure Outputs | publish, support_reply, send_email |
| 5 | Unauthorized Actions | permission_escalation, agent_action, deployment |
| 6 | Data Leakage | send_email, publish, support_reply |
| 7 | Supply Chain Attacks | code_merge, deployment |
| 8 | Denial of Wallet | send_email, agent_action |
| 9 | Unintended Autonomy | agent_action, permission_escalation |
| 10 | Trust Boundary Violations | permission_escalation, code_merge, deployment |
What happens at runtime
Regex mode
safe("send_email", { to: "user@example.com", body: "Your invoice" })
│
├─ 1. Execute the tool function (the call runs first)
│
├─ 2. Review the output with 3 reviewers IN PARALLEL (rule-based, no LLM calls):
│ ├─ security: regex scan for destructive ops, secrets, injection
│ ├─ compliance: regex scan for SSN patterns, PII, email addresses
│ └─ user-impact: regex scan for mass ops, irreversible actions
│
├─ 3. Each reviewer returns: { score: 0-1, rationale: string, block?: boolean }
│
├─ 4. Aggregate scores via policy (majority/supermajority/unanimous/threshold)
│ └─ Any reviewer can hard-block (block: true) regardless of policy
│
├─ 5. Decision: allow | block | retry | escalate
│
├─ 6. Write audit artifact (to memory or configured storage)
│
└─ 7. Return output or throw ConsensusBlockedErrorNote: In regex mode, the tool executes BEFORE reviewers run. Reviewers evaluate the output, not the input. If blocked, the output is discarded but the side effect already occurred. For pre-execution governance (input screening before the tool runs), use LLM Persona Mode.
Latency: Sub-millisecond for the review phase. Reviewers are synchronous regex evaluators. Total overhead includes tool execution time.
Cost: Zero for the governance layer. No API calls, no tokens consumed.
DecisionResult object:
{
action: "allow", // "allow" | "block" | "retry" | "escalate"
output: { ... }, // return value from the tool executor
scores: [
{ score: 0.9, rationale: "No security concerns", block: false },
{ score: 0.5, rationale: "Email PII detected", block: false },
{ score: 0.9, rationale: "Low user impact", block: false },
],
aggregateScore: 0.77, // arithmetic mean across reviewers
attempt: 1, // retry count
}LLM Persona mode
safe("deploy_to_prod", { service: "api", version: "2.1.0" })
│
├─ 1. Await reputation load (first call only, if reputationStore configured)
│
├─ 2. Regex pre-screen (sub-ms, same as regex mode)
│
├─ 3. Classify tool risk tier:
│ ├─ Low-risk ("get_user") → fast-path regex only, skip LLM
│ └─ High-risk ("deploy_to_prod") → continue to LLM deliberation
│
├─ 4. Call N personas IN PARALLEL (with per-persona timeout):
│ ├─ Each persona receives: system prompt + tool name + args (truncated to 2000 chars)
│ ├─ Each returns: VOTE: YES|NO|REWRITE, CONFIDENCE: 0-1, rationale
│ └─ On timeout/parse failure → regex fallback (fail-closed, votes NO)
│
├─ 5. Synthesize votes into ConsensusInput (single submission, N votes)
│ └─ YES → +1, NO → -1, REWRITE → 0
│
├─ 6. Resolve consensus via policy (resolveConsensus)
│ └─ APPROVAL_VOTE: threshold check (empty winners → block)
│ └─ MAJORITY/WEIGHTED: score-based (positive → allow)
│ └─ Rewrite majority → escalate
│
├─ 7. Record decision for feedback correlation (capped at 500 entries)
│
├─ 8. Write audit entry to storage
│
├─ 9. Fire onDecision callback (errors routed to onError, never affect decision)
│
└─ 10. enforce mode → block/throw | shadow mode → always executeLatency: Depends on model latency. Timeout per persona defaults to 3000ms. Low-risk tools bypass LLM entirely.
Cost: N model calls per high-risk tool invocation (where N = number of personas, default 3).
LlmDecisionResult object:
{
decisionId: "a1b2c3d4-...", // unique ID for feedback correlation
action: "allow", // "allow" | "block" | "escalate"
votes: [
{
personaId: "sec",
personaName: "Security Analyst",
vote: "YES",
confidence: 0.85,
rationale: "No injection patterns detected",
source: "llm", // "llm" or "regex_fallback"
},
// ...
],
policy: "WEIGHTED_REPUTATION",
consensusTrace: { ... }, // full trace from resolveConsensus
aggregateScore: 0.82,
}Exports
| Export | Kind | Description |
|---|---|---|
consensus |
Object | Main facade with .wrap(), .langchain(), .aiSdk(), .mcp() |
resolveWrappable |
Function | Resolves a Wrappable to a plain ToolExecutor |
policyToStrategy |
Function | Maps a policy name string to a StrategyConfig (regex mode) |
resolvePolicyType |
Function | Maps a policy string to a core policy type name (LLM mode) |
createLogger |
Function | Creates structured lifecycle log hooks |
ReputationManager |
Class | Per-persona reputation tracking with respawn (LLM mode) |
classifyTool |
Function | Classifies a tool name into a risk tier (LLM mode) |
deliberate |
Function | LLM persona deliberation engine (LLM mode) |
DEFAULTS |
Const | Default config values |
DEFAULT_GUARD |
Const | "agent_action" |
DEFAULT_POLICY |
Const | "majority" |
DEFAULT_PERSONA_TRIO |
Const | ["security", "compliance", "user-impact"] |
DEFAULT_PERSONA_COUNT |
Const | 3 |
DEFAULT_PACK |
Const | "default" |
ConsensusBlockedError |
Class | Thrown when deliberation blocks and failPolicy is "closed" |
MissingDependencyError |
Class | Thrown when an optional peer dep is not installed |
ConfigError |
Class | Thrown for invalid configuration (e.g., unknown policy name) |
AugmentedExecutor |
Type | ToolExecutor & { feedback(signal: FeedbackSignal): void } |
Wrappable |
Type | ToolExecutor | { execute } | { invoke } | { call } |
ToolExecutor |
Type | (toolName: string, args: Record<string, unknown>) => Promise<unknown> |
UniversalConfig |
Type | Full configuration interface |
ModelAdapter |
Type | (messages: ModelMessage[]) => Promise<string> |
ModelMessage |
Type | { role: "system" | "user"; content: string } |
LlmDecisionResult |
Type | Decision result from LLM persona deliberation |
FeedbackSignal |
Type | Human feedback signal for reputation updates |
ExecutionMode |
Type | "enforce" | "shadow" |
FailPolicy |
Type | "closed" | "open" |
RiskTier |
Type | "low" | "high" |
RiskTierMap |
Type | Record<string, RiskTier> |
LogEvent |
Type | { event: string; data: Record<string, unknown>; timestamp: number } |