gavio 0.1.0 → 0.3.0

package/dist/cjs/providers/azure-openai.js

@@ -0,0 +1,56 @@
+"use strict";
+/** azureOpenaiAdapter — Azure OpenAI deployment-based chat completions. */
+Object.defineProperty(exports, "__esModule", { value: true });
+exports.azureOpenaiAdapter = azureOpenaiAdapter;
+const errors_js_1 = require("../errors.js");
+const types_js_1 = require("../types.js");
+const base_js_1 = require("./base.js");
+const http_js_1 = require("./http.js");
+const DEFAULT_API_VERSION = '2024-06-01';
+class AzureOpenAIAdapter extends base_js_1.BaseProviderAdapter {
+    apiKey;
+    endpoint;
+    deployment;
+    apiVersion;
+    timeoutSeconds;
+    constructor(options = {}) {
+        super(options.pricing);
+        this.apiKey = options.apiKey ?? process.env['AZURE_OPENAI_API_KEY'];
+        this.endpoint = (options.endpoint ?? process.env['AZURE_OPENAI_ENDPOINT'] ?? '').replace(/\/+$/, '');
+        this.deployment = options.deployment ?? process.env['AZURE_OPENAI_DEPLOYMENT'];
+        this.apiVersion = options.apiVersion ?? DEFAULT_API_VERSION;
+        this.timeoutSeconds = (options.timeoutMs ?? 30_000) / 1000;
+    }
+    get providerName() {
+        return 'azure_openai';
+    }
+    url(request) {
+        const deployment = this.deployment ?? request.model;
+        return `${this.endpoint}/openai/deployments/${deployment}/chat/completions?api-version=${this.apiVersion}`;
+    }
+    async complete(request) {
+        if (!this.apiKey || !this.endpoint) {
+            throw new errors_js_1.ConfigurationError('AZURE_OPENAI_API_KEY and AZURE_OPENAI_ENDPOINT must be set');
+        }
+        const started = performance.now();
+        const payload = {
+            messages: request.messages,
+            temperature: request.temperature,
+            max_tokens: request.maxTokens,
+        };
+        const data = await (0, http_js_1.postJson)(this.url(request), payload, { 'api-key': this.apiKey }, this.timeoutSeconds);
+        const choices = data['choices'] ?? [];
+        const message = choices[0]?.['message'] ?? {};
+        const content = message['content'] ?? '';
+        const usageData = data['usage'] ?? {};
+        const usage = new types_js_1.TokenUsage(usageData['prompt_tokens'] ?? 0, usageData['completion_tokens'] ?? 0);
+        return this.buildResponse(request, content, usage, data['model'] ?? request.model, started);
+    }
+    async healthCheck() {
+        return !!(this.apiKey && this.endpoint);
+    }
+}
+/** Factory: build an Azure OpenAI provider adapter. */
+function azureOpenaiAdapter(options = {}) {
+    return new AzureOpenAIAdapter(options);
+}

package/dist/cjs/providers/base.js

@@ -4,6 +4,7 @@ Object.defineProperty(exports, "__esModule", { value: true });
 exports.BaseProviderAdapter = void 0;
 const pricing_js_1 = require("../pricing.js");
 const response_js_1 = require("../response.js");
+const types_js_1 = require("../types.js");
 /** Base class with a shared pricing provider and response builder. */
 class BaseProviderAdapter {
     pricing;
@@ -13,6 +14,14 @@ class BaseProviderAdapter {
     get reportedModelVersion() {
         return null;
     }
+    /**
+     * Build a response from a fully buffered stream (F-REL-06). Streamed chunks
+     * carry text only, so token usage is estimated from prompt + content.
+     */
+    buildStreamResponse(request, content, startedAt) {
+        const usage = new types_js_1.TokenUsage((0, pricing_js_1.estimateTokens)(request.promptText()), (0, pricing_js_1.estimateTokens)(content));
+        return this.buildResponse(request, content, usage, this.reportedModelVersion ?? request.model, startedAt);
+    }
     buildResponse(request, content, usage, modelVersion, startedAt) {
         const latencyMs = Math.floor(performance.now() - startedAt);
         return new response_js_1.GavioResponse({

package/dist/cjs/providers/gemini.js

@@ -0,0 +1,73 @@
+"use strict";
+/** geminiAdapter — Google Generative Language API (generateContent). */
+Object.defineProperty(exports, "__esModule", { value: true });
+exports.geminiToContents = geminiToContents;
+exports.geminiAdapter = geminiAdapter;
+const errors_js_1 = require("../errors.js");
+const types_js_1 = require("../types.js");
+const base_js_1 = require("./base.js");
+const http_js_1 = require("./http.js");
+const DEFAULT_BASE_URL = 'https://generativelanguage.googleapis.com/v1beta';
+/** Map Gavio messages to Gemini contents + a system instruction. */
+function geminiToContents(messages) {
+    let system = null;
+    const contents = [];
+    for (const m of messages) {
+        const text = m.content;
+        if (m.role === 'system') {
+            system = system ? `${system}\n${text}` : text;
+            continue;
+        }
+        contents.push({ role: m.role === 'assistant' ? 'model' : 'user', parts: [{ text }] });
+    }
+    return { system, contents };
+}
+class GeminiAdapter extends base_js_1.BaseProviderAdapter {
+    apiKey;
+    baseUrl;
+    timeoutSeconds;
+    constructor(options = {}) {
+        super(options.pricing);
+        this.apiKey =
+            options.apiKey ?? process.env['GEMINI_API_KEY'] ?? process.env['GOOGLE_API_KEY'];
+        this.baseUrl = (options.baseUrl ?? DEFAULT_BASE_URL).replace(/\/+$/, '');
+        this.timeoutSeconds = (options.timeoutMs ?? 30_000) / 1000;
+    }
+    get providerName() {
+        return 'gemini';
+    }
+    payload(request) {
+        const { system, contents } = geminiToContents(request.messages);
+        const payload = {
+            contents,
+            generationConfig: {
+                temperature: request.temperature,
+                maxOutputTokens: request.maxTokens,
+            },
+        };
+        if (system)
+            payload['systemInstruction'] = { parts: [{ text: system }] };
+        return payload;
+    }
+    async complete(request) {
+        if (!this.apiKey)
+            throw new errors_js_1.ConfigurationError('GEMINI_API_KEY not set');
+        const started = performance.now();
+        const url = `${this.baseUrl}/models/${request.model}:generateContent?key=${this.apiKey}`;
+        const data = await (0, http_js_1.postJson)(url, this.payload(request), {}, this.timeoutSeconds);
+        const candidates = data['candidates'] ?? [{}];
+        const contentObj = candidates[0]?.['content'] ?? {};
+        const parts = contentObj['parts'] ?? [];
+        const content = parts.map((p) => p.text ?? '').join('');
+        const um = data['usageMetadata'] ?? {};
+        const usage = new types_js_1.TokenUsage(um['promptTokenCount'] ?? 0, um['candidatesTokenCount'] ?? 0);
+        return this.buildResponse(request, content, usage, request.model, started);
+    }
+    async healthCheck() {
+        return !!this.apiKey;
+    }
+}
+/** Factory: build a Gemini provider adapter. */
+function geminiAdapter(options = {}) {
+    return new GeminiAdapter(options);
+}

package/dist/cjs/providers/index.js

@@ -1,12 +1,15 @@
 "use strict";
 /** Provider adapters and the provider registry. */
 Object.defineProperty(exports, "__esModule", { value: true });
-exports.Provider = exports.anthropicAdapter = exports.openaiAdapter = exports.mockProvider = exports.BaseProviderAdapter = void 0;
+exports.Provider = exports.ollamaAdapter = exports.azureOpenaiAdapter = exports.geminiAdapter = exports.anthropicAdapter = exports.openaiAdapter = exports.mockProvider = exports.BaseProviderAdapter = void 0;
 exports.buildAdapter = buildAdapter;
 const errors_js_1 = require("../errors.js");
 const types_js_1 = require("../types.js");
 const anthropic_js_1 = require("./anthropic.js");
+const azure_openai_js_1 = require("./azure-openai.js");
+const gemini_js_1 = require("./gemini.js");
 const mock_js_1 = require("./mock.js");
+const ollama_js_1 = require("./ollama.js");
 const openai_js_1 = require("./openai.js");
 var base_js_1 = require("./base.js");
 Object.defineProperty(exports, "BaseProviderAdapter", { enumerable: true, get: function () { return base_js_1.BaseProviderAdapter; } });
@@ -16,19 +19,32 @@ var openai_js_2 = require("./openai.js");
 Object.defineProperty(exports, "openaiAdapter", { enumerable: true, get: function () { return openai_js_2.openaiAdapter; } });
 var anthropic_js_2 = require("./anthropic.js");
 Object.defineProperty(exports, "anthropicAdapter", { enumerable: true, get: function () { return anthropic_js_2.anthropicAdapter; } });
+var gemini_js_2 = require("./gemini.js");
+Object.defineProperty(exports, "geminiAdapter", { enumerable: true, get: function () { return gemini_js_2.geminiAdapter; } });
+var azure_openai_js_2 = require("./azure-openai.js");
+Object.defineProperty(exports, "azureOpenaiAdapter", { enumerable: true, get: function () { return azure_openai_js_2.azureOpenaiAdapter; } });
+var ollama_js_2 = require("./ollama.js");
+Object.defineProperty(exports, "ollamaAdapter", { enumerable: true, get: function () { return ollama_js_2.ollamaAdapter; } });
 var types_js_2 = require("../types.js");
 Object.defineProperty(exports, "Provider", { enumerable: true, get: function () { return types_js_2.Provider; } });
-/** Instantiate the default adapter for a provider id. v0.1.0: OpenAI, Anthropic, Mock. */
+/** Instantiate the default adapter for a provider id. */
 function buildAdapter(provider, pricing) {
     const p = (0, types_js_1.coerceProvider)(provider);
+    const opts = pricing ? { pricing } : {};
     switch (p) {
         case types_js_1.Provider.OPENAI:
-            return (0, openai_js_1.openaiAdapter)(pricing ? { pricing } : {});
+            return (0, openai_js_1.openaiAdapter)(opts);
         case types_js_1.Provider.ANTHROPIC:
-            return (0, anthropic_js_1.anthropicAdapter)(pricing ? { pricing } : {});
+            return (0, anthropic_js_1.anthropicAdapter)(opts);
+        case types_js_1.Provider.GEMINI:
+            return (0, gemini_js_1.geminiAdapter)(opts);
+        case types_js_1.Provider.AZURE_OPENAI:
+            return (0, azure_openai_js_1.azureOpenaiAdapter)(opts);
+        case types_js_1.Provider.OLLAMA:
+            return (0, ollama_js_1.ollamaAdapter)(opts);
         case types_js_1.Provider.MOCK:
-            return (0, mock_js_1.mockProvider)(pricing ? { pricing } : {});
+            return (0, mock_js_1.mockProvider)(opts);
         default:
-            throw new errors_js_1.ConfigurationError(`Provider '${p}' is not available in v0.1.0 (available: openai, anthropic, mock)`);
+            throw new errors_js_1.ConfigurationError(`Provider '${p}' is not available (v0.3.0 adds bedrock, cohere)`);
     }
 }

package/dist/cjs/providers/ollama.js

@@ -0,0 +1,41 @@
+"use strict";
+/** ollamaAdapter — local models via the Ollama chat API. */
+Object.defineProperty(exports, "__esModule", { value: true });
+exports.ollamaAdapter = ollamaAdapter;
+const types_js_1 = require("../types.js");
+const base_js_1 = require("./base.js");
+const http_js_1 = require("./http.js");
+const DEFAULT_BASE_URL = 'http://localhost:11434';
+class OllamaAdapter extends base_js_1.BaseProviderAdapter {
+    baseUrl;
+    timeoutSeconds;
+    constructor(options = {}) {
+        super(options.pricing);
+        this.baseUrl = (options.baseUrl ?? process.env['OLLAMA_HOST'] ?? DEFAULT_BASE_URL).replace(/\/+$/, '');
+        this.timeoutSeconds = (options.timeoutMs ?? 60_000) / 1000;
+    }
+    get providerName() {
+        return 'ollama';
+    }
+    async complete(request) {
+        const started = performance.now();
+        const payload = {
+            model: request.model,
+            messages: request.messages,
+            stream: false,
+            options: { temperature: request.temperature },
+        };
+        const data = await (0, http_js_1.postJson)(`${this.baseUrl}/api/chat`, payload, {}, this.timeoutSeconds);
+        const message = data['message'] ?? {};
+        const content = message['content'] ?? '';
+        const usage = new types_js_1.TokenUsage(data['prompt_eval_count'] ?? 0, data['eval_count'] ?? 0);
+        return this.buildResponse(request, content, usage, data['model'] ?? request.model, started);
+    }
+    async healthCheck() {
+        return true;
+    }
+}
+/** Factory: build an Ollama provider adapter. */
+function ollamaAdapter(options = {}) {
+    return new OllamaAdapter(options);
+}

package/dist/cjs/request.js

@@ -19,6 +19,7 @@ class GavioRequest {
     sessionId;
     options;
     metadata;
+    lineage;
     constructor(init) {
         this.messages = init.messages;
         this.model = init.model;
@@ -29,6 +30,7 @@ class GavioRequest {
         this.sessionId = init.sessionId ?? null;
         this.options = init.options ?? {};
         this.metadata = init.metadata ?? {};
+        this.lineage = init.lineage != null ? types_js_1.PromptLineage.from(init.lineage) : null;
     }
     get temperature() {
         const t = this.options['temperature'];
@@ -54,6 +56,7 @@ class GavioRequest {
             sessionId: this.sessionId,
             options: { ...this.options },
             metadata: { ...this.metadata },
+            lineage: this.lineage,
         });
     }
 }

package/dist/cjs/shim/openai.js

@@ -0,0 +1,57 @@
+"use strict";
+/**
+ * OpenAI drop-in shim (F-DX-04) — point existing OpenAI SDK code at Gavio.
+ *
+ *   import { Gateway } from 'gavio'
+ *   import { GavioOpenAI } from 'gavio/shim/openai'
+ *
+ *   const client = new GavioOpenAI(new Gateway({ provider: 'openai', model: 'gpt-4o' }))
+ *   const resp = await client.chat.completions.create({
+ *     model: 'gpt-4o', messages: [{ role: 'user', content: 'hi' }],
+ *   })
+ *   console.log(resp.choices[0].message.content)
+ */
+Object.defineProperty(exports, "__esModule", { value: true });
+exports.GavioOpenAI = void 0;
+function toCompletion(resp) {
+    return {
+        id: resp.traceId,
+        object: 'chat.completion',
+        model: resp.modelVersion || resp.model,
+        choices: [
+            { index: 0, message: { role: 'assistant', content: resp.content }, finish_reason: 'stop' },
+        ],
+        usage: {
+            prompt_tokens: resp.usage.promptTokens,
+            completion_tokens: resp.usage.completionTokens,
+            total_tokens: resp.usage.totalTokens,
+        },
+        gavio: {
+            costUsd: resp.costUsd,
+            cacheHit: resp.cacheHit,
+            interceptorsFired: resp.interceptorsFired,
+        },
+    };
+}
+class Completions {
+    gw;
+    constructor(gw) {
+        this.gw = gw;
+    }
+    async create(params) {
+        const resp = await this.gw.complete({
+            messages: params.messages,
+            model: params.model,
+            options: { temperature: params.temperature ?? 0.7, maxTokens: params.max_tokens ?? 1024 },
+        });
+        return toCompletion(resp);
+    }
+}
+/** OpenAI-client-shaped facade over a Gavio Gateway. */
+class GavioOpenAI {
+    chat;
+    constructor(gateway) {
+        this.chat = { completions: new Completions(gateway) };
+    }
+}
+exports.GavioOpenAI = GavioOpenAI;

package/dist/cjs/types.js

@@ -1,7 +1,7 @@
 "use strict";
 /** Shared enums (as string unions / const objects) and utility types. */
 Object.defineProperty(exports, "__esModule", { value: true });
-exports.TokenUsage = exports.GuardrailOutcome = exports.Sensitivity = exports.PiiMode = exports.CacheType = exports.Provider = void 0;
+exports.PromptLineage = exports.RagChunk = exports.TokenUsage = exports.GuardrailOutcome = exports.Sensitivity = exports.PiiMode = exports.CacheType = exports.Provider = void 0;
 exports.coerceProvider = coerceProvider;
 /** Supported LLM providers. String-valued for easy config + logging. */
 exports.Provider = {
@@ -59,3 +59,55 @@ class TokenUsage {
     }
 }
 exports.TokenUsage = TokenUsage;
+/**
+ * A single retrieved source that contributed to a prompt. Carries a *reference*
+ * to the source — never the retrieved text — so prompt lineage stays within the
+ * audit record's metadata-only contract.
+ */
+class RagChunk {
+    source;
+    chunkId;
+    score;
+    constructor(init) {
+        this.source = init.source;
+        this.chunkId = init.chunkId ?? null;
+        this.score = init.score ?? null;
+    }
+    toJSON() {
+        return { source: this.source, chunkId: this.chunkId, score: this.score };
+    }
+}
+exports.RagChunk = RagChunk;
+/**
+ * Provenance for a rendered prompt (F-OBS-04): the template, the variable
+ * bindings interpolated into it, and the RAG chunk sources retrieved for it.
+ *
+ * Attached to a GavioRequest by the caller and copied into the AuditRecord so
+ * any prompt can be reconstructed and debugged. RAG chunk text is never stored
+ * — only source references (see {@link RagChunk}).
+ */
+class PromptLineage {
+    templateId;
+    templateVersion;
+    variables;
+    ragChunks;
+    constructor(init = {}) {
+        this.templateId = init.templateId ?? null;
+        this.templateVersion = init.templateVersion ?? null;
+        this.variables = init.variables ?? {};
+        this.ragChunks = (init.ragChunks ?? []).map((c) => c instanceof RagChunk ? c : new RagChunk(c));
+    }
+    /** Coerce a PromptLineage instance or plain init object into a PromptLineage. */
+    static from(value) {
+        return value instanceof PromptLineage ? value : new PromptLineage(value);
+    }
+    toJSON() {
+        return {
+            templateId: this.templateId,
+            templateVersion: this.templateVersion,
+            variables: this.variables,
+            ragChunks: this.ragChunks.map((c) => c.toJSON()),
+        };
+    }
+}
+exports.PromptLineage = PromptLineage;

package/dist/esm/config.d.ts

@@ -0,0 +1,12 @@
+/**
+ * Config loader (F-DX-05) — build a Gateway from an object or a JSON file.
+ *
+ *   const gw = await Gateway.fromConfig('gateway.json')
+ *
+ * JSON is supported out of the box; string values expand ${ENV_VAR}.
+ */
+import { Gateway } from './gateway.js';
+type Cfg = Record<string, unknown>;
+export declare function loadConfig(path: string): Cfg;
+export declare function buildFromConfig(config: Cfg): Gateway;
+export {};

package/dist/esm/config.js

@@ -0,0 +1,102 @@
+/**
+ * Config loader (F-DX-05) — build a Gateway from an object or a JSON file.
+ *
+ *   const gw = await Gateway.fromConfig('gateway.json')
+ *
+ * JSON is supported out of the box; string values expand ${ENV_VAR}.
+ */
+import { readFileSync } from 'node:fs';
+import { ConfigurationError } from './errors.js';
+import { Gateway } from './gateway.js';
+import { auditInterceptor } from './interceptors/audit/index.js';
+import { hashingEmbedder, semanticCache } from './interceptors/cache/index.js';
+import { costControl, modelPolicy, rateLimiter } from './interceptors/governance/index.js';
+import { promptInjectionGuard } from './interceptors/injection.js';
+import { piiGuard } from './interceptors/pii/index.js';
+import { retryInterceptor, timeoutPolicy } from './interceptors/reliability/index.js';
+export function loadConfig(path) {
+    const text = readFileSync(path, 'utf8');
+    if (!path.endsWith('.json')) {
+        throw new ConfigurationError('JS config loader supports JSON only (use .json)');
+    }
+    return expand(JSON.parse(text));
+}
+function expand(obj) {
+    if (Array.isArray(obj))
+        return obj.map(expand);
+    if (obj && typeof obj === 'object') {
+        return Object.fromEntries(Object.entries(obj).map(([k, v]) => [k, expand(v)]));
+    }
+    if (typeof obj === 'string') {
+        return obj.replace(/\$\{(\w+)\}/g, (_, v) => process.env[v] ?? '');
+    }
+    return obj;
+}
+export function buildFromConfig(config) {
+    const gatewayOptions = {};
+    if (config['provider'])
+        gatewayOptions['provider'] = config['provider'];
+    if (config['model'])
+        gatewayOptions['model'] = config['model'];
+    if (config['devMode'] ?? config['dev_mode'])
+        gatewayOptions['devMode'] = true;
+    if (config['dryRun'] ?? config['dry_run'])
+        gatewayOptions['dryRun'] = true;
+    let gw = new Gateway(gatewayOptions);
+    const ic = config['interceptors'] ?? {};
+    const cfg = (name) => {
+        const entry = ic[name];
+        return entry && entry['enabled'] !== false ? entry : null;
+    };
+    let c;
+    if ((c = cfg('audit'))) {
+        gw = gw.use(auditInterceptor({
+            sink: c['sink'] ?? 'stdout',
+            hashChain: Boolean(c['hashChain'] ?? c['hash_chain']),
+        }));
+    }
+    if ((c = cfg('prompt_injection'))) {
+        gw = gw.use(promptInjectionGuard({ action: c['action'] ?? 'block' }));
+    }
+    if ((c = cfg('pii_guard'))) {
+        gw = gw.use(piiGuard({
+            sensitivity: c['sensitivity'] ?? 'strict',
+            mode: c['mode'] ?? 'redact',
+        }));
+    }
+    if ((c = cfg('cost_control'))) {
+        gw = gw.use(costControl({
+            hardCapUsd: Number(c['hardCapUsd'] ?? c['hard_cap_usd']),
+            softCapUsd: (c['softCapUsd'] ?? c['soft_cap_usd']),
+            scope: c['scope'] ?? 'global',
+            window: c['window'] ?? 'day',
+        }));
+    }
+    if ((c = cfg('rate_limiter'))) {
+        gw = gw.use(rateLimiter({
+            maxRequestsPerMinute: (c['maxRequestsPerMinute'] ?? c['max_requests_per_minute']),
+            maxTokensPerMinute: (c['maxTokensPerMinute'] ?? c['max_tokens_per_minute']),
+            scope: c['scope'] ?? 'global',
+        }));
+    }
+    if ((c = cfg('model_policy'))) {
+        gw = gw.use(modelPolicy({ roles: c['roles'] ?? {} }));
+    }
+    if ((c = cfg('semantic_cache'))) {
+        const embedder = (c['enableSemantic'] ?? c['enable_semantic']) ? hashingEmbedder() : undefined;
+        gw = gw.use(semanticCache({
+            embedder,
+            similarityThreshold: Number(c['similarityThreshold'] ?? c['similarity_threshold'] ?? 0.95),
+        }));
+    }
+    if ((c = cfg('timeout'))) {
+        gw = gw.use(timeoutPolicy({ timeoutSeconds: Number(c['timeoutSeconds'] ?? c['timeout_seconds'] ?? 30) }));
+    }
+    if ((c = cfg('retry'))) {
+        gw = gw.use(retryInterceptor({
+            maxAttempts: Number(c['maxAttempts'] ?? c['max_attempts'] ?? 3),
+            baseDelayMs: Number(c['baseDelayMs'] ?? c['base_delay_ms'] ?? 500),
+        }));
+    }
+    return gw;
+}

package/dist/esm/errors.d.ts

@@ -31,6 +31,23 @@ export declare class PiiBlockedError extends GavioError {
 /** A hard budget cap was exceeded. Never swallow this — surface to user. */
 export declare class BudgetExceededError extends GavioError {
 }
+/** The circuit breaker is open; the call was rejected without hitting the provider. */
+export declare class CircuitOpenError extends ProviderUnavailableError {
+}
+/** A local rate limit (requests/tokens per minute) was exceeded. */
+export declare class RateLimitExceededError extends GavioError {
+}
+/** The caller's role is not permitted to use the requested model (RBAC). */
+export declare class ModelNotAllowedError extends GavioError {
+    readonly role: string;
+    readonly model: string;
+    constructor(role: string, model: string);
+}
 /** Output failed a guardrail validator with onFailure='error'. */
 export declare class GuardrailViolationError extends GavioError {
 }
+/** A prompt-injection attempt was detected and the guard is in block mode. */
+export declare class PromptInjectionError extends GavioError {
+    readonly patterns: string[];
+    constructor(patterns: string[]);
+}

package/dist/esm/errors.js

@@ -39,6 +39,30 @@ export class PiiBlockedError extends GavioError {
 /** A hard budget cap was exceeded. Never swallow this — surface to user. */
 export class BudgetExceededError extends GavioError {
 }
+/** The circuit breaker is open; the call was rejected without hitting the provider. */
+export class CircuitOpenError extends ProviderUnavailableError {
+}
+/** A local rate limit (requests/tokens per minute) was exceeded. */
+export class RateLimitExceededError extends GavioError {
+}
+/** The caller's role is not permitted to use the requested model (RBAC). */
+export class ModelNotAllowedError extends GavioError {
+    role;
+    model;
+    constructor(role, model) {
+        super(`role ${JSON.stringify(role)} may not use model ${JSON.stringify(model)}`);
+        this.role = role;
+        this.model = model;
+    }
+}
 /** Output failed a guardrail validator with onFailure='error'. */
 export class GuardrailViolationError extends GavioError {
 }
+/** A prompt-injection attempt was detected and the guard is in block mode. */
+export class PromptInjectionError extends GavioError {
+    patterns;
+    constructor(patterns) {
+        super(`prompt injection detected: ${patterns.join(', ')}`);
+        this.patterns = patterns;
+    }
+}

package/dist/esm/gateway.d.ts

@@ -4,7 +4,7 @@ import { PricingProvider } from './pricing.js';
 import type { ProviderAdapter } from './providers/base.js';
 import type { GavioResponse } from './response.js';
 import { Provider } from './types.js';
-import type { Message } from './types.js';
+import type { Message, PromptLineage, PromptLineageInit } from './types.js';
 export interface GatewayOptions {
     provider?: Provider | string;
     model?: string;
@@ -22,6 +22,8 @@ export interface CompleteOptions {
     metadata?: Record<string, unknown>;
     /** Provider sampling options (temperature, maxTokens, etc.). */
     options?: Record<string, unknown>;
+    /** Prompt provenance (F-OBS-04): template, variables, and RAG chunk sources. */
+    lineage?: PromptLineage | PromptLineageInit | null;
 }
 /**
  * Routes a request through the interceptor pipeline to a provider.
@@ -39,6 +41,11 @@ export declare class Gateway {
     private readonly pricing;
     private readonly interceptors;
     constructor(options?: GatewayOptions);
+    /**
+     * Build a Gateway from a config object or a JSON file path (F-DX-05).
+     * Async so the config module loads lazily (avoids a circular import).
+     */
+    static fromConfig(config: string | Record<string, unknown>): Promise<Gateway>;
     /** Register an interceptor or executor policy. First-registered = outermost. */
     use(interceptor: Interceptor): this;
     /** Supply a provider adapter explicitly (overrides `provider`). */
@@ -46,6 +53,16 @@ export declare class Gateway {
     get model(): string;
     get providerName(): string;
     complete(opts: CompleteOptions): Promise<GavioResponse>;
+    /**
+     * Stream a completion, buffering the provider stream (F-REL-06).
+     *
+     * The provider stream is buffered in full so the post-interceptor pipeline
+     * (guardrails, PII restore, audit) runs on the complete response before any
+     * chunk reaches the caller. Pre/post interceptors run via the chain; executor
+     * policies (retry, circuit breaker, cache) are not applied to the streaming
+     * path.
+     */
+    stream(opts: CompleteOptions): AsyncGenerator<string>;
     healthCheck(): Promise<boolean>;
     private buildPipeline;
     private wrapPolicy;