@tangle-network/agent-runtime 0.94.11 → 0.94.13

package/dist/backends-BO4Jqoe2.d.ts

@@ -0,0 +1,134 @@
+import { A as AgentBackendInput, O as OpenAIChatTool, a as OpenAIChatToolChoice, b as OpenAIChatResponseFormat, c as AgentExecutionBackend, d as AgentBackendContext, R as RuntimeStreamEvent } from './types-BwoZWq-i.js';
+
+/**
+ *
+ * Backend factories for `runAgentTaskStream`. Three shapes ship in core:
+ *
+ *  - `createIterableBackend` — wrap any custom async iterable into a backend
+ *  - `createSandboxPromptBackend` — sandbox / sidecar `streamPrompt` clients
+ *  - `createOpenAICompatibleBackend` — OpenAI-style chat completions endpoints
+ *
+ * Adapters stay thin: domain repos own auth, model selection, and the concrete
+ * tool surface. The factories handle session creation, stream normalization,
+ * and graceful end-of-stream signalling.
+ *
+ * @stable
+ */
+
+/** Wrap any custom async-iterable stream into a typed `AgentExecutionBackend`. @stable */
+declare function createIterableBackend<TInput extends AgentBackendInput>(options: {
+    kind: string;
+    start?: AgentExecutionBackend<TInput>['start'];
+    resume?: AgentExecutionBackend<TInput>['resume'];
+    stream: AgentExecutionBackend<TInput>['stream'];
+    stop?: AgentExecutionBackend<TInput>['stop'];
+}): AgentExecutionBackend<TInput>;
+/** Build an `AgentExecutionBackend` backed by a sandbox/sidecar `streamPrompt` call. @stable */
+declare function createSandboxPromptBackend<TBox, TInput extends AgentBackendInput = AgentBackendInput>(options: {
+    kind?: string;
+    getBox(input: TInput, context: Omit<AgentBackendContext, 'session'>): Promise<TBox> | TBox;
+    streamPrompt(box: TBox, message: string, context: AgentBackendContext): AsyncIterable<unknown>;
+    mapEvent?: (event: unknown, context: AgentBackendContext) => RuntimeStreamEvent | undefined;
+    getSessionId?: (box: TBox, input: TInput) => string | undefined;
+}): AgentExecutionBackend<TInput>;
+/** @stable */
+/**
+ * Retry policy for transient transport errors (rate limits, upstream
+ * timeouts). Defaults to 5 attempts with exponential backoff starting at
+ * 1s, ±25% jitter, capped at 30s. Set `maxAttempts: 1` to disable retries.
+ *
+ * Retried status codes:
+ *   - 408 Request Timeout
+ *   - 425 Too Early
+ *   - 429 Too Many Requests
+ *   - 500 / 502 / 503 / 504 — upstream transient failures
+ *
+ * Hard failures (401, 403, 4xx other than the above) propagate immediately.
+ */
+interface BackendRetryPolicy {
+    /** Total attempts including the first try. Default 5. */
+    maxAttempts?: number;
+    /** Initial backoff in ms before the second attempt. Default 1000. */
+    initialBackoffMs?: number;
+    /** Hard ceiling on backoff in ms. Default 30000. */
+    maxBackoffMs?: number;
+    /** Jitter fraction in [0, 1]. Default 0.25 (±25%). */
+    jitter?: number;
+    /** Status codes that trigger a retry. Default: 408, 425, 429, 500, 502, 503, 504. */
+    retryStatuses?: ReadonlyArray<number>;
+    /**
+     * Per-attempt wall-clock deadline in ms. If a single fetch attempt does
+     * not return headers within this window the attempt is aborted and
+     * retried. Default 120000 (2 min). Without this a hung upstream blocks
+     * the attempt indefinitely — observed in production as a 15-minute
+     * `fetch failed` that burned an entire eval persona. Set to 0 to disable.
+     */
+    requestTimeoutMs?: number;
+}
+/**
+ *
+ * OpenAI-compat streaming backend. Routes `runAgentTaskStream` through any
+ * `POST /chat/completions` endpoint that speaks OpenAI's SSE protocol —
+ * Tangle Router, OpenAI direct, OpenRouter, Groq, DeepSeek, Together. The
+ * router also fronts Anthropic models in Anthropic-native SSE shape; this
+ * backend handles both.
+ *
+ * ### Tool calls
+ *
+ * Pass `tools` (and optionally `toolChoice`) to forward an OpenAI Chat
+ * Completions `tools[]` array on every request. Streamed `tool_call` chunks
+ * are buffered until the model finalizes them (either `finish_reason:
+ * 'tool_calls'` for OpenAI shape or a `content_block_stop` for Anthropic
+ * `tool_use` blocks proxied through the router), then emitted as a single
+ * `tool_call` RuntimeStreamEvent with the assembled `args`.
+ *
+ * The backend does NOT execute tools — it surfaces calls for the caller's
+ * own dispatcher (typically the product's MCP / sandbox runtime) to fulfill
+ * and feed back as a subsequent `messages` turn. This keeps the transport
+ * thin and lets the agent host own tool dispatch policy.
+ *
+ * ### Fail-loud errors
+ *
+ * Non-success HTTP responses (4xx/5xx) and exhausted retry budgets throw
+ * `BackendTransportError` from inside the `stream()` generator. The runtime
+ * catches the throw, yields a `backend_error` with a typed `error` field
+ * (`kind`, `status`, truncated `body`) and a terminal `final` event with
+ * `status: 'failed'` carrying the same detail. Consumers MUST map
+ * `final.error` onto their `RunRecord.error` — silently treating an empty
+ * `finalText` as "agent produced nothing" hides credit exhaustion, auth
+ * failure, and upstream outages.
+ *
+ * @stable
+ */
+declare function createOpenAICompatibleBackend<TInput extends AgentBackendInput = AgentBackendInput>(options: {
+    apiKey: string;
+    baseUrl: string;
+    model: string;
+    kind?: string;
+    /**
+     * OpenAI Chat Completions `tools[]` definitions surfaced to the model on
+     * every request. Omit to send a tool-free request (existing behavior).
+     * The runtime makes no assumption about the dispatcher — calls stream out
+     * as `tool_call` events and the caller is responsible for executing them
+     * and feeding `tool_result` messages back on a follow-up turn.
+     */
+    tools?: ReadonlyArray<OpenAIChatTool>;
+    /**
+     * OpenAI Chat Completions `tool_choice`. Default `undefined` (request
+     * omits the field; provider falls back to its own default — typically
+     * `'auto'`).
+     */
+    toolChoice?: OpenAIChatToolChoice;
+    /**
+     * OpenAI Chat Completions `response_format`. Omit for provider default text.
+     */
+    responseFormat?: OpenAIChatResponseFormat;
+    /** OpenAI Chat Completions `temperature`. Omit for provider default. */
+    temperature?: number;
+    /** Maximum completion tokens, sent as OpenAI-compatible `max_tokens`. Omit for provider default. */
+    maxTokens?: number;
+    fetchImpl?: typeof fetch;
+    retry?: BackendRetryPolicy;
+}): AgentExecutionBackend<TInput>;
+
+export { createIterableBackend as a, createSandboxPromptBackend as b, createOpenAICompatibleBackend as c };

package/dist/candidate-execution/index.d.ts

@@ -1,7 +1,7 @@
 import { CodeSurface } from '@tangle-network/agent-eval/campaign';
 import { AgentCandidateCodeDisabled, AgentCandidateCodeNoOp, AgentCandidateGitHubRepository, AgentProfile, AgentProfileDiff, AgentCandidateProfile, AgentCandidateExecution, AgentCandidateKnowledge, AgentCandidateMemoryPolicy, AgentCandidateLineage, Sha256Digest, AgentCandidateArtifactRef, AgentCandidateWorkspaceSnapshotEvidence } from '@tangle-network/agent-interface';
-import { a9 as sealAgentCandidateBundle, a0 as PreparedAgentCandidateExecution, j as AgentCandidateExecutionClaim, l as AgentCandidateExecutionClaimStore, k as AgentCandidateExecutionClaimResult, i as AgentCandidateExecutionAttemptRef, h as AgentCandidateExecutionAttemptRecord, p as AgentCandidateExecutionLease, r as AgentCandidateExecutionPhaseResult, v as AgentCandidateExecutionTerminalResult, t as AgentCandidateExecutionStageResult, o as AgentCandidateExecutionFinishResult, s as AgentCandidateExecutionRecoveryEvidence, M as AgentCandidateOutputArtifactPort, N as AgentCandidateOutputPurpose, L as AgentCandidateModelPort, R as AgentCandidateProtectedModelReservation, O as AgentCandidateProtectedModelActivation, S as AgentCandidateProtectedModelSettlement, z as AgentCandidateExecutorPort, b as AgentCandidateExecutionPorts, W as AgentCandidateVerificationPorts, a5 as VerifiedAgentCandidate, G as AgentCandidateExecutorWorkspaceFile, X as AgentCandidateWorkspacePort } from '../prepare-DIeVgQG3.js';
-export { d as AgentCandidateArtifactPort, e as AgentCandidateBenchmarkGraderIdentity, f as AgentCandidateBenchmarkGraderPort, A as AgentCandidateBundleInput, g as AgentCandidateContainerPort, m as AgentCandidateExecutionCleanupHandles, n as AgentCandidateExecutionFailureClass, q as AgentCandidateExecutionPhase, u as AgentCandidateExecutionTerminalRecord, w as AgentCandidateExecutionUsage, x as AgentCandidateExecutorFinalCapture, y as AgentCandidateExecutorMemoryCapture, B as AgentCandidateExecutorProfileFile, C as AgentCandidateExecutorRequest, D as AgentCandidateExecutorStopRequest, F as AgentCandidateExecutorTaskOutcomeCapture, H as AgentCandidateExecutorWorkspaceInput, I as AgentCandidateMemoryPort, J as AgentCandidateMemoryResetResult, K as AgentCandidateModelLimits, Q as AgentCandidateProtectedModelCall, T as AgentCandidateProtectedRunCapture, U as AgentCandidateRepositoryPort, V as AgentCandidateRetryRejection, c as AgentCandidateRunFinalization, a as AgentCandidateTaskExecution, Y as CANDIDATE_TRACE_ENV, Z as CANDIDATE_TRACE_TAGS, _ as CanonicalCandidateDocument, E as ExecutePreparedAgentCandidateOptions, $ as InMemoryAgentCandidateExecutionClaimStore, P as PrepareAgentCandidateExecutionOptions, a1 as PreparedAgentCandidateInstruction, a2 as PreparedAgentCandidateLaunch, a3 as PreparedAgentCandidateTrace, a4 as ResolvedAgentCandidateContainer, a6 as VerifiedAgentCandidateTaskOutcome, a7 as executePreparedAgentCandidate, a8 as prepareAgentCandidateExecution } from '../prepare-DIeVgQG3.js';
+import { aa as sealAgentCandidateBundle, a0 as PreparedAgentCandidateExecution, j as AgentCandidateExecutionClaim, l as AgentCandidateExecutionClaimStore, k as AgentCandidateExecutionClaimResult, i as AgentCandidateExecutionAttemptRef, h as AgentCandidateExecutionAttemptRecord, p as AgentCandidateExecutionLease, r as AgentCandidateExecutionPhaseResult, v as AgentCandidateExecutionTerminalResult, t as AgentCandidateExecutionStageResult, o as AgentCandidateExecutionFinishResult, s as AgentCandidateExecutionRecoveryEvidence, M as AgentCandidateOutputArtifactPort, N as AgentCandidateOutputPurpose, L as AgentCandidateModelPort, R as AgentCandidateProtectedModelReservation, O as AgentCandidateProtectedModelActivation, S as AgentCandidateProtectedModelSettlement, z as AgentCandidateExecutorPort, b as AgentCandidateExecutionPorts, W as AgentCandidateVerificationPorts, a6 as VerifiedAgentCandidate, G as AgentCandidateExecutorWorkspaceFile, X as AgentCandidateWorkspacePort } from '../prepare-CtdtsFNG.js';
+export { d as AgentCandidateArtifactPort, e as AgentCandidateBenchmarkGraderIdentity, f as AgentCandidateBenchmarkGraderPort, A as AgentCandidateBundleInput, g as AgentCandidateContainerPort, m as AgentCandidateExecutionCleanupHandles, n as AgentCandidateExecutionFailureClass, q as AgentCandidateExecutionPhase, u as AgentCandidateExecutionTerminalRecord, w as AgentCandidateExecutionUsage, x as AgentCandidateExecutorFinalCapture, y as AgentCandidateExecutorMemoryCapture, B as AgentCandidateExecutorProfileFile, C as AgentCandidateExecutorRequest, D as AgentCandidateExecutorStopRequest, F as AgentCandidateExecutorTaskOutcomeCapture, H as AgentCandidateExecutorWorkspaceInput, I as AgentCandidateMemoryPort, J as AgentCandidateMemoryResetResult, K as AgentCandidateModelLimits, Q as AgentCandidateProtectedModelCall, T as AgentCandidateProtectedRunCapture, U as AgentCandidateRepositoryPort, V as AgentCandidateRetryRejection, c as AgentCandidateRunFinalization, a as AgentCandidateTaskExecution, Y as CANDIDATE_TRACE_ENV, Z as CANDIDATE_TRACE_TAGS, _ as CanonicalCandidateDocument, E as ExecutePreparedAgentCandidateOptions, $ as InMemoryAgentCandidateExecutionClaimStore, P as PrepareAgentCandidateExecutionOptions, a1 as PreparedAgentCandidateInstruction, a2 as PreparedAgentCandidateKnowledge, a3 as PreparedAgentCandidateLaunch, a4 as PreparedAgentCandidateTrace, a5 as ResolvedAgentCandidateContainer, a7 as VerifiedAgentCandidateTaskOutcome, a8 as executePreparedAgentCandidate, a9 as prepareAgentCandidateExecution } from '../prepare-CtdtsFNG.js';
 import { TraceStore } from '@tangle-network/agent-eval';
 
 /** A complete profile that can be frozen without losing behavior. */

package/dist/candidate-execution/index.js

@@ -7,7 +7,7 @@ import {
   createProtectedAgentCandidateModelPort,
   disposePreparedAgentCandidateExecution,
   recoverExpiredAgentCandidateExecution
-} from "../chunk-FTXM7Q7P.js";
+} from "../chunk-TVJQAYQM.js";
 import {
   CANDIDATE_TRACE_ENV,
   CANDIDATE_TRACE_TAGS,
@@ -21,7 +21,7 @@ import {
   prepareAgentCandidateExecution,
   sealAgentCandidateBundle,
   verifyAgentCandidateBundle
-} from "../chunk-37V4KQDC.js";
+} from "../chunk-CNH7DF7Z.js";
 import "../chunk-WIPGQ4GT.js";
 import "../chunk-3MDZX7YU.js";
 import "../chunk-DGUM43GV.js";