@deepstrike/sdk 0.2.17 → 0.2.18

package/dist/harness/harness.js

@@ -84,7 +84,6 @@ export class HarnessLoop {
     }
     async *stream(request) {
         const kernel = getKernel();
-        const pipeline = new kernel.EvalPipeline({ extractSkillOnPass: true });
         const criteria = request.criteria ?? [];
         let currentGoal = request.goal;
         let lastIterations = 0;
@@ -126,11 +125,10 @@ export class HarnessLoop {
                 }
             }
             yield { type: "supervising" };
-            const evalAction = pipeline.feedOutcome(request.goal, criteria, lastResult, attempt);
-            if (evalAction.kind !== "evaluate")
-                break;
+            // #6 (0.5.0): the eval/verdict compute is the kernel's stateless free functions (was the
+            // EvalPipeline state machine). Build the eval prompt, call the eval LLM, parse the verdict.
+            const evalMsgs = kernel.buildEvalMessages(request.goal, criteria, lastResult, attempt, true);
             let evalText = "";
-            const evalMsgs = evalAction.messages ?? [];
             const evalContext = {
                 systemText: evalMsgs.filter((m) => m.role === "system").map((m) => m.content).join("\n\n"),
                 turns: evalMsgs.filter((m) => m.role !== "system"),
@@ -139,18 +137,16 @@ export class HarnessLoop {
                 if (evt.type === "text_delta")
                     evalText += evt.delta;
             }
-            const doneAction = pipeline.feedEvalResult(evalText);
-            if (doneAction.kind !== "done")
-                break;
+            const parsed = kernel.parseVerdict(evalText);
             const verdict = {
-                passed: doneAction.passed ?? false,
-                overallScore: doneAction.overallScore ?? 0,
-                feedback: doneAction.feedback ?? "",
-                details: doneAction.details ?? [],
+                passed: parsed.passed,
+                overallScore: parsed.overallScore,
+                feedback: parsed.feedback,
+                details: parsed.details ?? [],
             };
             if (verdict.passed) {
-                if (doneAction.skillCandidate && this.skillDir) {
-                    const { name, description, whenToUse, content } = doneAction.skillCandidate;
+                if (parsed.skillCandidate && this.skillDir) {
+                    const { name, description, whenToUse, content } = parsed.skillCandidate;
                     const fm = ["---", `name: ${name}`, `description: ${description}`,
                         whenToUse ? `when_to_use: ${whenToUse}` : null, "---", ""]
                         .filter(Boolean).join("\n");
@@ -162,7 +158,6 @@ export class HarnessLoop {
             yield { type: "revising", verdict };
             currentGoal = `${request.goal}\n\n[Attempt ${attempt} feedback: ${verdict.feedback}]`;
             lastResult = "";
-            pipeline.reset();
         }
         yield { type: "max_attempts_reached" };
     }

package/dist/index.d.ts

@@ -68,7 +68,7 @@ export { SinglePassHarness, EvalLoopHarness, HarnessLoop } from "./harness/harne
 export type { HarnessRequest, HarnessOutcome, HarnessLoopOptions, QualityGate } from "./harness/harness.js";
 export type { Message, ToolCall, ToolResult, ToolSchema, ContentPart, TextPart, ImagePart, AudioPart, StreamEvent, TextDelta, ThinkingDelta, ToolCallEvent, ToolChunk, ToolDeltaEvent, ToolSuspendEvent, ToolResultEvent, DoneEvent, ErrorEvent, PermissionRequestEvent, PermissionResolvedEvent, PermissionResponse, LLMProvider, RetryConfig, TokenUsage, ProviderToolSpec, ProviderRunState, ProviderReplay, RenderedContext, ReplayabilityAssessment, } from "./types.js";
 export type { AgentCapabilityFilter, AgentIdentity, AgentIsolation, AgentRunSpec, AgentProcessChangedObservation, ContextInheritance, KernelAgentRole, LoopResult, MilestoneCheckResult, MilestoneContract, MilestonePhase, MilestonePolicy, SubAgentResult, TerminationReason, WorkflowSpec, WorkflowNodeSpec, WorkflowTaskSpec, WorkflowSpawnInfo, } from "./types/agent.js";
-export { agentIdentitySub, agentRunSpecToKernel, milestoneCheckFail, milestoneCheckPass, milestoneCheckResultToKernel, subAgentResultToKernel, workflowSpecToKernel, workflowNodeSpecToKernel, submitWorkflowNodesToKernel, submitWorkflowToKernel, submitWorkflowNodesTool, startWorkflowTool, fanoutSynthesize, generateAndFilter, verifyRules, } from "./types/agent.js";
+export { agentIdentitySub, agentRunSpecToKernel, milestoneCheckFail, milestoneCheckPass, milestoneCheckResultToKernel, subAgentResultToKernel, workflowSpecToKernel, workflowNodeSpecToKernel, submitWorkflowNodesToKernel, submitWorkflowToKernel, submitWorkflowNodesTool, startWorkflowTool, fanoutSynthesize, generateAndFilter, verifyRules, genEval, } from "./types/agent.js";
 export type { AcceptanceCriterion, VerificationContract, ContractCheckResult, } from "./collaboration/contract.js";
 export { ContractBuilder, formatContractForSystemPrompt, contractToCriteriaStrings, } from "./collaboration/contract.js";
 export { AgentPool } from "./collaboration/pool.js";

package/dist/index.js

@@ -46,7 +46,7 @@ export { PermissionManager, PermissionMode } from "./safety/permissions.js";
 export { Governance, governancePolicyToKernelEvent } from "./governance.js";
 // ── Harness ────────────────────────────────────────────────────────────────
 export { SinglePassHarness, EvalLoopHarness, HarnessLoop } from "./harness/harness.js";
-export { agentIdentitySub, agentRunSpecToKernel, milestoneCheckFail, milestoneCheckPass, milestoneCheckResultToKernel, subAgentResultToKernel, workflowSpecToKernel, workflowNodeSpecToKernel, submitWorkflowNodesToKernel, submitWorkflowToKernel, submitWorkflowNodesTool, startWorkflowTool, fanoutSynthesize, generateAndFilter, verifyRules, } from "./types/agent.js";
+export { agentIdentitySub, agentRunSpecToKernel, milestoneCheckFail, milestoneCheckPass, milestoneCheckResultToKernel, subAgentResultToKernel, workflowSpecToKernel, workflowNodeSpecToKernel, submitWorkflowNodesToKernel, submitWorkflowToKernel, submitWorkflowNodesTool, startWorkflowTool, fanoutSynthesize, generateAndFilter, verifyRules, genEval, } from "./types/agent.js";
 export { ContractBuilder, formatContractForSystemPrompt, contractToCriteriaStrings, } from "./collaboration/contract.js";
 export { AgentPool } from "./collaboration/pool.js";
 export { KERNEL_ROLE_MAP } from "./collaboration/pool.js";

package/dist/kernel.d.ts

@@ -77,13 +77,11 @@ interface NativeCriterion {
     required: boolean;
     weight?: number;
 }
-interface EvalPipelineAction {
-    kind: "evaluate" | "done";
-    messages?: Message[];
-    passed?: boolean;
-    overallScore?: number;
-    feedback?: string;
-    details?: Array<{
+export interface Verdict {
+    passed: boolean;
+    overallScore: number;
+    feedback: string;
+    details: Array<{
         criterion: string;
         passed: boolean;
         score: number;
@@ -96,11 +94,6 @@ interface EvalPipelineAction {
         content: string;
     };
 }
-interface EvalPipelineInstance {
-    feedOutcome(goal: string, criteria: NativeCriterion[], result: string, attempt: number): EvalPipelineAction;
-    feedEvalResult(content: string): EvalPipelineAction;
-    reset(): void;
-}
 interface IdlePipelineAction {
     kind: "synthesize_insights" | "commit_memories" | "noop" | "aborted";
     messages?: Message[];
@@ -156,9 +149,9 @@ interface KernelModule {
         timeoutMs?: bigint;
     }) => KernelRuntimeInstance;
     SignalRouter: new (maxQueueSize: number) => SignalRouterInstance;
-    EvalPipeline: new (options?: {
-        extractSkillOnPass?: boolean;
-    }) => EvalPipelineInstance;
+    buildEvalMessages(goal: string, criteria: NativeCriterion[], result: string, attempt: number, extractSkillOnPass: boolean): Message[];
+    parseVerdict(content: string): Verdict;
+    verdictOutputSchema(extractSkillOnPass: boolean): string;
     IdlePipeline: new (agentId: string) => IdlePipelineInstance;
 }
 export declare function getKernel(): KernelModule;

package/dist/types/agent.d.ts

@@ -225,6 +225,13 @@ export declare function workflowNodeToManifest(node: WorkflowSpawnInfo, parentSe
 export declare function fanoutSynthesize(workers: WorkflowTaskSpec[], synthesize: WorkflowTaskSpec): WorkflowSpec;
 /** N parallel Implement generators feeding a single Verify filter/dedupe step (barrier). */
 export declare function generateAndFilter(generators: WorkflowTaskSpec[], filter: WorkflowTaskSpec): WorkflowSpec;
+/**
+ * Generate→evaluate quality gate (the EvalPipeline successor, #6): a `loop` worker node (re-run up
+ * to `maxIters`, stopping early on a `loop_continue=false` self-signal) + a bias-resistant `verify`
+ * eval node gated on it, carrying the kernel's verdict `outputSchema`. Mirrors the kernel `gen_eval`
+ * template. For the iterative retry-with-feedback variant, drive it with `HarnessLoop`.
+ */
+export declare function genEval(worker: WorkflowTaskSpec, evaluate: WorkflowTaskSpec, maxIters?: number, extractSkillOnPass?: boolean): WorkflowSpec;
 /**
  * One fresh-context verifier per rule/claim (parallel) + optional skeptic that depends on all and
  * re-checks flags. Verifiers run read-only with no inherited author context (bias-resistant).

package/dist/types/agent.js

@@ -1,3 +1,4 @@
+import { getKernel } from "../kernel.js";
 /** Map kernel spawn observation → host manifest. */
 export function spawnObservationToManifest(obs, spec, parentSessionId) {
     const o = obs;
@@ -409,6 +410,34 @@ export function generateAndFilter(generators, filter) {
     });
     return { nodes };
 }
+/**
+ * Generate→evaluate quality gate (the EvalPipeline successor, #6): a `loop` worker node (re-run up
+ * to `maxIters`, stopping early on a `loop_continue=false` self-signal) + a bias-resistant `verify`
+ * eval node gated on it, carrying the kernel's verdict `outputSchema`. Mirrors the kernel `gen_eval`
+ * template. For the iterative retry-with-feedback variant, drive it with `HarnessLoop`.
+ */
+export function genEval(worker, evaluate, maxIters = 3, extractSkillOnPass = true) {
+    const schema = JSON.parse(getKernel().verdictOutputSchema(extractSkillOnPass));
+    return {
+        nodes: [
+            {
+                task: asTask(worker),
+                role: "implement",
+                isolation: "worktree",
+                contextInheritance: "full",
+                loop: { maxIters: Math.max(1, maxIters) },
+            },
+            {
+                task: asTask(evaluate),
+                role: "verify",
+                isolation: "read_only",
+                contextInheritance: "none",
+                dependsOn: [0],
+                outputSchema: schema,
+            },
+        ],
+    };
+}
 /**
  * One fresh-context verifier per rule/claim (parallel) + optional skeptic that depends on all and
  * re-checks flags. Verifiers run read-only with no inherited author context (bias-resistant).

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@deepstrike/sdk",
-  "version": "0.2.17",
+  "version": "0.2.18",
   "description": "DeepStrike Node.js SDK",
   "type": "module",
   "main": "dist/index.js",
@@ -20,7 +20,7 @@
   },
   "dependencies": {
     "@anthropic-ai/sdk": "^0.99.0",
-    "@deepstrike/core": "0.2.17",
+    "@deepstrike/core": "0.2.18",
     "@google/generative-ai": "^0.24.1",
     "openai": "^5.23.2"
   },