@deepstrike/sdk 0.2.28 → 0.2.31

package/README.md

@@ -76,9 +76,25 @@ const reply = await collectText(runner.run({ sessionId: "chat-1", goal: "What is
 
 Use `InMemorySessionLog` for process-local sessions or `FileSessionLog` when replay should survive restarts. `wake(sessionId)` resumes from the event log without inserting a duplicate `run_started` event.
 
+### Package layout (v0.2.30)
+
+The root export is the **intent layer** — what you reach for to run an agent, run a workflow, author a tool, or pick a provider (~30 symbols). Advanced machinery lives behind subpaths, so the common surface stays small and tree-shakeable:
+
+| Import | Contains |
+|--------|----------|
+| `@deepstrike/sdk` | `runAgent` · `runFanout` · `RuntimeRunner` · `tool` · `LocalExecutionPlane` · `InMemorySessionLog`/`FileSessionLog` · `AnthropicProvider`/`OpenAIProvider`/`OpenAIResponsesProvider` · `createProvider` · `Governance` · `AgentPool` · core types |
+| `@deepstrike/sdk/providers` | backend factories (`deepseek`, `kimi`, `qwen`, `glm`, `minimax`, `gemini`, `ollama`), profiles, `CircuitBreaker` |
+| `@deepstrike/sdk/workflow` | `SubAgentOrchestrator`, `spawnStandalone`, reducers, contracts, handoff/modes, agent + spec types |
+| `@deepstrike/sdk/planes` | `WorktreeExecutionPlane`, `ProcessSandboxPlane`, `McpProxyPlane`, `RemoteVpcPlane`, archive/credential stores |
+| `@deepstrike/sdk/memory` | `DreamStore`, `WorkingMemory`, `InMemoryDreamStore`, `KnowledgeSource` |
+| `@deepstrike/sdk/harness` | `SinglePassHarness`, `EvalLoopHarness`, `HarnessLoop`, `judge` |
+| `@deepstrike/sdk/os` | profiles, `KernelPrimitivesDashboard`, signals, `PermissionManager`, replay-testing utilities |
+
+> **Migration from 0.2.x:** the kernel-lowering converters (`*ToKernel`), low-level prompt/eval builders, and the `OpenAIChatProvider` alias are no longer exported from root; backend providers, planes, memory, harness, and OS utilities moved to the subpaths above. See [`MIGRATION-v0.2.300.md`](./MIGRATION-v0.2.300.md).
+
 ### Recipes — the canonical entry points
 
-The package exports a large surface, but most apps need one of three shapes. Start with the facades and drop down to `RuntimeRunner` only when you need streaming, signals, memory, or governance hooks.
+Most apps need one of three shapes. Start with the facades and drop down to `RuntimeRunner` only when you need streaming, signals, memory, or governance hooks.
 
 ```typescript
 import { runAgent, runFanout } from "@deepstrike/sdk"
@@ -232,36 +248,44 @@ A node's `kind` selects the control-flow shape; the same executor drives them al
 
 ## Providers
 
-| Class | Backend | Notes |
-|-------|---------|-------|
-| `OpenAIChatProvider` | OpenAI Chat Completions API | SSE tool-call accumulation |
-| `OpenAIProvider` | OpenAI Chat Completions API | Compatibility alias for `OpenAIChatProvider` |
-| `OpenAIResponsesProvider` | OpenAI Responses API | Native `previous_response_id` continuation |
-| `AnthropicProvider` | Anthropic API | Native SSE, `ThinkingDelta` support |
-| `QwenProvider` | DashScope | `enable_thinking` via extensions |
-| `DeepSeekProvider` | DeepSeek API | V4 thinking controls + reasoning replay across tool turns |
-| `MiniMaxProvider` | MiniMax API | Anthropic-compatible M2.7/M2.5 path |
-| `OllamaProvider` | Local Ollama | `http://localhost:11434` default |
-| `KimiProvider` | Moonshot API | K2.6 default; K2.5 also supported |
+The root package exports the three base providers — `AnthropicProvider`, `OpenAIProvider`,
+`OpenAIResponsesProvider` — plus `createProvider`. **Every other backend is a factory function** in
+`@deepstrike/sdk/providers`: one per backend, with a `protocol` option where a backend speaks both the
+OpenAI- and Anthropic-compatible wire.
 
-All providers accept `RetryConfig` for exponential backoff and share a `CircuitBreaker`.
+```typescript
+import { deepseek, kimi, minimax } from "@deepstrike/sdk/providers"
 
-`extensions` are forwarded by every provider in both `complete()` and `stream()` while SDK-owned structural fields such as `model`, `messages`, `tools`, and streaming flags remain protected.
+const ds = deepseek({ apiKey })                          // OpenAI-compatible wire (default)
+const dsA = deepseek({ apiKey, protocol: "anthropic" })  // Anthropic-compatible wire
+const mm = minimax({ apiKey })                           // MiniMax defaults to the Anthropic wire
+```
+
+| Entry | Import from | Backend |
+|-------|-------------|---------|
+| `OpenAIProvider` | root | OpenAI Chat Completions (and any OpenAI-compatible `/v1`) |
+| `OpenAIResponsesProvider` | root | OpenAI Responses API (`previous_response_id` continuation) |
+| `AnthropicProvider` | root | Anthropic Messages API (`ThinkingDelta` support) |
+| `deepseek` · `kimi` · `qwen` · `glm` · `minimax` · `gemini` · `ollama` | `@deepstrike/sdk/providers` | the respective vendor (factory functions) |
 
-**Custom OpenAI-compatible endpoint** (MiMo, DeepSeek, Kimi, Qwen, GLM via their `/v1` base URL): use `OpenAIChatProvider` (alias `OpenAIProvider`) and pass the base URL as the **4th** argument — the 3rd is `RetryConfig`:
+Providers take an **options object** and share a `CircuitBreaker`. `extensions` are forwarded in both
+`complete()` and `stream()`; SDK-owned fields (`model`, `messages`, `tools`, streaming flags) stay protected.
+
+**Custom OpenAI-compatible endpoint** (MiMo, DeepSeek, Kimi, Qwen, GLM via their `/v1` base URL): construct
+`OpenAIProvider` with an options object — no more positional `baseURL` hole:
 
 ```typescript
 import { OpenAIProvider } from "@deepstrike/sdk"
 
-const provider = new OpenAIProvider(
+const provider = new OpenAIProvider({
   apiKey,
-  "mimo-v2.5-pro",
-  { maxRetries: 3, baseDelay: 1000 }, // RetryConfig (pass undefined for defaults)
-  "https://token-plan-cn.xiaomimimo.com/v1",
-)
+  model: "mimo-v2.5-pro",
+  baseURL: "https://token-plan-cn.xiaomimimo.com/v1",
+})
 ```
 
-Prefer a dedicated `*Provider` (e.g. `DeepSeekProvider`, `KimiProvider`) when one exists for your backend — they default the base URL and add backend-specific reasoning handling. OpenAI itself can also be selected through the provider catalog:
+Prefer a dedicated backend class from `@deepstrike/sdk/providers` when one exists — they default the base
+URL and add backend-specific reasoning handling. Any model can also be selected through the catalog: `createProvider` picks the protocol/endpoint for you:
 
 ```typescript
 import { createProvider } from "@deepstrike/sdk"

package/dist/harness/harness.d.ts

@@ -126,10 +126,11 @@ export interface HarnessLoopOptions {
 }
 export declare class HarnessLoop {
     private runner;
-    private evalProvider;
     private maxAttempts;
     private skillDir?;
-    private verdictFn?;
+    /** How each attempt is judged. Hybrid (host verdictFn → LLM eval) when a verdictFn is supplied,
+     *  otherwise the built-in LLM eval. The loop just calls `this.judge.judge(...)`. */
+    private judge;
     constructor(runner: RuntimeRunner, evalProvider: import("../types.js").LLMProvider, options?: HarnessLoopOptions);
     run(request: HarnessRequest): Promise<HarnessOutcome>;
     stream(request: HarnessRequest): AsyncIterable<HarnessEvent>;

package/dist/harness/harness.js

@@ -1,7 +1,7 @@
 import { collectText } from "../runtime/runner.js";
 import { writeFile } from "fs/promises";
 import path from "path";
-import { getKernel } from "../kernel.js";
+import { HybridJudge, VerdictFnJudge, LlmEvalJudge } from "./judge.js";
 async function runOnce(runner, req) {
     let text = "";
     let done;
@@ -58,16 +58,19 @@ export class EvalLoopHarness {
 }
 export class HarnessLoop {
     runner;
-    evalProvider;
     maxAttempts;
     skillDir;
-    verdictFn;
+    /** How each attempt is judged. Hybrid (host verdictFn → LLM eval) when a verdictFn is supplied,
+     *  otherwise the built-in LLM eval. The loop just calls `this.judge.judge(...)`. */
+    judge;
     constructor(runner, evalProvider, options = {}) {
         this.runner = runner;
-        this.evalProvider = evalProvider;
         this.maxAttempts = options.maxAttempts ?? 3;
         this.skillDir = options.skillDir;
-        this.verdictFn = options.verdictFn;
+        const llmJudge = new LlmEvalJudge(evalProvider);
+        this.judge = options.verdictFn
+            ? new HybridJudge(new VerdictFnJudge(options.verdictFn), llmJudge)
+            : llmJudge;
     }
     async run(request) {
         let last;
@@ -93,7 +96,6 @@ export class HarnessLoop {
         };
     }
     async *stream(request) {
-        const kernel = getKernel();
         const criteria = request.criteria ?? [];
         let currentGoal = request.goal;
         let lastIterations = 0;
@@ -135,35 +137,13 @@ export class HarnessLoop {
                 }
             }
             yield { type: "supervising" };
-            // I3.2 (A2/A3): host-supplied `verdictFn` short-circuits the LLM eval. When it returns a
-            // Verdict, use it; when it returns undefined, defer to the built-in eval (hybrid path).
-            let verdict;
-            let skillCandidate;
-            if (this.verdictFn) {
-                verdict = await this.verdictFn({ goal: request.goal, criteria, attempt, result: lastResult });
-            }
-            if (!verdict) {
-                // #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 evalContext = {
-                    systemText: evalMsgs.filter((m) => m.role === "system").map((m) => m.content).join("\n\n"),
-                    turns: evalMsgs.filter((m) => m.role !== "system"),
-                };
-                for await (const evt of this.evalProvider.stream(evalContext, [], undefined)) {
-                    if (evt.type === "text_delta")
-                        evalText += evt.delta;
-                }
-                const parsed = kernel.parseVerdict(evalText);
-                verdict = {
-                    passed: parsed.passed,
-                    overallScore: parsed.overallScore,
-                    feedback: parsed.feedback,
-                    details: parsed.details ?? [],
-                };
-                skillCandidate = parsed.skillCandidate;
-            }
+            // I3.2 (A2/A3): the judge Strategy encapsulates "host verdictFn short-circuit → built-in LLM
+            // eval". HarnessLoop's judge always terminates in LlmEvalJudge, so a verdict is guaranteed.
+            const judged = await this.judge.judge({ goal: request.goal, criteria, attempt, result: lastResult });
+            const verdict = judged?.verdict;
+            const skillCandidate = judged?.skillCandidate;
+            if (!verdict)
+                throw new Error("HarnessLoop: judge produced no verdict");
             if (verdict.passed) {
                 if (skillCandidate && this.skillDir) {
                     const { name, description, whenToUse, content } = skillCandidate;

package/dist/harness/judge.d.ts

@@ -0,0 +1,42 @@
+import type { LLMProvider } from "../types.js";
+import { getKernel } from "../kernel.js";
+import type { Criterion, Verdict, VerdictFn } from "./harness.js";
+export type SkillCandidate = ReturnType<ReturnType<typeof getKernel>["parseVerdict"]>["skillCandidate"];
+export interface JudgeContext {
+    goal: string;
+    criteria: Criterion[];
+    attempt: number;
+    result: string;
+}
+export interface JudgeResult {
+    verdict: Verdict;
+    /** Skill the judge proposes extracting on pass (LLM-eval path only). */
+    skillCandidate?: SkillCandidate;
+}
+/** Decides whether one attempt's result meets the criteria. Returning `undefined` defers to a
+ *  fallback judge (enables hybrid host/LLM judgment). */
+export interface AttemptJudge {
+    judge(ctx: JudgeContext): Promise<JudgeResult | undefined>;
+}
+/** Wraps a host-supplied `VerdictFn`. Returns `undefined` (defer) when the function does. */
+export declare class VerdictFnJudge implements AttemptJudge {
+    private readonly fn;
+    constructor(fn: VerdictFn);
+    judge(ctx: JudgeContext): Promise<JudgeResult | undefined>;
+}
+/** The built-in LLM eval: render the kernel eval prompt, stream the eval provider, parse the
+ *  verdict. Always produces a JudgeResult (never defers). */
+export declare class LlmEvalJudge implements AttemptJudge {
+    private readonly evalProvider;
+    private readonly extractSkillOnPass;
+    constructor(evalProvider: LLMProvider, extractSkillOnPass?: boolean);
+    judge(ctx: JudgeContext): Promise<JudgeResult>;
+}
+/** Try `primary`; if it defers (`undefined`), use `fallback`. Models HarnessLoop's
+ *  "verdictFn short-circuits, else built-in LLM eval" hybrid judgment. */
+export declare class HybridJudge implements AttemptJudge {
+    private readonly primary;
+    private readonly fallback;
+    constructor(primary: AttemptJudge, fallback: AttemptJudge);
+    judge(ctx: JudgeContext): Promise<JudgeResult | undefined>;
+}

package/dist/harness/judge.js

@@ -0,0 +1,58 @@
+import { getKernel } from "../kernel.js";
+/** Wraps a host-supplied `VerdictFn`. Returns `undefined` (defer) when the function does. */
+export class VerdictFnJudge {
+    fn;
+    constructor(fn) {
+        this.fn = fn;
+    }
+    async judge(ctx) {
+        const verdict = await this.fn({ goal: ctx.goal, criteria: ctx.criteria, attempt: ctx.attempt, result: ctx.result });
+        return verdict ? { verdict } : undefined;
+    }
+}
+/** The built-in LLM eval: render the kernel eval prompt, stream the eval provider, parse the
+ *  verdict. Always produces a JudgeResult (never defers). */
+export class LlmEvalJudge {
+    evalProvider;
+    extractSkillOnPass;
+    constructor(evalProvider, extractSkillOnPass = true) {
+        this.evalProvider = evalProvider;
+        this.extractSkillOnPass = extractSkillOnPass;
+    }
+    async judge(ctx) {
+        const kernel = getKernel();
+        const evalMsgs = kernel.buildEvalMessages(ctx.goal, ctx.criteria, ctx.result, ctx.attempt, this.extractSkillOnPass);
+        const evalContext = {
+            systemText: evalMsgs.filter((m) => m.role === "system").map((m) => m.content).join("\n\n"),
+            turns: evalMsgs.filter((m) => m.role !== "system"),
+        };
+        let evalText = "";
+        for await (const evt of this.evalProvider.stream(evalContext, [], undefined)) {
+            if (evt.type === "text_delta")
+                evalText += evt.delta;
+        }
+        const parsed = kernel.parseVerdict(evalText);
+        return {
+            verdict: {
+                passed: parsed.passed,
+                overallScore: parsed.overallScore,
+                feedback: parsed.feedback,
+                details: parsed.details ?? [],
+            },
+            skillCandidate: parsed.skillCandidate,
+        };
+    }
+}
+/** Try `primary`; if it defers (`undefined`), use `fallback`. Models HarnessLoop's
+ *  "verdictFn short-circuits, else built-in LLM eval" hybrid judgment. */
+export class HybridJudge {
+    primary;
+    fallback;
+    constructor(primary, fallback) {
+        this.primary = primary;
+        this.fallback = fallback;
+    }
+    async judge(ctx) {
+        return (await this.primary.judge(ctx)) ?? (await this.fallback.judge(ctx));
+    }
+}

package/dist/harness/public.d.ts

@@ -0,0 +1,4 @@
+export { SinglePassHarness, EvalLoopHarness, HarnessLoop } from "./harness.js";
+export type { HarnessRequest, HarnessOutcome, HarnessLoopOptions, QualityGate, CriterionResult, HarnessEvent, VerdictFn, } from "./harness.js";
+export { judge } from "../runtime/eval.js";
+export type { Criterion, Verdict, VerdictDetail, JudgeArgs } from "../runtime/eval.js";

package/dist/harness/public.js

@@ -0,0 +1,3 @@
+// `@deepstrike/sdk/harness` — the evaluation framework: single-pass / eval-loop harnesses and the judge.
+export { SinglePassHarness, EvalLoopHarness, HarnessLoop } from "./harness.js";
+export { judge } from "../runtime/eval.js";

package/dist/index.d.ts

@@ -1,92 +1,33 @@
 export { runAgent, runFanout } from "./runtime/facade.js";
 export type { RunAgentOptions, RunFanoutOptions } from "./runtime/facade.js";
 export { RuntimeRunner, collectText } from "./runtime/runner.js";
-export type { RuntimeOptions, SchedulerBudget } from "./runtime/runner.js";
-export { builtinReducers, resolveReducer } from "./runtime/reducers.js";
-export type { Reducer, ReducerRegistry, ReducerInput } from "./runtime/reducers.js";
-export { loopInstruction, classifyInstruction, judgeGoal, extractLoopContinue, extractClassifyBranch, extractJudgeWinner, } from "./runtime/workflow-control-flow.js";
-export { WorktreeExecutionPlane, GitWorktreeManager } from "./runtime/worktree-plane.js";
-export type { WorktreeManager } from "./runtime/worktree-plane.js";
-export { FileWorkflowStore } from "./runtime/workflow-store.js";
-export type { MemoryPolicy, MemoryWriteRateLimit, ResourceQuota } from "./kernel.js";
-export { KernelPrimitivesDashboard } from "./runtime/kernel-primitives-dashboard.js";
-export { FilteredExecutionPlane } from "./runtime/filtered-plane.js";
-export { SubAgentOrchestrator, defaultSubAgentOrchestrator, spawnStandalone } from "./runtime/sub-agent-orchestrator.js";
-export type { SubAgentRunContext } from "./runtime/sub-agent-orchestrator.js";
+export type { RuntimeOptions } from "./runtime/runner.js";
 export { LocalExecutionPlane } from "./runtime/execution-plane.js";
 export type { ExecutionPlane, RunContext } from "./runtime/execution-plane.js";
 export { InMemorySessionLog, FileSessionLog } from "./runtime/session-log.js";
 export type { SessionLog, SessionEvent } from "./runtime/session-log.js";
-export { ReplayProvider } from "./runtime/replay-provider.js";
-export type { ReplayProviderOpts } from "./runtime/replay-provider.js";
-export { extractRecordedMessages } from "./runtime/replay-fixture.js";
-export { judge, buildEvalMessages, parseVerdict, verdictOutputSchema } from "./runtime/eval.js";
-export type { Criterion, Verdict, VerdictDetail, JudgeArgs } from "./runtime/eval.js";
-export { DEFAULT_NATIVE_ATTENTION_POLICY, DEFAULT_NATIVE_GOVERNANCE_POLICY, assertNativeProfile, osProfile, } from "./runtime/os-profile.js";
-export type { NativeOsProfile, OsProfileId } from "./runtime/os-profile.js";
-export { rebuildOsSnapshotFromSessionEvents, sessionLogHasRequiredCategories, } from "./runtime/os-snapshot.js";
-export type { OsSnapshot } from "./runtime/os-snapshot.js";
-export { categoryForKind, kernelObservationToSessionEvent } from "./runtime/kernel-event-log.js";
-export type { KernelEventCategory } from "./runtime/kernel-event-log.js";
-export { NullArchiveStore, FileArchiveStore } from "./runtime/archive.js";
-export type { ArchiveStore } from "./runtime/archive.js";
-export { EnvCredentialVault, InMemoryCredentialVault, ChainedCredentialVault } from "./runtime/credential-vault.js";
-export type { CredentialVault } from "./runtime/credential-vault.js";
-export { ProcessSandboxPlane } from "./runtime/process-sandbox-plane.js";
-export type { SandboxOptions } from "./runtime/process-sandbox-plane.js";
-export { McpProxyPlane } from "./runtime/mcp-proxy-plane.js";
-export type { McpServerConfig } from "./runtime/mcp-proxy-plane.js";
-export { RemoteVpcPlane } from "./runtime/remote-vpc-plane.js";
-export type { RemoteVpcOptions } from "./runtime/remote-vpc-plane.js";
-export { AnthropicProvider } from "./providers/anthropic.js";
-export { OpenAIChatProvider, OpenAIProvider } from "./providers/openai.js";
-export { DeepSeekProvider, DeepSeekAnthropicProvider } from "./providers/deepseek.js";
-export { KimiProvider, KimiAnthropicProvider } from "./providers/kimi.js";
-export { QwenProvider, QwenAnthropicProvider } from "./providers/qwen.js";
-export { GLMProvider, GLMAnthropicProvider } from "./providers/glm.js";
-export { GeminiProvider } from "./providers/gemini.js";
-export { MiniMaxAnthropicProvider, MiniMaxOpenAIProvider } from "./providers/minimax.js";
-export { OllamaProvider } from "./providers/ollama.js";
-export { CircuitBreaker, normalizeToolCall } from "./providers/base.js";
-export { OpenAIChatAdapter } from "./providers/openai-chat.js";
-export { OpenAIResponsesAdapter, OpenAIResponsesProvider } from "./providers/openai-responses.js";
-export type { OpenAIResponsesRunState } from "./providers/openai-responses.js";
-export { endpointProfiles, modelProfiles, getModelProfile } from "./providers/profiles.js";
-export type { ModelProfileId, ProviderId } from "./providers/profiles.js";
-export { createProvider } from "./providers/catalog.js";
-export type { CreateProviderOptions, EndpointProfileId } from "./providers/catalog.js";
-export { ProviderReplayValidationError, DEGRADED_REASONING_PLACEHOLDER } from "./providers/replay-validator.js";
-export { assessProviderReplayability, peekProviderReplay, seedProviderReplayFromEvents, isReplayCompatibleWithProvider, } from "./runtime/provider-replay.js";
-export { tool, streamingTool, executeTools, readFile, validateToolArguments } from "./tools/index.js";
+export { InMemoryGroupBudgetStore, SessionLogGroupBudgetStore } from "./runtime/run-group.js";
+export type { RunGroup, GroupBudgetStore, GroupLedger, GroupCharge, GroupMember } from "./runtime/run-group.js";
+export { InMemoryEventStream, isVisibleTo } from "./runtime/event-stream.js";
+export type { EventStream, BlackboardEvent, EventViewer } from "./runtime/event-stream.js";
+export { reactByMention, directorDriven, roundRobin, firstNonEmpty, union } from "./runtime/turn-policy.js";
+export type { TurnPolicy, PeerView } from "./runtime/turn-policy.js";
+export { ReactiveSession, readRecentTool } from "./runtime/reactive-session.js";
+export type { ReactiveSessionOptions, ReactivePeerSpec, EmitEvent, Reaction, ReactorTurn, ReactorContext } from "./runtime/reactive-session.js";
+export { tool, streamingTool } from "./tools/index.js";
 export type { RegisteredTool, ToolExecContext } from "./tools/index.js";
 export { safeTool, ok, fail, ToolError, formatToolError } from "./tools/errors.js";
 export type { ToolEnvelope, ToolEnvelopeOk, ToolEnvelopeFail } from "./tools/errors.js";
-export { scanSkillDir, readSkillFile } from "./skills/loader.js";
-export type { SkillMetadata } from "./skills/loader.js";
-export { WorkingMemory } from "./memory/working.js";
-export { InMemoryDreamStore } from "./memory/in-memory-store.js";
-export type { DreamStore, DreamResult, SessionData, SessionMessage, MemoryEntry, CurationResult, CurationStats, MemoryWriteRequest, MemoryQuery, MemoryRetrieval, MemoryMetadata, MemoryKind, } from "./memory/protocols.js";
-export type { KnowledgeSource } from "./knowledge/source.js";
-export { ScheduledPrompt } from "./signals/scheduled.js";
-export { SignalGateway } from "./signals/gateway.js";
-export type { RuntimeSignal, SignalSource } from "./signals/types.js";
-export { PermissionManager, PermissionMode } from "./safety/permissions.js";
-export type { PermissionDecision, Permission } from "./safety/permissions.js";
-export { Governance, governancePolicyToKernelEvent } from "./governance.js";
+export { AnthropicProvider } from "./providers/anthropic.js";
+export type { AnthropicProviderConfig } from "./providers/anthropic.js";
+export { OpenAIProvider } from "./providers/openai.js";
+export type { OpenAIProviderOptions } from "./providers/openai.js";
+export { OpenAIResponsesProvider } from "./providers/openai-responses.js";
+export { createProvider } from "./providers/catalog.js";
+export type { CreateProviderOptions, EndpointProfileId } from "./providers/catalog.js";
+export { Governance } from "./governance.js";
 export type { GovernanceVerdict, GovernancePolicy, GovernanceConstraint } from "./governance.js";
-export { SinglePassHarness, EvalLoopHarness, HarnessLoop } from "./harness/harness.js";
-export type { HarnessRequest, HarnessOutcome, HarnessLoopOptions, QualityGate, CriterionResult, HarnessEvent, VerdictFn } from "./harness/harness.js";
-export type { Message, ToolCall, ToolResult, ToolSchema, ContentPart, TextPart, ImagePart, AudioPart, StreamEvent, TextDelta, ThinkingDelta, ToolCallEvent, ToolChunk, ToolDeltaEvent, ToolSuspendEvent, ToolResultEvent, ToolAuditFailedEvent, DoneEvent, ErrorEvent, PermissionRequestEvent, PermissionResolvedEvent, PermissionResponse, LLMProvider, RetryConfig, TokenUsage, ProviderToolSpec, ProviderRunState, ProviderReplay, RenderedContext, ReplayabilityAssessment, CacheBreakpointStrategy, } 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, 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";
-export type { AgentRole, IsolatedVerifierContext, CoordinatorConfig } from "./collaboration/pool.js";
-export { KERNEL_ROLE_MAP } from "./collaboration/pool.js";
-export { ContractDrivenHarness } from "./collaboration/harness.js";
-export type { ContractOutcome, ContractHarnessOptions, Violation } from "./collaboration/harness.js";
-export { HandoffBus } from "./collaboration/handoff.js";
-export type { HandoffArtifact, ContractOutcomeInput } from "./collaboration/handoff.js";
-export { CreatorVerifierMode, OrchestrationMode } from "./collaboration/modes/creator-verifier.js";
-export type { CreatorVerifierMetrics } from "./collaboration/modes/creator-verifier.js";
+export type { RuntimeSignal, SignalSource } from "./signals/types.js";
+export type { Message, ToolCall, ToolResult, ToolSchema, ContentPart, TextPart, ImagePart, AudioPart, StreamEvent, TextDelta, ThinkingDelta, ToolCallEvent, ToolChunk, ToolDeltaEvent, ToolSuspendEvent, ToolResultEvent, ToolAuditFailedEvent, DoneEvent, ErrorEvent, PermissionRequestEvent, PermissionResolvedEvent, PermissionResponse, LLMProvider, RetryConfig, TokenUsage, } from "./types.js";
+export type { WorkflowSpec, WorkflowNodeSpec, } from "./types/agent.js";

package/dist/index.js

@@ -1,69 +1,37 @@
 // ╔══════════════════════════════════════════════════════════════════════════╗
-// ║ START HERE — the canonical entry points for the common cases.              ║
-// ║   runAgent   → one prompt, one model, the text back.                       ║
-// ║   runFanout  → run N tasks in parallel, then synthesize (kernel-gated DAG).║
-// ║   RuntimeRunner → drop down to this for streaming, tools, signals, memory, ║
-// ║                   governance, and the standalone `runWorkflow` driver.     ║
-// ║ Everything below the providers block is advanced / opt-in surface.         ║
+// ║ @deepstrike/sdk — root surface (v0.2.30).                                      ║
+// ║                                                                            ║
+// ║ This is the intent layer: run an agent, run a workflow, author a tool,     ║
+// ║ pick a provider. Advanced machinery lives behind subpaths:                 ║
+// ║   @deepstrike/sdk/providers  — backend provider classes + profiles         ║
+// ║   @deepstrike/sdk/workflow   — orchestration, reducers, contracts, specs   ║
+// ║   @deepstrike/sdk/planes     — worktree / sandbox / mcp / vpc planes        ║
+// ║   @deepstrike/sdk/memory     — dream + working memory, knowledge sources    ║
+// ║   @deepstrike/sdk/harness    — eval harnesses + judge                       ║
+// ║   @deepstrike/sdk/os         — profiles, diagnostics, signals, replay tests ║
 // ╚══════════════════════════════════════════════════════════════════════════╝
+// ── Start here: the canonical entry points ─────────────────────────────────
 export { runAgent, runFanout } from "./runtime/facade.js";
-// ── Runtime (Layer 1.5) ────────────────────────────────────────────────────
 export { RuntimeRunner, collectText } from "./runtime/runner.js";
-export { builtinReducers, resolveReducer } from "./runtime/reducers.js";
-export { loopInstruction, classifyInstruction, judgeGoal, extractLoopContinue, extractClassifyBranch, extractJudgeWinner, } from "./runtime/workflow-control-flow.js";
-export { WorktreeExecutionPlane, GitWorktreeManager } from "./runtime/worktree-plane.js";
-export { FileWorkflowStore } from "./runtime/workflow-store.js";
-export { KernelPrimitivesDashboard } from "./runtime/kernel-primitives-dashboard.js";
-export { FilteredExecutionPlane } from "./runtime/filtered-plane.js";
-export { SubAgentOrchestrator, defaultSubAgentOrchestrator, spawnStandalone } from "./runtime/sub-agent-orchestrator.js";
+// ── Execution plane + session log (the defaults) ────────────────────────────
 export { LocalExecutionPlane } from "./runtime/execution-plane.js";
 export { InMemorySessionLog, FileSessionLog } from "./runtime/session-log.js";
-export { ReplayProvider } from "./runtime/replay-provider.js";
-export { extractRecordedMessages } from "./runtime/replay-fixture.js";
-export { judge, buildEvalMessages, parseVerdict, verdictOutputSchema } from "./runtime/eval.js";
-export { DEFAULT_NATIVE_ATTENTION_POLICY, DEFAULT_NATIVE_GOVERNANCE_POLICY, assertNativeProfile, osProfile, } from "./runtime/os-profile.js";
-export { rebuildOsSnapshotFromSessionEvents, sessionLogHasRequiredCategories, } from "./runtime/os-snapshot.js";
-export { categoryForKind, kernelObservationToSessionEvent } from "./runtime/kernel-event-log.js";
-export { NullArchiveStore, FileArchiveStore } from "./runtime/archive.js";
-export { EnvCredentialVault, InMemoryCredentialVault, ChainedCredentialVault } from "./runtime/credential-vault.js";
-export { ProcessSandboxPlane } from "./runtime/process-sandbox-plane.js";
-export { McpProxyPlane } from "./runtime/mcp-proxy-plane.js";
-export { RemoteVpcPlane } from "./runtime/remote-vpc-plane.js";
-// ── Providers ─────────────────────────────────────────────────────────────
+export { InMemoryGroupBudgetStore, SessionLogGroupBudgetStore } from "./runtime/run-group.js";
+export { InMemoryEventStream, isVisibleTo } from "./runtime/event-stream.js";
+export { reactByMention, directorDriven, roundRobin, firstNonEmpty, union } from "./runtime/turn-policy.js";
+export { ReactiveSession, readRecentTool } from "./runtime/reactive-session.js";
+// ── Tool authoring ──────────────────────────────────────────────────────────
+export { tool, streamingTool } from "./tools/index.js";
+export { safeTool, ok, fail, ToolError, formatToolError } from "./tools/errors.js";
+// ── Providers (base classes + the universal factory) ────────────────────────
+// Any backend — including a custom OpenAI-compatible endpoint — is reachable via `createProvider`.
+// Backend-specific classes (DeepSeek/Kimi/Qwen/GLM/Gemini/Ollama/MiniMax) live in `@deepstrike/sdk/providers`.
 export { AnthropicProvider } from "./providers/anthropic.js";
-export { OpenAIChatProvider, OpenAIProvider } from "./providers/openai.js";
-export { DeepSeekProvider, DeepSeekAnthropicProvider } from "./providers/deepseek.js";
-export { KimiProvider, KimiAnthropicProvider } from "./providers/kimi.js";
-export { QwenProvider, QwenAnthropicProvider } from "./providers/qwen.js";
-export { GLMProvider, GLMAnthropicProvider } from "./providers/glm.js";
-export { GeminiProvider } from "./providers/gemini.js";
-export { MiniMaxAnthropicProvider, MiniMaxOpenAIProvider } from "./providers/minimax.js";
-export { OllamaProvider } from "./providers/ollama.js";
-export { CircuitBreaker, normalizeToolCall } from "./providers/base.js";
-export { OpenAIChatAdapter } from "./providers/openai-chat.js";
-export { OpenAIResponsesAdapter, OpenAIResponsesProvider } from "./providers/openai-responses.js";
-export { endpointProfiles, modelProfiles, getModelProfile } from "./providers/profiles.js";
+export { OpenAIProvider } from "./providers/openai.js";
+export { OpenAIResponsesProvider } from "./providers/openai-responses.js";
 export { createProvider } from "./providers/catalog.js";
-export { ProviderReplayValidationError, DEGRADED_REASONING_PLACEHOLDER } from "./providers/replay-validator.js";
-export { assessProviderReplayability, peekProviderReplay, seedProviderReplayFromEvents, isReplayCompatibleWithProvider, } from "./runtime/provider-replay.js";
-// ── Tools & Skills ─────────────────────────────────────────────────────────
-export { tool, streamingTool, executeTools, readFile, validateToolArguments } from "./tools/index.js";
-export { safeTool, ok, fail, ToolError, formatToolError } from "./tools/errors.js";
-export { scanSkillDir, readSkillFile } from "./skills/loader.js";
-// ── Memory ─────────────────────────────────────────────────────────────────
-export { WorkingMemory } from "./memory/working.js";
-export { InMemoryDreamStore } from "./memory/in-memory-store.js";
-export { ScheduledPrompt } from "./signals/scheduled.js";
-export { SignalGateway } from "./signals/gateway.js";
-// ── Safety & Governance ────────────────────────────────────────────────────
-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, genEval, } from "./types/agent.js";
-export { ContractBuilder, formatContractForSystemPrompt, contractToCriteriaStrings, } from "./collaboration/contract.js";
+// ── Governance ──────────────────────────────────────────────────────────────
+export { Governance } from "./governance.js";
+// ── Multi-agent primitive ───────────────────────────────────────────────────
+// Parallel fan-out / sub-agent delegation. The full orchestration layer is in `@deepstrike/sdk/workflow`.
 export { AgentPool } from "./collaboration/pool.js";
-export { KERNEL_ROLE_MAP } from "./collaboration/pool.js";
-export { ContractDrivenHarness } from "./collaboration/harness.js";
-export { HandoffBus } from "./collaboration/handoff.js";
-export { CreatorVerifierMode, OrchestrationMode } from "./collaboration/modes/creator-verifier.js";

package/dist/kernel.d.ts

@@ -17,8 +17,12 @@ export interface MemoryWriteRateLimit {
     windowMs: number;
 }
 export interface ResourceQuota {
-    /** Max sub-agents in the `running` state at once; further spawns are denied while at cap. */
+    /** Max sub-agents in the `running` state at once; further spawns are denied while at cap.
+     *  Instantaneous — vehicle-scoped (cannot span stateless replicas). */
     maxConcurrentSubagents?: number;
+    /** L1 (RunGroup): max sub-agents spawned *cumulatively* across the governance domain. With a
+     *  `runGroup`, this spans N stateless top-level runs (seeded/charged via the group ledger). */
+    maxTotalSubagents?: number;
     /** Max sub-agent nesting depth (direct children of the root loop are depth 1). */
     maxSpawnDepth?: number;
     /** Rolling-window memory-write rate limit: at most `maxWrites` per any `windowMs` span. */
@@ -135,6 +139,8 @@ export interface KernelRuntimeInstance {
     step(inputJson: string): string;
     isTerminal(): boolean;
     turn(): number;
+    /** L1 (RunGroup): cumulative sub-agent spawns this run, for charging the group ledger at run end. */
+    localSubagentsSpawned(): number;
     recoveryContentBytes(): number;
     render(): RenderedContext;
     drainNewMessages(): Message[];

package/dist/memory/public.d.ts

@@ -0,0 +1,4 @@
+export { WorkingMemory } from "./working.js";
+export { InMemoryDreamStore } from "./in-memory-store.js";
+export type { DreamStore, DreamResult, SessionData, SessionMessage, MemoryEntry, CurationResult, CurationStats, MemoryWriteRequest, MemoryQuery, MemoryRetrieval, MemoryMetadata, MemoryKind, } from "./protocols.js";
+export type { KnowledgeSource } from "../knowledge/source.js";

package/dist/memory/public.js

@@ -0,0 +1,3 @@
+// `@deepstrike/sdk/memory` — long-term (dream) and working memory, plus the knowledge-source interface.
+export { WorkingMemory } from "./working.js";
+export { InMemoryDreamStore } from "./in-memory-store.js";

package/dist/os/public.d.ts

@@ -0,0 +1,18 @@
+export { DEFAULT_NATIVE_ATTENTION_POLICY, DEFAULT_NATIVE_GOVERNANCE_POLICY, assertNativeProfile, osProfile, } from "../runtime/os-profile.js";
+export type { NativeOsProfile, OsProfileId } from "../runtime/os-profile.js";
+export { rebuildOsSnapshotFromSessionEvents, sessionLogHasRequiredCategories } from "../runtime/os-snapshot.js";
+export type { OsSnapshot } from "../runtime/os-snapshot.js";
+export type { KernelEventCategory } from "../runtime/kernel-event-log.js";
+export { KernelPrimitivesDashboard } from "../runtime/kernel-primitives-dashboard.js";
+export type { MemoryPolicy, MemoryWriteRateLimit, ResourceQuota } from "../kernel.js";
+export type { SchedulerBudget } from "../runtime/runner.js";
+export { ScheduledPrompt } from "../signals/scheduled.js";
+export { SignalGateway } from "../signals/gateway.js";
+export { PermissionManager, PermissionMode } from "../safety/permissions.js";
+export type { PermissionDecision, Permission } from "../safety/permissions.js";
+export { ReplayProvider } from "../runtime/replay-provider.js";
+export type { ReplayProviderOpts } from "../runtime/replay-provider.js";
+export { extractRecordedMessages } from "../runtime/replay-fixture.js";
+export { ProviderReplayValidationError, DEGRADED_REASONING_PLACEHOLDER } from "../providers/replay-validator.js";
+export { assessProviderReplayability, peekProviderReplay, seedProviderReplayFromEvents, isReplayCompatibleWithProvider, } from "../runtime/provider-replay.js";
+export type { ReplayabilityAssessment } from "../types.js";

package/dist/os/public.js

@@ -0,0 +1,14 @@
+// `@deepstrike/sdk/os` — Agent-OS diagnostics, profiles, signal/permission machinery, replay-testing,
+// and the scheduler/quota/policy types referenced by advanced `RuntimeOptions` fields.
+export { DEFAULT_NATIVE_ATTENTION_POLICY, DEFAULT_NATIVE_GOVERNANCE_POLICY, assertNativeProfile, osProfile, } from "../runtime/os-profile.js";
+export { rebuildOsSnapshotFromSessionEvents, sessionLogHasRequiredCategories } from "../runtime/os-snapshot.js";
+export { KernelPrimitivesDashboard } from "../runtime/kernel-primitives-dashboard.js";
+// Signals + SDK-side permissions.
+export { ScheduledPrompt } from "../signals/scheduled.js";
+export { SignalGateway } from "../signals/gateway.js";
+export { PermissionManager, PermissionMode } from "../safety/permissions.js";
+// Replay-based testing utilities.
+export { ReplayProvider } from "../runtime/replay-provider.js";
+export { extractRecordedMessages } from "../runtime/replay-fixture.js";
+export { ProviderReplayValidationError, DEGRADED_REASONING_PLACEHOLDER } from "../providers/replay-validator.js";
+export { assessProviderReplayability, peekProviderReplay, seedProviderReplayFromEvents, isReplayCompatibleWithProvider, } from "../runtime/provider-replay.js";