@deepstrike/sdk 0.2.6 → 0.2.8

This diff represents the content of publicly available package versions that have been released to one of the supported registries. The information contained in this diff is provided for informational purposes only and reflects changes between package versions as they appear in their respective public registries.
Files changed (40) hide show
  1. package/dist/index.d.ts +8 -4
  2. package/dist/index.js +5 -2
  3. package/dist/kernel.d.ts +54 -0
  4. package/dist/kernel.js +8 -0
  5. package/dist/providers/anthropic.d.ts +4 -1
  6. package/dist/providers/anthropic.js +51 -0
  7. package/dist/providers/base.d.ts +6 -0
  8. package/dist/providers/base.js +10 -1
  9. package/dist/providers/catalog.js +5 -2
  10. package/dist/providers/deepseek.d.ts +4 -1
  11. package/dist/providers/deepseek.js +79 -8
  12. package/dist/providers/glm.d.ts +2 -1
  13. package/dist/providers/glm.js +7 -0
  14. package/dist/providers/kimi.d.ts +2 -1
  15. package/dist/providers/kimi.js +7 -0
  16. package/dist/providers/minimax.d.ts +28 -2
  17. package/dist/providers/minimax.js +200 -1
  18. package/dist/providers/openai-chat.d.ts +18 -2
  19. package/dist/providers/openai-chat.js +37 -3
  20. package/dist/providers/openai.d.ts +14 -1
  21. package/dist/providers/openai.js +48 -7
  22. package/dist/providers/profiles.d.ts +6 -0
  23. package/dist/providers/profiles.js +6 -0
  24. package/dist/providers/qwen.d.ts +3 -1
  25. package/dist/providers/qwen.js +20 -2
  26. package/dist/providers/replay-validator.d.ts +32 -0
  27. package/dist/providers/replay-validator.js +90 -0
  28. package/dist/runtime/kernel-event-log.js +22 -0
  29. package/dist/runtime/kernel-step.d.ts +13 -0
  30. package/dist/runtime/provider-replay.d.ts +16 -1
  31. package/dist/runtime/provider-replay.js +47 -4
  32. package/dist/runtime/runner.d.ts +22 -1
  33. package/dist/runtime/runner.js +68 -2
  34. package/dist/runtime/session-log.d.ts +22 -0
  35. package/dist/runtime/session-repair.d.ts +26 -3
  36. package/dist/runtime/session-repair.js +33 -32
  37. package/dist/types/agent.d.ts +50 -0
  38. package/dist/types/agent.js +110 -0
  39. package/dist/types.d.ts +38 -0
  40. package/package.json +2 -2
package/dist/index.d.ts CHANGED
@@ -1,6 +1,8 @@
1
1
  export { RuntimeRunner, collectText } from "./runtime/runner.js";
2
2
  export type { RuntimeOptions, SchedulerBudget } from "./runtime/runner.js";
3
3
  export type { MemoryPolicy, MemoryWriteRateLimit, ResourceQuota } from "./kernel.js";
4
+ export { createTournament, createLoopUntilDone } from "./kernel.js";
5
+ export type { TournamentMatch, TournamentAction, StopConditionSpec, RoundReport, LoopAction, } from "./kernel.js";
4
6
  export { KernelPrimitivesDashboard } from "./runtime/kernel-primitives-dashboard.js";
5
7
  export { FilteredExecutionPlane } from "./runtime/filtered-plane.js";
6
8
  export { SubAgentOrchestrator, defaultSubAgentOrchestrator, spawnStandalone } from "./runtime/sub-agent-orchestrator.js";
@@ -31,7 +33,7 @@ export { DeepSeekProvider } from "./providers/deepseek.js";
31
33
  export { KimiProvider } from "./providers/kimi.js";
32
34
  export { QwenProvider } from "./providers/qwen.js";
33
35
  export { GeminiProvider } from "./providers/gemini.js";
34
- export { MiniMaxProvider } from "./providers/minimax.js";
36
+ export { MiniMaxAnthropicProvider, MiniMaxOpenAIProvider } from "./providers/minimax.js";
35
37
  export { OllamaProvider } from "./providers/ollama.js";
36
38
  export { CircuitBreaker, normalizeToolCall } from "./providers/base.js";
37
39
  export { OpenAIChatAdapter } from "./providers/openai-chat.js";
@@ -41,6 +43,8 @@ export { endpointProfiles, modelProfiles, getModelProfile } from "./providers/pr
41
43
  export type { ModelProfileId, ProviderId } from "./providers/profiles.js";
42
44
  export { createProvider } from "./providers/catalog.js";
43
45
  export type { CreateProviderOptions, EndpointProfileId } from "./providers/catalog.js";
46
+ export { ProviderReplayValidationError, DEGRADED_REASONING_PLACEHOLDER } from "./providers/replay-validator.js";
47
+ export { assessProviderReplayability, peekProviderReplay, seedProviderReplayFromEvents, isReplayCompatibleWithProvider, } from "./runtime/provider-replay.js";
44
48
  export { tool, streamingTool, executeTools, readFile, validateToolArguments } from "./tools/index.js";
45
49
  export type { RegisteredTool } from "./tools/index.js";
46
50
  export { scanSkillDir, readSkillFile } from "./skills/loader.js";
@@ -57,9 +61,9 @@ export { Governance, governancePolicyToKernelEvent } from "./governance.js";
57
61
  export type { GovernanceVerdict, GovernancePolicy, GovernanceConstraint } from "./governance.js";
58
62
  export { SinglePassHarness, EvalLoopHarness, HarnessLoop } from "./harness/harness.js";
59
63
  export type { HarnessRequest, HarnessOutcome, HarnessLoopOptions, QualityGate } from "./harness/harness.js";
60
- 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, } from "./types.js";
61
- export type { AgentCapabilityFilter, AgentIdentity, AgentIsolation, AgentRunSpec, AgentProcessChangedObservation, ContextInheritance, KernelAgentRole, LoopResult, MilestoneCheckResult, MilestoneContract, MilestonePhase, MilestonePolicy, SubAgentResult, TerminationReason, } from "./types/agent.js";
62
- export { agentIdentitySub, agentRunSpecToKernel, milestoneCheckFail, milestoneCheckPass, milestoneCheckResultToKernel, subAgentResultToKernel, } from "./types/agent.js";
64
+ 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";
65
+ export type { AgentCapabilityFilter, AgentIdentity, AgentIsolation, AgentRunSpec, AgentProcessChangedObservation, ContextInheritance, KernelAgentRole, LoopResult, MilestoneCheckResult, MilestoneContract, MilestonePhase, MilestonePolicy, SubAgentResult, TerminationReason, WorkflowSpec, WorkflowNodeSpec, WorkflowTaskSpec, WorkflowSpawnInfo, } from "./types/agent.js";
66
+ export { agentIdentitySub, agentRunSpecToKernel, milestoneCheckFail, milestoneCheckPass, milestoneCheckResultToKernel, subAgentResultToKernel, workflowSpecToKernel, fanoutSynthesize, generateAndFilter, verifyRules, } from "./types/agent.js";
63
67
  export type { AcceptanceCriterion, VerificationContract, ContractCheckResult, } from "./collaboration/contract.js";
64
68
  export { ContractBuilder, formatContractForSystemPrompt, contractToCriteriaStrings, } from "./collaboration/contract.js";
65
69
  export { AgentPool } from "./collaboration/pool.js";
package/dist/index.js CHANGED
@@ -1,5 +1,6 @@
1
1
  // ── Runtime (Layer 1.5) ────────────────────────────────────────────────────
2
2
  export { RuntimeRunner, collectText } from "./runtime/runner.js";
3
+ export { createTournament, createLoopUntilDone } from "./kernel.js";
3
4
  export { KernelPrimitivesDashboard } from "./runtime/kernel-primitives-dashboard.js";
4
5
  export { FilteredExecutionPlane } from "./runtime/filtered-plane.js";
5
6
  export { SubAgentOrchestrator, defaultSubAgentOrchestrator, spawnStandalone } from "./runtime/sub-agent-orchestrator.js";
@@ -20,13 +21,15 @@ export { DeepSeekProvider } from "./providers/deepseek.js";
20
21
  export { KimiProvider } from "./providers/kimi.js";
21
22
  export { QwenProvider } from "./providers/qwen.js";
22
23
  export { GeminiProvider } from "./providers/gemini.js";
23
- export { MiniMaxProvider } from "./providers/minimax.js";
24
+ export { MiniMaxAnthropicProvider, MiniMaxOpenAIProvider } from "./providers/minimax.js";
24
25
  export { OllamaProvider } from "./providers/ollama.js";
25
26
  export { CircuitBreaker, normalizeToolCall } from "./providers/base.js";
26
27
  export { OpenAIChatAdapter } from "./providers/openai-chat.js";
27
28
  export { OpenAIResponsesAdapter, OpenAIResponsesProvider } from "./providers/openai-responses.js";
28
29
  export { endpointProfiles, modelProfiles, getModelProfile } from "./providers/profiles.js";
29
30
  export { createProvider } from "./providers/catalog.js";
31
+ export { ProviderReplayValidationError, DEGRADED_REASONING_PLACEHOLDER } from "./providers/replay-validator.js";
32
+ export { assessProviderReplayability, peekProviderReplay, seedProviderReplayFromEvents, isReplayCompatibleWithProvider, } from "./runtime/provider-replay.js";
30
33
  // ── Tools & Skills ─────────────────────────────────────────────────────────
31
34
  export { tool, streamingTool, executeTools, readFile, validateToolArguments } from "./tools/index.js";
32
35
  export { scanSkillDir, readSkillFile } from "./skills/loader.js";
@@ -39,7 +42,7 @@ export { PermissionManager, PermissionMode } from "./safety/permissions.js";
39
42
  export { Governance, governancePolicyToKernelEvent } from "./governance.js";
40
43
  // ── Harness ────────────────────────────────────────────────────────────────
41
44
  export { SinglePassHarness, EvalLoopHarness, HarnessLoop } from "./harness/harness.js";
42
- export { agentIdentitySub, agentRunSpecToKernel, milestoneCheckFail, milestoneCheckPass, milestoneCheckResultToKernel, subAgentResultToKernel, } from "./types/agent.js";
45
+ export { agentIdentitySub, agentRunSpecToKernel, milestoneCheckFail, milestoneCheckPass, milestoneCheckResultToKernel, subAgentResultToKernel, workflowSpecToKernel, fanoutSynthesize, generateAndFilter, verifyRules, } from "./types/agent.js";
43
46
  export { ContractBuilder, formatContractForSystemPrompt, contractToCriteriaStrings, } from "./collaboration/contract.js";
44
47
  export { AgentPool } from "./collaboration/pool.js";
45
48
  export { KERNEL_ROLE_MAP } from "./collaboration/pool.js";
package/dist/kernel.d.ts CHANGED
@@ -147,6 +147,54 @@ export interface KernelRuntimeInstance {
147
147
  drainNewMessages(): Message[];
148
148
  preservedRefs(): string[];
149
149
  }
150
+ /** One pairwise match-up in a tournament round. */
151
+ export interface TournamentMatch {
152
+ id: number;
153
+ left: string;
154
+ right: string;
155
+ }
156
+ /** Discriminated action returned by {@link TournamentInstance} methods. */
157
+ export interface TournamentAction {
158
+ kind: "judgeRound" | "done";
159
+ /** `judgeRound`: 1-based round number. */
160
+ round?: number;
161
+ /** `judgeRound`: run one fresh-context judge per match (parallelisable). */
162
+ matches?: TournamentMatch[];
163
+ /** `done`: the winning entrant id. */
164
+ winner?: string;
165
+ /** `done`: number of rounds played. */
166
+ roundsUsed?: number;
167
+ }
168
+ interface TournamentInstance {
169
+ start(): TournamentAction;
170
+ feedRound(winners: string[]): TournamentAction;
171
+ isDone(): boolean;
172
+ }
173
+ /** A single loop stop predicate. `maxRounds` is required when `kind === "maxRounds"`. */
174
+ export interface StopConditionSpec {
175
+ kind: "noNewFindings" | "noErrors" | "maxRounds";
176
+ maxRounds?: number;
177
+ }
178
+ /** What the SDK reports after running a loop round's worker. */
179
+ export interface RoundReport {
180
+ newFindings: number;
181
+ errors: number;
182
+ }
183
+ /** Discriminated action returned by {@link LoopUntilDoneInstance} methods. */
184
+ export interface LoopAction {
185
+ kind: "spawn" | "done";
186
+ /** `spawn`: 1-based round number to run. */
187
+ round?: number;
188
+ /** `done`: number of rounds run. */
189
+ roundsUsed?: number;
190
+ /** `done`: which condition fired. */
191
+ reason?: "noNewFindings" | "noErrors" | "maxRounds";
192
+ }
193
+ interface LoopUntilDoneInstance {
194
+ start(): LoopAction;
195
+ feed(report: RoundReport): LoopAction;
196
+ isDone(): boolean;
197
+ }
150
198
  interface KernelModule {
151
199
  Governance: new (defaultAction?: "allow" | "deny" | "ask_user") => GovernanceInstance;
152
200
  KernelRuntime: new (policy: {
@@ -160,6 +208,12 @@ interface KernelModule {
160
208
  extractSkillOnPass?: boolean;
161
209
  }) => EvalPipelineInstance;
162
210
  IdlePipeline: new (agentId: string) => IdlePipelineInstance;
211
+ Tournament: new (entrants: string[]) => TournamentInstance;
212
+ LoopUntilDone: new (conditions: StopConditionSpec[]) => LoopUntilDoneInstance;
163
213
  }
214
+ /** Create a single-elimination tournament (pairwise comparative judging). Throws if `entrants` is empty. */
215
+ export declare function createTournament(entrants: string[]): TournamentInstance;
216
+ /** Create a loop-until-done state machine. A `maxRounds` backstop is injected if none is given. */
217
+ export declare function createLoopUntilDone(conditions: StopConditionSpec[]): LoopUntilDoneInstance;
164
218
  export declare function getKernel(): KernelModule;
165
219
  export {};
package/dist/kernel.js CHANGED
@@ -2,6 +2,14 @@ import { createRequire } from "module";
2
2
  import { existsSync } from "node:fs";
3
3
  import { dirname, join } from "node:path";
4
4
  import { fileURLToPath } from "node:url";
5
+ /** Create a single-elimination tournament (pairwise comparative judging). Throws if `entrants` is empty. */
6
+ export function createTournament(entrants) {
7
+ return new (getKernel().Tournament)(entrants);
8
+ }
9
+ /** Create a loop-until-done state machine. A `maxRounds` backstop is injected if none is given. */
10
+ export function createLoopUntilDone(conditions) {
11
+ return new (getKernel().LoopUntilDone)(conditions);
12
+ }
5
13
  const cjsRequire = createRequire(import.meta.url);
6
14
  let cachedKernel;
7
15
  function resolveCoreModule() {
@@ -1,4 +1,4 @@
1
- import type { Message, ProviderReplay, RenderedContext, ToolSchema, StreamEvent, LLMProvider, RuntimePolicy } from "../types.js";
1
+ import type { Message, ProviderDescriptor, ProviderReplay, RenderedContext, ToolSchema, StreamEvent, LLMProvider, RuntimePolicy } from "../types.js";
2
2
  interface AnthropicProviderOptions {
3
3
  baseURL?: string;
4
4
  authMode?: "api-key" | "bearer";
@@ -15,6 +15,9 @@ export declare class AnthropicProvider implements LLMProvider {
15
15
  baseDelay: number;
16
16
  }, options?: AnthropicProviderOptions);
17
17
  runtimePolicy(): RuntimePolicy;
18
+ /** Identity advertised in the descriptor; overridden by Anthropic-compatible vendors (e.g. MiniMax). */
19
+ protected providerName(): string;
20
+ descriptor(): ProviderDescriptor;
18
21
  peekProviderReplay(message: Pick<Message, "content" | "toolCalls">): ProviderReplay | undefined;
19
22
  seedProviderReplay(message: Pick<Message, "content" | "toolCalls">, replay: ProviderReplay): void;
20
23
  private buildTools;
@@ -34,6 +34,26 @@ export class AnthropicProvider {
34
34
  runtimePolicy() {
35
35
  return CLAUDE_POLICIES[this.model] ?? {};
36
36
  }
37
+ /** Identity advertised in the descriptor; overridden by Anthropic-compatible vendors (e.g. MiniMax). */
38
+ providerName() {
39
+ return "anthropic";
40
+ }
41
+ descriptor() {
42
+ return {
43
+ provider: this.providerName(),
44
+ protocol: "anthropic-messages",
45
+ model: this.model,
46
+ reasoning: {
47
+ supported: true,
48
+ preserveAcrossToolTurns: true,
49
+ requiresReplayForToolTurns: true,
50
+ },
51
+ toolCalls: {
52
+ supported: true,
53
+ requiresStrictPairing: true,
54
+ },
55
+ };
56
+ }
37
57
  peekProviderReplay(message) {
38
58
  const blocks = this.nativeAssistantBlocks.get(assistantReplayKey(message));
39
59
  return blocks?.length ? { native_blocks: blocks } : undefined;
@@ -41,7 +61,14 @@ export class AnthropicProvider {
41
61
  seedProviderReplay(message, replay) {
42
62
  if (replay.native_blocks?.length) {
43
63
  this.nativeAssistantBlocks.set(assistantReplayKey(message), replay.native_blocks);
64
+ return;
44
65
  }
66
+ // Legacy log without persisted native blocks: reconstruct neutral
67
+ // text + tool_use blocks from the transcript so a tool-use turn can be
68
+ // replayed. Thinking blocks were never persisted, so they are not recovered.
69
+ const blocks = reconstructAnthropicBlocks(message);
70
+ if (blocks.length)
71
+ this.nativeAssistantBlocks.set(assistantReplayKey(message), blocks);
45
72
  }
46
73
  buildTools(tools) {
47
74
  return tools.map((t, i) => ({
@@ -206,3 +233,27 @@ export class AnthropicProvider {
206
233
  this.nativeAssistantBlocks.set(assistantReplayKey(message), blocks);
207
234
  }
208
235
  }
236
+ /**
237
+ * Reconstruct Anthropic assistant content blocks from a neutral transcript when
238
+ * no provider replay was persisted. Only meaningful for tool-use turns: a plain
239
+ * text turn needs no native blocks to replay.
240
+ */
241
+ function reconstructAnthropicBlocks(message) {
242
+ const toolCalls = message.toolCalls ?? [];
243
+ if (!toolCalls.length)
244
+ return [];
245
+ const blocks = [];
246
+ if (message.content)
247
+ blocks.push({ type: "text", text: message.content });
248
+ for (const tc of toolCalls) {
249
+ let input = {};
250
+ try {
251
+ input = JSON.parse(tc.arguments || "{}");
252
+ }
253
+ catch {
254
+ input = {};
255
+ }
256
+ blocks.push({ type: "tool_use", id: tc.id, name: tc.name, input });
257
+ }
258
+ return blocks;
259
+ }
@@ -9,6 +9,12 @@ export declare class CircuitBreaker {
9
9
  recordSuccess(): void;
10
10
  recordFailure(): void;
11
11
  }
12
+ /**
13
+ * Internal control flags that steer DeepStrike's own serialization/validation
14
+ * and must never be forwarded to any provider's wire request, regardless of the
15
+ * per-call omit list.
16
+ */
17
+ export declare const INTERNAL_EXTENSION_KEYS: readonly string[];
12
18
  export declare function omitExtensionKeys(extensions: Record<string, unknown> | undefined, keys: readonly string[]): Record<string, unknown>;
13
19
  export declare function normalizeToolCall(id: string, name: string, args: unknown): {
14
20
  id: string;
@@ -26,10 +26,19 @@ export class CircuitBreaker {
26
26
  this.openedAt = Date.now();
27
27
  }
28
28
  }
29
+ /**
30
+ * Internal control flags that steer DeepStrike's own serialization/validation
31
+ * and must never be forwarded to any provider's wire request, regardless of the
32
+ * per-call omit list.
33
+ */
34
+ export const INTERNAL_EXTENSION_KEYS = [
35
+ "__deepstrikeThinkingEnabled",
36
+ "degradeMissingReasoningReplay",
37
+ ];
29
38
  export function omitExtensionKeys(extensions, keys) {
30
39
  if (!extensions)
31
40
  return {};
32
- const blocked = new Set(keys);
41
+ const blocked = new Set([...keys, ...INTERNAL_EXTENSION_KEYS]);
33
42
  return Object.fromEntries(Object.entries(extensions).filter(([key]) => !blocked.has(key)));
34
43
  }
35
44
  export function normalizeToolCall(id, name, args) {
@@ -3,7 +3,7 @@ import { OpenAIChatProvider } from "./openai.js";
3
3
  import { DeepSeekProvider } from "./deepseek.js";
4
4
  import { KimiProvider } from "./kimi.js";
5
5
  import { OpenAIResponsesProvider } from "./openai-responses.js";
6
- import { MiniMaxProvider } from "./minimax.js";
6
+ import { MiniMaxAnthropicProvider, MiniMaxOpenAIProvider } from "./minimax.js";
7
7
  import { QwenProvider } from "./qwen.js";
8
8
  import { GeminiProvider } from "./gemini.js";
9
9
  import { GLMProvider } from "./glm.js";
@@ -43,7 +43,10 @@ export function createProvider(options) {
43
43
  }
44
44
  }
45
45
  if (providerId === "minimax" && endpoint.protocol === "anthropic-messages") {
46
- return new MiniMaxProvider(options.apiKey, model, options.retry, baseURL);
46
+ return new MiniMaxAnthropicProvider(options.apiKey, model, options.retry, baseURL);
47
+ }
48
+ if (providerId === "minimax" && endpoint.protocol === "openai-chat") {
49
+ return new MiniMaxOpenAIProvider(options.apiKey, model, options.retry, baseURL);
47
50
  }
48
51
  if (providerId === "deepseek" && endpoint.protocol === "openai-chat") {
49
52
  return new DeepSeekProvider(options.apiKey, model, options.retry, baseURL);
@@ -1,4 +1,4 @@
1
- import type { Message, RenderedContext, ToolSchema, StreamEvent, RuntimePolicy } from "../types.js";
1
+ import type { Message, ProviderDescriptor, RenderedContext, ToolSchema, StreamEvent, RuntimePolicy } from "../types.js";
2
2
  import { OpenAIChatProvider } from "./openai.js";
3
3
  export declare class DeepSeekProvider extends OpenAIChatProvider {
4
4
  constructor(apiKey: string, model?: string, retry?: {
@@ -6,6 +6,9 @@ export declare class DeepSeekProvider extends OpenAIChatProvider {
6
6
  baseDelay: number;
7
7
  }, baseURL?: string);
8
8
  runtimePolicy(): RuntimePolicy;
9
+ descriptor(): ProviderDescriptor;
10
+ protected requireNonEmptyReasoningReplayForToolTurns(extensions?: Record<string, unknown>): boolean;
9
11
  complete(context: RenderedContext, tools: ToolSchema[], extensions?: Record<string, unknown>): Promise<Message>;
10
12
  stream(context: RenderedContext, tools: ToolSchema[], extensions?: Record<string, unknown>): AsyncIterable<StreamEvent>;
13
+ private rememberDeepSeekReplay;
11
14
  }
@@ -15,20 +15,74 @@ export class DeepSeekProvider extends OpenAIChatProvider {
15
15
  runtimePolicy() {
16
16
  return DEEPSEEK_POLICIES[this.model] ?? {};
17
17
  }
18
+ descriptor() {
19
+ return {
20
+ provider: "deepseek",
21
+ protocol: "openai-chat",
22
+ model: this.model,
23
+ reasoning: {
24
+ supported: true,
25
+ preserveAcrossToolTurns: true,
26
+ requiresReplayForToolTurns: true,
27
+ },
28
+ toolCalls: {
29
+ supported: true,
30
+ requiresStrictPairing: true,
31
+ },
32
+ };
33
+ }
34
+ requireNonEmptyReasoningReplayForToolTurns(extensions) {
35
+ if (extensions?.__deepstrikeThinkingEnabled === false)
36
+ return false;
37
+ return extensions?.thinking !== false;
38
+ }
18
39
  async complete(context, tools, extensions) {
19
40
  const thinking = extensions?.thinking === false ? "disabled" : "enabled";
41
+ const thinkingEnabled = thinking !== "disabled";
20
42
  const reasoningEffort = extensions?.reasoningEffort === "max" ? "max" : "high";
21
- return super.complete(context, tools, {
43
+ const requestExtensions = {
22
44
  ...omitExtensionKeys(extensions, ["thinking", "reasoningEffort", "exposeReasoning", "extra_body", "reasoning_effort"]),
45
+ __deepstrikeThinkingEnabled: thinkingEnabled,
46
+ // Re-thread the degrade control flag (omitExtensionKeys strips internal
47
+ // keys) so buildChatMessages can honor it; the wire-request omit drops it.
48
+ ...(extensions?.degradeMissingReasoningReplay === true ? { degradeMissingReasoningReplay: true } : {}),
23
49
  reasoning_effort: reasoningEffort,
24
50
  extra_body: { thinking: { type: thinking } },
25
- });
51
+ };
52
+ if (this.circuit.isOpen())
53
+ throw new Error("Circuit breaker open");
54
+ const msgs = this.buildChatMessages(context, requestExtensions);
55
+ let lastErr;
56
+ for (let i = 0; i < this.maxRetries; i++) {
57
+ try {
58
+ const resp = await this.client.chat.completions.create({
59
+ ...this.requestExtensions(requestExtensions),
60
+ model: this.model,
61
+ messages: msgs,
62
+ ...(tools.length ? { tools: this.chat.buildTools(tools) } : {}),
63
+ });
64
+ this.circuit.recordSuccess();
65
+ const choice = resp.choices[0].message;
66
+ const nativeToolCalls = choice.tool_calls ?? [];
67
+ const toolCalls = this.chat.normalizeToolCalls(nativeToolCalls);
68
+ const content = choice.content ?? "";
69
+ this.rememberDeepSeekReplay(content, toolCalls, choice.reasoning_content, nativeToolCalls);
70
+ return { role: "assistant", content, tokenCount: resp.usage?.completion_tokens ?? resp.usage?.total_tokens, toolCalls };
71
+ }
72
+ catch (err) {
73
+ lastErr = err;
74
+ this.circuit.recordFailure();
75
+ if (i < this.maxRetries - 1)
76
+ await new Promise(r => setTimeout(r, this.baseDelay * 2 ** i));
77
+ }
78
+ }
79
+ throw lastErr;
26
80
  }
27
81
  async *stream(context, tools, extensions) {
28
82
  const exposeReasoning = extensions?.exposeReasoning ?? false;
29
83
  const thinking = extensions?.thinking === false ? "disabled" : "enabled";
30
84
  const reasoningEffort = extensions?.reasoningEffort === "max" ? "max" : "high";
31
- const msgs = this.chat.buildMessages(context);
85
+ const msgs = this.buildChatMessages(context, extensions);
32
86
  const toolCallBufs = {};
33
87
  const emittedToolCallIndexes = new Set();
34
88
  let reasoningContent = "";
@@ -36,7 +90,7 @@ export class DeepSeekProvider extends OpenAIChatProvider {
36
90
  const stream = await this.client.chat.completions.create({
37
91
  ...omitExtensionKeys(extensions, [
38
92
  "model", "messages", "tools", "stream", "stream_options", "extra_body", "reasoning_effort",
39
- "exposeReasoning", "thinking", "reasoningEffort",
93
+ "exposeReasoning", "thinking", "reasoningEffort", "__deepstrikeThinkingEnabled",
40
94
  ]),
41
95
  model: this.model,
42
96
  messages: msgs,
@@ -83,7 +137,7 @@ export class DeepSeekProvider extends OpenAIChatProvider {
83
137
  const toolCalls = Object.values(toolCallBufs).map(tb => ({
84
138
  id: tb.id, name: tb.name, arguments: tb.argsBuf || "{}",
85
139
  }));
86
- this.chat.rememberReplayFields({ content: finalText, toolCalls }, { reasoning_content: reasoningContent });
140
+ this.rememberDeepSeekReplay(finalText, toolCalls, reasoningContent, nativeToolCallsFromBuffers(toolCallBufs));
87
141
  for (const [index, tb] of Object.entries(toolCallBufs)) {
88
142
  const idx = Number(index);
89
143
  if (emittedToolCallIndexes.has(idx))
@@ -103,9 +157,7 @@ export class DeepSeekProvider extends OpenAIChatProvider {
103
157
  const toolCalls = Object.values(toolCallBufs).map(tb => ({
104
158
  id: tb.id, name: tb.name, arguments: tb.argsBuf || "{}",
105
159
  }));
106
- if (toolCalls.length || reasoningContent) {
107
- this.chat.rememberReplayFields({ content: finalText, toolCalls }, { reasoning_content: reasoningContent });
108
- }
160
+ this.rememberDeepSeekReplay(finalText, toolCalls, reasoningContent, nativeToolCallsFromBuffers(toolCallBufs));
109
161
  for (const [index, tb] of Object.entries(toolCallBufs)) {
110
162
  const idx = Number(index);
111
163
  if (emittedToolCallIndexes.has(idx))
@@ -123,4 +175,23 @@ export class DeepSeekProvider extends OpenAIChatProvider {
123
175
  if (totalTokens > 0)
124
176
  yield { type: "usage", totalTokens, inputTokens, outputTokens };
125
177
  }
178
+ rememberDeepSeekReplay(content, toolCalls, reasoningContent, nativeToolCalls) {
179
+ if (typeof reasoningContent !== "string" || !reasoningContent.trim())
180
+ return;
181
+ this.chat.rememberReplayFields({ content, toolCalls }, {
182
+ schema_version: 2,
183
+ provider: "deepseek",
184
+ protocol: "openai-chat",
185
+ model: this.model,
186
+ reasoning_content: reasoningContent,
187
+ ...(nativeToolCalls.length ? { tool_calls: nativeToolCalls } : {}),
188
+ });
189
+ }
190
+ }
191
+ function nativeToolCallsFromBuffers(toolCallBufs) {
192
+ return Object.values(toolCallBufs).map(tb => ({
193
+ id: tb.id,
194
+ type: "function",
195
+ function: { name: tb.name, arguments: tb.argsBuf || "{}" },
196
+ }));
126
197
  }
@@ -1,4 +1,4 @@
1
- import type { RuntimePolicy } from "../types.js";
1
+ import type { ProviderDescriptor, RuntimePolicy } from "../types.js";
2
2
  import { OpenAIChatProvider } from "./openai.js";
3
3
  export declare class GLMProvider extends OpenAIChatProvider {
4
4
  constructor(apiKey: string, model?: string, retry?: {
@@ -6,4 +6,5 @@ export declare class GLMProvider extends OpenAIChatProvider {
6
6
  baseDelay: number;
7
7
  }, baseURL?: string);
8
8
  runtimePolicy(): RuntimePolicy;
9
+ descriptor(): ProviderDescriptor;
9
10
  }
@@ -18,4 +18,11 @@ export class GLMProvider extends OpenAIChatProvider {
18
18
  runtimePolicy() {
19
19
  return GLM_POLICIES[this.model] ?? {};
20
20
  }
21
+ descriptor() {
22
+ return {
23
+ ...super.descriptor(),
24
+ provider: "glm",
25
+ model: this.model,
26
+ };
27
+ }
21
28
  }
@@ -1,4 +1,4 @@
1
- import type { RuntimePolicy } from "../types.js";
1
+ import type { ProviderDescriptor, RuntimePolicy } from "../types.js";
2
2
  import { OpenAIChatProvider } from "./openai.js";
3
3
  export declare class KimiProvider extends OpenAIChatProvider {
4
4
  constructor(apiKey: string, model?: string, retry?: {
@@ -6,4 +6,5 @@ export declare class KimiProvider extends OpenAIChatProvider {
6
6
  baseDelay: number;
7
7
  }, baseURL?: string);
8
8
  runtimePolicy(): RuntimePolicy;
9
+ descriptor(): ProviderDescriptor;
9
10
  }
@@ -17,4 +17,11 @@ export class KimiProvider extends OpenAIChatProvider {
17
17
  runtimePolicy() {
18
18
  return KIMI_POLICIES[this.model] ?? {};
19
19
  }
20
+ descriptor() {
21
+ return {
22
+ ...super.descriptor(),
23
+ provider: "kimi",
24
+ model: this.model,
25
+ };
26
+ }
20
27
  }
@@ -1,9 +1,35 @@
1
- import type { RuntimePolicy } from "../types.js";
1
+ import type { Message, ProviderDescriptor, RenderedContext, RuntimePolicy, StreamEvent, ToolSchema } from "../types.js";
2
2
  import { AnthropicProvider } from "./anthropic.js";
3
- export declare class MiniMaxProvider extends AnthropicProvider {
3
+ import { OpenAIChatProvider } from "./openai.js";
4
+ /**
5
+ * MiniMax over its Anthropic-compatible endpoint. Replay is carried as Anthropic
6
+ * `native_blocks` (thinking/text/tool_use), identical to the first-party
7
+ * Anthropic provider.
8
+ */
9
+ export declare class MiniMaxAnthropicProvider extends AnthropicProvider {
4
10
  constructor(apiKey: string, model?: string, retry?: {
5
11
  maxRetries: number;
6
12
  baseDelay: number;
7
13
  }, baseURL?: string);
14
+ protected providerName(): string;
8
15
  runtimePolicy(): RuntimePolicy;
9
16
  }
17
+ /**
18
+ * MiniMax over its OpenAI-compatible endpoint. Replay is carried as
19
+ * `reasoning_content` / `reasoning_details` (split reasoning), and requests
20
+ * default to `reasoning_split: true` so reasoning is returned out-of-band rather
21
+ * than embedded in the message content.
22
+ */
23
+ export declare class MiniMaxOpenAIProvider extends OpenAIChatProvider {
24
+ constructor(apiKey: string, model?: string, retry?: {
25
+ maxRetries: number;
26
+ baseDelay: number;
27
+ }, baseURL?: string);
28
+ runtimePolicy(): RuntimePolicy;
29
+ descriptor(): ProviderDescriptor;
30
+ protected requireNonEmptyReasoningReplayForToolTurns(extensions?: Record<string, unknown>): boolean;
31
+ private buildRequestExtensions;
32
+ complete(context: RenderedContext, tools: ToolSchema[], extensions?: Record<string, unknown>): Promise<Message>;
33
+ stream(context: RenderedContext, tools: ToolSchema[], extensions?: Record<string, unknown>): AsyncIterable<StreamEvent>;
34
+ private rememberMiniMaxReplay;
35
+ }