@elevasis/sdk 1.43.0 → 1.44.0

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.
@@ -274,152 +274,6 @@ interface IExecutionLogger {
274
274
  error(message: string, context?: LogContext): void;
275
275
  }
276
276
 
277
- /**
278
- * Model Configuration
279
- * Centralized model information, configuration, options, constraints, and validation
280
- * Single source of truth for all model-related definitions
281
- * Update manually when pricing changes or new models are added
282
- */
283
-
284
- /**
285
- * Supported Open AI models (direct SDK access)
286
- */
287
- type OpenAIModel = 'gpt-5' | 'gpt-5.4-mini' | 'gpt-5.4-nano';
288
- /**
289
- * Supported OpenRouter models (explicit union for type safety)
290
- */
291
- type OpenRouterModel = 'openrouter/z-ai/glm-5';
292
- /**
293
- * Supported Anthropic models (direct SDK access via @anthropic-ai/sdk)
294
- */
295
- type AnthropicModel = 'claude-opus-5' | 'claude-sonnet-5' | 'claude-haiku-4-5-20251001' | 'claude-haiku-4-5';
296
- /** Supported LLM models */
297
- type LLMModel = OpenAIModel | OpenRouterModel | AnthropicModel | 'mock';
298
- /**
299
- * GPT-5 model options schema
300
- */
301
- declare const GPT5OptionsSchema: z.ZodObject<{
302
- reasoning_effort: z.ZodOptional<z.ZodEnum<{
303
- minimal: "minimal";
304
- low: "low";
305
- medium: "medium";
306
- high: "high";
307
- }>>;
308
- verbosity: z.ZodOptional<z.ZodEnum<{
309
- low: "low";
310
- medium: "medium";
311
- high: "high";
312
- }>>;
313
- }, z.core.$strip>;
314
- /**
315
- * OpenRouter model options schema
316
- * OpenRouter-specific options for routing and transforms
317
- */
318
- declare const OpenRouterOptionsSchema: z.ZodObject<{
319
- transforms: z.ZodOptional<z.ZodArray<z.ZodString>>;
320
- route: z.ZodOptional<z.ZodEnum<{
321
- fallback: "fallback";
322
- }>>;
323
- }, z.core.$strip>;
324
- /**
325
- * Anthropic model options schema
326
- * Currently empty - future options must be added per supported model family
327
- */
328
- declare const AnthropicOptionsSchema: z.ZodObject<{}, z.core.$strict>;
329
- /**
330
- * Infer TypeScript types from schemas
331
- */
332
- type GPT5Options = z.infer<typeof GPT5OptionsSchema>;
333
- type MockOptions = Record<string, never>;
334
- type OpenRouterOptions = z.infer<typeof OpenRouterOptionsSchema>;
335
- type AnthropicOptions = z.infer<typeof AnthropicOptionsSchema>;
336
- type ModelSpecificOptions = GPT5Options | MockOptions | OpenRouterOptions | AnthropicOptions;
337
- /**
338
- * Model configuration for LLM execution
339
- * Belongs in resource definition (AgentDefinition, WorkflowDefinition, etc.)
340
- */
341
- interface ModelConfig {
342
- model: LLMModel;
343
- provider: 'openai' | 'anthropic' | 'openrouter' | 'mock';
344
- apiKey: string;
345
- temperature?: number;
346
- /** Maximum output tokens per LLM call. NOT the model's context window — see ModelInfo.maxTokens for that. */
347
- maxOutputTokens?: number;
348
- topP?: number;
349
- /**
350
- * Model-specific options (flat structure)
351
- * Options are model-specific, not vendor-specific
352
- * Available options defined in MODEL_INFO per model
353
- * Validated at build time via validateModelOptions()
354
- */
355
- modelOptions?: ModelSpecificOptions;
356
- }
357
-
358
- /**
359
- * Types for the schema compiler. `compile.ts` walks a `JsonSchema` once, driven entirely by a
360
- * `ProviderDialect`, and every server adapter compiles through it.
361
- */
362
- /**
363
- * What happened to `strict` on a request, recorded per call rather than inferred.
364
- *
365
- * `applied` and `notAttempted` are the two states that a refusal-only field cannot tell apart --
366
- * both leave `strictRefusalReasons` empty. Recording the verdict positively is what makes "was
367
- * this agent's output actually enforced?" answerable from an `ai_calls` row.
368
- */
369
- type StrictStatus = 'applied' | 'refused' | 'compileRejected' | 'notAttempted';
370
- /**
371
- * A JSON Schema node, typed enough to be useful without pretending to validate the spec.
372
- *
373
- * The compiler has to accept schemas that arrive OUTSIDE the strict subset (that is the whole
374
- * point of a dialect that can refuse or rewrite them) as well as the `$ref`/`$defs`/`const`/
375
- * `$schema` shapes the strict subset has no vocabulary for at all. The index signature exists
376
- * because tenant schemas carry keywords (`minLength`, `pattern`, `minimum`, ...) this compiler
377
- * drops or refuses on, and they still need somewhere to type-check while they pass through
378
- * `Object.entries`.
379
- */
380
- interface JsonSchema {
381
- type?: string | string[];
382
- /**
383
- * The value is `JsonSchema | undefined`, not `JsonSchema`, because a property really can be
384
- * declared with nothing describing it. `buildIterationResponseSchema` emits one per tool as
385
- * `input: tool.inputSchema`, and `ToolDefinition.inputSchema` is typed `unknown` -- a tool
386
- * deployed without one puts `undefined` under a key that exists.
387
- *
388
- * Both readers already handle it: `compileSchema` passes each value through `convertNode`, which
389
- * takes `unknown`, and `collectErrors` opens with `if (!schema || typeof schema !== 'object')`
390
- * above a comment naming this exact case. Declaring the value non-optional only hid that they
391
- * were right to.
392
- */
393
- properties?: Record<string, JsonSchema | undefined>;
394
- items?: JsonSchema;
395
- anyOf?: JsonSchema[];
396
- oneOf?: JsonSchema[];
397
- allOf?: JsonSchema[];
398
- required?: string[];
399
- additionalProperties?: boolean | JsonSchema;
400
- minItems?: number;
401
- maxItems?: number;
402
- format?: string;
403
- enum?: unknown[];
404
- const?: unknown;
405
- description?: string;
406
- default?: unknown;
407
- $ref?: string;
408
- $defs?: Record<string, JsonSchema>;
409
- definitions?: Record<string, JsonSchema>;
410
- $schema?: string;
411
- $id?: string;
412
- $anchor?: string;
413
- /**
414
- * OpenAPI's nullability spelling, which is not JSON Schema's. It is declared because
415
- * `response-schema-validator.ts` READS it (`schemaPermitsNull`) -- Google's schema dialect is
416
- * OpenAPI-derived, so a schema that reaches the validator can carry it. No dialect in `compile.ts`
417
- * writes or rewrites it; the canonical spelling this compiler emits is `type: ['x', 'null']`.
418
- */
419
- nullable?: boolean;
420
- [key: string]: unknown;
421
- }
422
-
423
277
  declare const ResourceGovernanceStatusSchema: z.ZodEnum<{
424
278
  active: "active";
425
279
  deprecated: "deprecated";
@@ -782,165 +636,194 @@ type AgentResourceEntry = z.infer<typeof AgentResourceEntrySchema$1>;
782
636
  type ResourceEntry = z.infer<typeof ResourceEntrySchema>;
783
637
 
784
638
  /**
785
- * Shared form field types for dynamic form generation
786
- * Used by: Command Queue, Execution Runner UI, future form-based features
787
- */
788
- /**
789
- * Supported form field types for action payloads
790
- * Maps to Mantine form components
791
- */
792
- type FormFieldType = 'text' | 'textarea' | 'number' | 'select' | 'checkbox' | 'radio' | 'richtext';
793
- /**
794
- * Form field definition
639
+ * Memory type definitions
640
+ * Types for agent memory management with semantic entry types
795
641
  */
796
- interface FormField {
797
- /** Field key in payload object */
798
- name: string;
799
- /** Field label for UI */
800
- label: string;
801
- /** Field type (determines UI component) */
802
- type: FormFieldType;
803
- /** Default value */
804
- defaultValue?: unknown;
805
- /** Required field */
806
- required?: boolean;
807
- /** Placeholder text */
808
- placeholder?: string;
809
- /** Help text */
810
- description?: string;
811
- /** Options for select/radio */
812
- options?: Array<{
813
- label: string;
814
- value: string | number;
815
- }>;
816
- /** Min/max for number */
817
- min?: number;
818
- max?: number;
819
- /** Path to context value for pre-filling (dot notation, e.g., 'proposal.summary') */
820
- defaultValueFromContext?: string;
821
- }
822
642
  /**
823
- * Form schema for action payload collection
643
+ * Semantic memory entry types
644
+ * Use-case agnostic types that describe the purpose of each entry
645
+ * Memory types mirror action types for clarity and filtering
824
646
  */
825
- interface FormSchema {
826
- /** Form title */
827
- title?: string;
828
- /** Form description */
829
- description?: string;
830
- /** Form fields */
831
- fields: FormField[];
832
- }
833
-
647
+ type MemoryEntryType = 'context' | 'input' | 'reasoning' | 'tool-result' | 'error';
834
648
  /**
835
- * Execution interface configuration
836
- * Defines how a resource is executed via the UI (forms, scheduling, webhooks)
837
- * Applies to both agents and workflows
649
+ * Who authored an entry's content.
650
+ *
651
+ * This is what lets the assembled prompt tell framework-authored text apart from text that
652
+ * originated outside the trust boundary. `'framework'` content is ours; the other three are not
653
+ * and are rendered inside the JSON data envelope (see `MemoryManager.toContextParts`).
838
654
  */
839
- interface ExecutionInterface {
840
- /** Form configuration for execution inputs */
841
- form: ExecutionFormSchema;
842
- /** Optional: Schedule configuration */
843
- schedule?: ScheduleConfig;
844
- /** Optional: Webhook trigger configuration */
845
- webhook?: WebhookConfig;
846
- }
655
+ type MemoryEntrySource = 'framework' | 'user' | 'tool' | 'model';
847
656
  /**
848
- * Execution form schema
849
- * Extends FormSchema with execution-specific fields
657
+ * Memory entry - represents a single entry in agent memory
658
+ * Stored in agent memory, translated by adapters to vendor-specific formats
850
659
  */
851
- interface ExecutionFormSchema extends FormSchema {
660
+ interface MemoryEntry {
661
+ type: MemoryEntryType;
662
+ content: string;
663
+ timestamp: number;
664
+ turnNumber: number | null;
665
+ iterationNumber: number | null;
852
666
  /**
853
- * Field mappings to resource input schema
854
- * Maps form field names to contract input paths
855
- * If omitted, field names must match contract input keys exactly
667
+ * Provenance. **Optional on purpose** `undefined` means unknown, which is what every
668
+ * pre-existing snapshot and every not-yet-redeployed tenant bundle produces. Read sites MUST
669
+ * test `== null`, never `=== undefined`: the `inTurnScope` predicate in `manager.ts` is the
670
+ * cautionary precedent, where a `=== undefined` check silently dropped every `null`-stamped
671
+ * entry. `isMemoryEntry` is deliberately NOT tightened to require this field; doing so would
672
+ * make every stored snapshot fail validation, and `restoreSessionMemory` fails open by
673
+ * starting the agent with empty memory rather than throwing.
856
674
  */
857
- fieldMappings?: Record<string, string>;
675
+ source?: MemoryEntrySource;
858
676
  /**
859
- * Submit button configuration
860
- * Default: { label: 'Run', loadingLabel: 'Running...' }
677
+ * Which tool produced this entry. Set on `tool-result` entries so the model can tell N parallel
678
+ * results apart -- the framework instructs batching independent tool calls in one iteration, and
679
+ * an anonymous result is unattributable the moment two land in the same iteration. `addToolError`
680
+ * already carries this (folded into its `content` JSON); this is the same fact for the success
681
+ * path, carried as a real field instead of prose the caller has to parse back out.
861
682
  */
862
- submitButton?: {
863
- label?: string;
864
- loadingLabel?: string;
865
- confirmMessage?: string;
683
+ toolName?: string;
684
+ /**
685
+ * Present when `truncateContent` cut this entry's `content` to fit its token budget. A sibling
686
+ * field, never text appended into `content` -- the notice used to be spliced into the string
687
+ * itself, which could (and did) land inside a JSON string literal `truncateContent` had just cut
688
+ * open, breaking `JSON.parse` on the far end. Absent means never truncated.
689
+ */
690
+ truncated?: {
691
+ omittedTokens: number;
866
692
  };
693
+ /**
694
+ * Prompt-injection warning types found in `content`, screened once here -- when the entry is
695
+ * written -- instead of by re-scanning the whole accumulated envelope on every iteration it gets
696
+ * re-sent for (`screenRequest`'s `data-envelope` slot used to do exactly that). Empty array means
697
+ * screened and clean; `undefined` means never screened (entries that bypass `addToHistory`/`set`,
698
+ * or pre-existing snapshots from before this field existed).
699
+ */
700
+ warnings?: string[];
867
701
  }
868
702
  /**
869
- * Schedule configuration for automated execution
870
- */
871
- interface ScheduleConfig {
872
- /** Whether scheduling is enabled for this resource */
873
- enabled: boolean;
874
- /** Default schedule (cron expression) */
875
- defaultSchedule?: string;
876
- /** Allowed schedule patterns (if restricted) */
877
- allowedPatterns?: string[];
703
+ * Agent memory - Self-orchestrated memory with session + working storage
704
+ * Agent has full control over what persists, framework handles auto-compaction
705
+ */
706
+ interface AgentMemory {
707
+ /**
708
+ * Session memory - Persists for session/conversation duration
709
+ * Never auto-trimmed by framework
710
+ * Agent-managed key-value store for critical information
711
+ * Agent provides strings, framework wraps in MemoryEntry
712
+ */
713
+ sessionMemory: Record<string, MemoryEntry>;
714
+ /**
715
+ * Working memory - Execution history
716
+ * Automatically compacted by framework when needed
717
+ * Agent doesn't control compaction
718
+ */
719
+ history: MemoryEntry[];
878
720
  }
879
721
  /**
880
- * Webhook configuration for external triggers
722
+ * Memory status for agent awareness
881
723
  */
882
- interface WebhookConfig {
883
- /** Whether webhook trigger is enabled */
884
- enabled: boolean;
885
- /** Expected payload schema (for documentation) */
886
- payloadSchema?: unknown;
887
- }
888
-
889
- interface WorkflowConfig extends ResourceDefinition {
890
- type: 'workflow';
891
- /** OM descriptor backing canonical identity and governance metadata. */
892
- resource?: WorkflowResourceEntry;
893
- }
894
- interface WorkflowStepDefinition {
895
- id: string;
896
- name: string;
897
- description: string;
898
- }
899
- type StepHandler = (input: unknown, context: ExecutionContext) => Promise<unknown>;
900
- interface LinearNext {
901
- type: 'linear';
902
- target: string;
903
- }
904
- interface ConditionalNext {
905
- type: 'conditional';
906
- routes: Array<{
907
- condition: (data: unknown) => boolean;
908
- target: string;
909
- }>;
910
- default: string;
911
- }
912
- type NextConfig = LinearNext | ConditionalNext | null;
913
- interface WorkflowStep extends WorkflowStepDefinition {
914
- handler: StepHandler;
915
- inputSchema: z.ZodSchema;
916
- outputSchema: z.ZodSchema;
917
- next: NextConfig;
918
- }
919
- interface WorkflowDefinition {
920
- config: WorkflowConfig;
921
- contract: Contract;
922
- steps: Record<string, WorkflowStep>;
923
- entryPoint: string;
724
+ interface MemoryStatus {
725
+ sessionMemoryKeys: number;
726
+ sessionMemoryLimit: number;
727
+ sessionMemoryTokens: number;
728
+ sessionMemoryTokenLimit: number;
924
729
  /**
925
- * Metrics configuration for ROI calculations
926
- * Optional: Only needed if tracking automation savings
730
+ * History tokens as a percentage of `historyBudget` — history ALONE, not history plus session
731
+ * memory. It previously reported the combined total under this name, so session memory growth
732
+ * read as history pressure and triggered history compaction that could not relieve it.
927
733
  */
928
- metricsConfig?: ResourceMetricsConfig;
734
+ historyPercent: number;
929
735
  /**
930
- * Execution interface configuration (optional)
931
- * If provided, workflow appears in Execution Runner UI
736
+ * Tokens the history entries **in scope for the requested turn** occupy — the same set
737
+ * `toContextParts` puts in the envelope. Equal to `storedHistoryTokens` when `getStatus` is
738
+ * called without a turn.
739
+ *
740
+ * This is the number the model is shown, and it is scoped because the model is handed a scoped
741
+ * set. Counting the whole cross-turn array here meant the framing quoted the size of a store
742
+ * while the envelope beside it carried one turn's worth of it.
932
743
  */
933
- interface?: ExecutionInterface;
744
+ historyTokens: number;
934
745
  /**
935
- * Lead-gen processing stage this workflow implements (optional).
936
- * Must match a key in the platform lead-gen stage catalog.
937
- * Used by org-os graph derivation to surface workflow→stage edges and
938
- * by pipeline_config validation to confirm each catalog stage has an
939
- * implementing workflow before a list is activated.
746
+ * Tokens the **entire** history array occupies, across every turn the session snapshot restored.
940
747
  *
941
- * Example: stageImplemented: 'verified' on the email-verification workflow.
748
+ * This is what compaction measures, because compaction trims that array. Scoping it to a turn
749
+ * would let the store grow without bound whenever the current turn happened to be small.
942
750
  */
943
- stageImplemented?: string;
751
+ storedHistoryTokens: number;
752
+ /** `storedHistoryTokens` as a percentage of `historyBudget`. The auto-compaction trigger. */
753
+ storedHistoryPercent: number;
754
+ historyBudget: number;
755
+ }
756
+ /**
757
+ * Memory constraints (optional limits)
758
+ */
759
+ interface MemoryConstraints {
760
+ maxSessionMemoryKeys?: number;
761
+ maxMemoryTokens?: number;
762
+ }
763
+
764
+ /**
765
+ * Types for the schema compiler. `compile.ts` walks a `JsonSchema` once, driven entirely by a
766
+ * `ProviderDialect`, and every server adapter compiles through it.
767
+ */
768
+ /**
769
+ * What happened to `strict` on a request, recorded per call rather than inferred.
770
+ *
771
+ * `applied` and `notAttempted` are the two states that a refusal-only field cannot tell apart --
772
+ * both leave `strictRefusalReasons` empty. Recording the verdict positively is what makes "was
773
+ * this agent's output actually enforced?" answerable from an `ai_calls` row.
774
+ */
775
+ type StrictStatus = 'applied' | 'refused' | 'compileRejected' | 'notAttempted';
776
+ /**
777
+ * A JSON Schema node, typed enough to be useful without pretending to validate the spec.
778
+ *
779
+ * The compiler has to accept schemas that arrive OUTSIDE the strict subset (that is the whole
780
+ * point of a dialect that can refuse or rewrite them) as well as the `$ref`/`$defs`/`const`/
781
+ * `$schema` shapes the strict subset has no vocabulary for at all. The index signature exists
782
+ * because tenant schemas carry keywords (`minLength`, `pattern`, `minimum`, ...) this compiler
783
+ * drops or refuses on, and they still need somewhere to type-check while they pass through
784
+ * `Object.entries`.
785
+ */
786
+ interface JsonSchema {
787
+ type?: string | string[];
788
+ /**
789
+ * The value is `JsonSchema | undefined`, not `JsonSchema`, because a property really can be
790
+ * declared with nothing describing it. `buildIterationResponseSchema` emits one per tool as
791
+ * `input: tool.inputSchema`, and `ToolDefinition.inputSchema` is typed `unknown` -- a tool
792
+ * deployed without one puts `undefined` under a key that exists.
793
+ *
794
+ * Both readers already handle it: `compileSchema` passes each value through `convertNode`, which
795
+ * takes `unknown`, and `collectErrors` opens with `if (!schema || typeof schema !== 'object')`
796
+ * above a comment naming this exact case. Declaring the value non-optional only hid that they
797
+ * were right to.
798
+ */
799
+ properties?: Record<string, JsonSchema | undefined>;
800
+ items?: JsonSchema;
801
+ anyOf?: JsonSchema[];
802
+ oneOf?: JsonSchema[];
803
+ allOf?: JsonSchema[];
804
+ required?: string[];
805
+ additionalProperties?: boolean | JsonSchema;
806
+ minItems?: number;
807
+ maxItems?: number;
808
+ format?: string;
809
+ enum?: unknown[];
810
+ const?: unknown;
811
+ description?: string;
812
+ default?: unknown;
813
+ $ref?: string;
814
+ $defs?: Record<string, JsonSchema>;
815
+ definitions?: Record<string, JsonSchema>;
816
+ $schema?: string;
817
+ $id?: string;
818
+ $anchor?: string;
819
+ /**
820
+ * OpenAPI's nullability spelling, which is not JSON Schema's. It is declared because
821
+ * `response-schema-validator.ts` READS it (`schemaPermitsNull`) -- Google's schema dialect is
822
+ * OpenAPI-derived, so a schema that reaches the validator can carry it. No dialect in `compile.ts`
823
+ * writes or rewrites it; the canonical spelling this compiler emits is `type: ['x', 'null']`.
824
+ */
825
+ nullable?: boolean;
826
+ [key: string]: unknown;
944
827
  }
945
828
 
946
829
  /**
@@ -955,6 +838,31 @@ interface WorkflowDefinition {
955
838
  interface LLMMessage {
956
839
  role: 'system' | 'user' | 'assistant';
957
840
  content: string;
841
+ /**
842
+ * Marks this message as the end of a byte-stable prefix worth an Anthropic cache breakpoint,
843
+ * beyond the one the system prompt already gets. Anthropic's rule is "everything up to and
844
+ * including the marked block is cached", so this only ever needs to sit on ONE message -- the
845
+ * last one before content that changes.
846
+ *
847
+ * `buildAgentMessages` sets it on the last replayed prior-turn message: conversation history is
848
+ * fixed for the whole turn (only the framing/envelope after it grow per iteration), so it is the
849
+ * only part of a session agent's messages, besides the system prompt, that is ever byte-identical
850
+ * call to call. A hint rather than a mechanism deliberately -- an adapter that does not read it
851
+ * (OpenAI, OpenRouter, any test stub) just ignores the extra property; only the Anthropic adapter
852
+ * turns it into a wire `cache_control` block.
853
+ */
854
+ cacheBreakpoint?: boolean;
855
+ /**
856
+ * Prompt-injection warning types already found in this message's content, when the caller has
857
+ * already screened it and wants `screenRequest` to use that verdict instead of re-scanning.
858
+ *
859
+ * Set only on the data-envelope message by `buildAgentMessages`, sourced from
860
+ * `MemoryContextParts.envelopeWarnings` -- itself an aggregate of `MemoryEntry.warnings` stamped
861
+ * once per fragment when it entered memory. `undefined` means "not pre-screened"; `screenRequest`
862
+ * falls back to scanning the content directly, which is what every other message role/slot still
863
+ * does and what a hand-built message (tests, other callers) gets by default.
864
+ */
865
+ envelopeWarnings?: string[];
958
866
  }
959
867
  /**
960
868
  * Generic LLM generation request
@@ -1120,112 +1028,84 @@ interface LLMAdapter {
1120
1028
  }
1121
1029
 
1122
1030
  /**
1123
- * Memory type definitions
1124
- * Types for agent memory management with semantic entry types
1031
+ * Model Configuration
1032
+ * Centralized model information, configuration, options, constraints, and validation
1033
+ * Single source of truth for all model-related definitions
1034
+ * Update manually when pricing changes or new models are added
1125
1035
  */
1036
+
1126
1037
  /**
1127
- * Semantic memory entry types
1128
- * Use-case agnostic types that describe the purpose of each entry
1129
- * Memory types mirror action types for clarity and filtering
1038
+ * Supported Open AI models (direct SDK access)
1130
1039
  */
1131
- type MemoryEntryType = 'context' | 'input' | 'reasoning' | 'tool-result' | 'error';
1040
+ type OpenAIModel = 'gpt-5' | 'gpt-5.4-mini' | 'gpt-5.4-nano';
1132
1041
  /**
1133
- * Who authored an entry's content.
1134
- *
1135
- * This is what lets the assembled prompt tell framework-authored text apart from text that
1136
- * originated outside the trust boundary. `'framework'` content is ours; the other three are not
1137
- * and are rendered inside the JSON data envelope (see `MemoryManager.toContextParts`).
1042
+ * Supported OpenRouter models (explicit union for type safety)
1138
1043
  */
1139
- type MemoryEntrySource = 'framework' | 'user' | 'tool' | 'model';
1044
+ type OpenRouterModel = 'openrouter/z-ai/glm-5';
1140
1045
  /**
1141
- * Memory entry - represents a single entry in agent memory
1142
- * Stored in agent memory, translated by adapters to vendor-specific formats
1046
+ * Supported Anthropic models (direct SDK access via @anthropic-ai/sdk)
1143
1047
  */
1144
- interface MemoryEntry {
1145
- type: MemoryEntryType;
1146
- content: string;
1147
- timestamp: number;
1148
- turnNumber: number | null;
1149
- iterationNumber: number | null;
1150
- /**
1151
- * Provenance. **Optional on purpose** — `undefined` means unknown, which is what every
1152
- * pre-existing snapshot and every not-yet-redeployed tenant bundle produces. Read sites MUST
1153
- * test `== null`, never `=== undefined`: the `inTurnScope` predicate in `manager.ts` is the
1154
- * cautionary precedent, where a `=== undefined` check silently dropped every `null`-stamped
1155
- * entry. `isMemoryEntry` is deliberately NOT tightened to require this field; doing so would
1156
- * make every stored snapshot fail validation, and `restoreSessionMemory` fails open by
1157
- * starting the agent with empty memory rather than throwing.
1158
- */
1159
- source?: MemoryEntrySource;
1160
- /**
1161
- * Which tool produced this entry. Set on `tool-result` entries so the model can tell N parallel
1162
- * results apart -- the framework instructs batching independent tool calls in one iteration, and
1163
- * an anonymous result is unattributable the moment two land in the same iteration. `addToolError`
1164
- * already carries this (folded into its `content` JSON); this is the same fact for the success
1165
- * path, carried as a real field instead of prose the caller has to parse back out.
1166
- */
1167
- toolName?: string;
1168
- }
1048
+ type AnthropicModel = 'claude-opus-5' | 'claude-sonnet-5' | 'claude-haiku-4-5-20251001' | 'claude-haiku-4-5';
1049
+ /** Supported LLM models */
1050
+ type LLMModel = OpenAIModel | OpenRouterModel | AnthropicModel | 'mock';
1169
1051
  /**
1170
- * Agent memory - Self-orchestrated memory with session + working storage
1171
- * Agent has full control over what persists, framework handles auto-compaction
1052
+ * GPT-5 model options schema
1172
1053
  */
1173
- interface AgentMemory {
1174
- /**
1175
- * Session memory - Persists for session/conversation duration
1176
- * Never auto-trimmed by framework
1177
- * Agent-managed key-value store for critical information
1178
- * Agent provides strings, framework wraps in MemoryEntry
1179
- */
1180
- sessionMemory: Record<string, MemoryEntry>;
1181
- /**
1182
- * Working memory - Execution history
1183
- * Automatically compacted by framework when needed
1184
- * Agent doesn't control compaction
1185
- */
1186
- history: MemoryEntry[];
1187
- }
1054
+ declare const GPT5OptionsSchema: z.ZodObject<{
1055
+ reasoning_effort: z.ZodOptional<z.ZodEnum<{
1056
+ minimal: "minimal";
1057
+ low: "low";
1058
+ medium: "medium";
1059
+ high: "high";
1060
+ }>>;
1061
+ verbosity: z.ZodOptional<z.ZodEnum<{
1062
+ low: "low";
1063
+ medium: "medium";
1064
+ high: "high";
1065
+ }>>;
1066
+ }, z.core.$strip>;
1188
1067
  /**
1189
- * Memory status for agent awareness
1068
+ * OpenRouter model options schema
1069
+ * OpenRouter-specific options for routing and transforms
1190
1070
  */
1191
- interface MemoryStatus {
1192
- sessionMemoryKeys: number;
1193
- sessionMemoryLimit: number;
1194
- sessionMemoryTokens: number;
1195
- sessionMemoryTokenLimit: number;
1196
- /**
1197
- * History tokens as a percentage of `historyBudget` — history ALONE, not history plus session
1198
- * memory. It previously reported the combined total under this name, so session memory growth
1199
- * read as history pressure and triggered history compaction that could not relieve it.
1200
- */
1201
- historyPercent: number;
1202
- /**
1203
- * Tokens the history entries **in scope for the requested turn** occupy — the same set
1204
- * `toContextParts` puts in the envelope. Equal to `storedHistoryTokens` when `getStatus` is
1205
- * called without a turn.
1206
- *
1207
- * This is the number the model is shown, and it is scoped because the model is handed a scoped
1208
- * set. Counting the whole cross-turn array here meant the framing quoted the size of a store
1209
- * while the envelope beside it carried one turn's worth of it.
1210
- */
1211
- historyTokens: number;
1212
- /**
1213
- * Tokens the **entire** history array occupies, across every turn the session snapshot restored.
1214
- *
1215
- * This is what compaction measures, because compaction trims that array. Scoping it to a turn
1216
- * would let the store grow without bound whenever the current turn happened to be small.
1217
- */
1218
- storedHistoryTokens: number;
1219
- /** `storedHistoryTokens` as a percentage of `historyBudget`. The auto-compaction trigger. */
1220
- storedHistoryPercent: number;
1221
- historyBudget: number;
1222
- }
1071
+ declare const OpenRouterOptionsSchema: z.ZodObject<{
1072
+ transforms: z.ZodOptional<z.ZodArray<z.ZodString>>;
1073
+ route: z.ZodOptional<z.ZodEnum<{
1074
+ fallback: "fallback";
1075
+ }>>;
1076
+ }, z.core.$strip>;
1223
1077
  /**
1224
- * Memory constraints (optional limits)
1078
+ * Anthropic model options schema
1079
+ * Currently empty - future options must be added per supported model family
1225
1080
  */
1226
- interface MemoryConstraints {
1227
- maxSessionMemoryKeys?: number;
1228
- maxMemoryTokens?: number;
1081
+ declare const AnthropicOptionsSchema: z.ZodObject<{}, z.core.$strict>;
1082
+ /**
1083
+ * Infer TypeScript types from schemas
1084
+ */
1085
+ type GPT5Options = z.infer<typeof GPT5OptionsSchema>;
1086
+ type MockOptions = Record<string, never>;
1087
+ type OpenRouterOptions = z.infer<typeof OpenRouterOptionsSchema>;
1088
+ type AnthropicOptions = z.infer<typeof AnthropicOptionsSchema>;
1089
+ type ModelSpecificOptions = GPT5Options | MockOptions | OpenRouterOptions | AnthropicOptions;
1090
+ /**
1091
+ * Model configuration for LLM execution
1092
+ * Belongs in resource definition (AgentDefinition, WorkflowDefinition, etc.)
1093
+ */
1094
+ interface ModelConfig {
1095
+ model: LLMModel;
1096
+ provider: 'openai' | 'anthropic' | 'openrouter' | 'mock';
1097
+ apiKey: string;
1098
+ temperature?: number;
1099
+ /** Maximum output tokens per LLM call. NOT the model's context window — see ModelInfo.maxTokens for that. */
1100
+ maxOutputTokens?: number;
1101
+ topP?: number;
1102
+ /**
1103
+ * Model-specific options (flat structure)
1104
+ * Options are model-specific, not vendor-specific
1105
+ * Available options defined in MODEL_INFO per model
1106
+ * Validated at build time via validateModelOptions()
1107
+ */
1108
+ modelOptions?: ModelSpecificOptions;
1229
1109
  }
1230
1110
 
1231
1111
  /**
@@ -1245,8 +1125,27 @@ interface MemoryConstraints {
1245
1125
  interface MemoryContextParts {
1246
1126
  /** Framework-authored. Memory status and a description of the envelope. Carries NO stored content. */
1247
1127
  framing: string;
1248
- /** Every stored fragment, JSON-encoded and source-tagged. Untrusted. */
1128
+ /** Every stored fragment, JSON-encoded. Untrusted. */
1249
1129
  dataEnvelope: string;
1130
+ /**
1131
+ * Union of prompt-injection warning types already found across every fragment `dataEnvelope`
1132
+ * actually carries this call, aggregated from verdicts stamped once when each fragment entered
1133
+ * memory (see `MemoryEntry.warnings`) rather than by re-scanning `dataEnvelope`'s text on every
1134
+ * iteration it gets rebuilt for. Elided fragments (see `ENVELOPE_FULL_RESULT_WINDOW`) contribute
1135
+ * nothing here — their original content isn't what gets sent once they're stubbed.
1136
+ *
1137
+ * This is metadata about the envelope, not part of it: folding a detector's own finding into the
1138
+ * model-visible JSON would hand a would-be attacker — plausibly the same person on the other end
1139
+ * of a session conversation — direct feedback on which pattern tripped. A caller wiring this up
1140
+ * (`screenRequest`'s `data-envelope` slot is the one that currently re-scans instead of reading
1141
+ * this) should treat it exactly the way `screenRequest` already treats cross-turn history: it
1142
+ * warns, but whether it blocks is that caller's decision to make, not this one's.
1143
+ *
1144
+ * Optional (not just possibly-empty): the fixture literals in `agent/reasoning/**` tests build
1145
+ * `MemoryContextParts` by hand without it, and requiring it would make this signature's landing
1146
+ * a forced edit across files this change does not otherwise touch.
1147
+ */
1148
+ envelopeWarnings?: string[];
1250
1149
  }
1251
1150
  /**
1252
1151
  * Memory Manager - Agent memory orchestration
@@ -1258,7 +1157,41 @@ declare class MemoryManager {
1258
1157
  private constraints;
1259
1158
  private logger?;
1260
1159
  private cachedSnapshot?;
1160
+ /**
1161
+ * Rolling correction for `estimateTokens`'s bias, learned from real provider usage.
1162
+ * `undefined` until the first `recordActualUsage` call -- the cold-start state, where
1163
+ * `estimate()` returns the raw `estimateTokens` output unscaled. See `recordActualUsage`.
1164
+ */
1165
+ private tokenCorrectionFactor?;
1261
1166
  constructor(memory: AgentMemory, constraints?: MemoryConstraints, logger?: AgentScopedLogger | undefined);
1167
+ /**
1168
+ * Record how far `estimateTokens` was from reality on a real provider call, and roll it into a
1169
+ * correction applied to every estimate this instance makes from here on -- `getStatus`'s three
1170
+ * token fields and `enforceSessionMemoryTokenLimit`'s eviction check, which is what
1171
+ * `autoCompact`/`enforceHardLimits` actually decide compaction from (C3 / Wave M3).
1172
+ *
1173
+ * `estimateTokens` is `chars / 3.5` -- a constant-ratio guess with no knowledge of JSON escaping,
1174
+ * key overhead, or real tokenizer behaviour. Every provider call already returns an EXACT count
1175
+ * (`usage.inputTokens`) that reaches `ai_calls` and is then dropped; this is where it stops being
1176
+ * dropped, without replacing the estimator outright -- a cold session still needs SOME number
1177
+ * before its first real call completes, so the estimator stays the prior and this only corrects
1178
+ * it once real data exists.
1179
+ *
1180
+ * `estimatedRequestTokens` must be `estimateTokens` applied to the SAME text `actualInputTokens`
1181
+ * was billed for -- the whole assembled request (system prompt, tools, conversation history, the
1182
+ * envelope, everything), not just what this class itself emits. `estimateTokens`'s bias is a
1183
+ * property of the heuristic, not of which slice of the request it is pointed at, so measuring it
1184
+ * against the full request (visible to the caller, not to this class) and applying the result to
1185
+ * this class's own estimates (which can only ever see its own slice) is a fair trade -- one ratio,
1186
+ * calibrated on real data, standing in for a per-segment breakdown nothing needs.
1187
+ *
1188
+ * Exponential moving average, not a straight replace: a single call's ratio is noisy, and a
1189
+ * straight replace lets one outlier swing every compaction decision made afterward. Each new
1190
+ * observation gets 30% weight, converging within a handful of calls without chasing one spike.
1191
+ */
1192
+ recordActualUsage(estimatedRequestTokens: number, actualInputTokens: number): void;
1193
+ /** `estimateTokens`, scaled by the learned correction once one exists. See `recordActualUsage`. */
1194
+ private estimate;
1262
1195
  /**
1263
1196
  * Set session memory entry (agent provides string, framework wraps it)
1264
1197
  * @param key - Session memory key
@@ -1354,7 +1287,15 @@ declare class MemoryManager {
1354
1287
  * treat "everything in this block" as data was also being handed the live question inside that
1355
1288
  * block.
1356
1289
  *
1357
- * Shows current iteration entries FIRST (reverse chronological) for LLM attention.
1290
+ * History entries stay chronological. They used to be split into a "current iteration" slot
1291
+ * (reverse chronological, for LLM positional bias) and an "earlier" slot -- but the LLM call
1292
+ * always happens BEFORE `addToHistory` writes that iteration's own entries, so the
1293
+ * current-iteration slot held nothing on any call that mattered. One chronological list replaces
1294
+ * both.
1295
+ *
1296
+ * Tool results (and tool errors) older than `ENVELOPE_FULL_RESULT_WINDOW` iterations are carried
1297
+ * as a short stub instead of their full content -- see `ENVELOPE_FULL_RESULT_WINDOW`. The STORE
1298
+ * (`this.memory.history`) is untouched; only what this call carries is capped.
1358
1299
  *
1359
1300
  * @param currentIteration - Current iteration number (0 = pre-iteration)
1360
1301
  * @param currentTurn - Current turn number (optional, for session context filtering)
@@ -1363,89 +1304,145 @@ declare class MemoryManager {
1363
1304
  }
1364
1305
 
1365
1306
  /**
1366
- * Agent-specific type definitions
1367
- * Types for autonomous agents with tools, memory, and constraints
1307
+ * Shared form field types for dynamic form generation
1308
+ * Used by: Command Queue, Execution Runner UI, future form-based features
1368
1309
  */
1310
+ /**
1311
+ * Supported form field types for action payloads
1312
+ * Maps to Mantine form components
1313
+ */
1314
+ type FormFieldType = 'text' | 'textarea' | 'number' | 'select' | 'checkbox' | 'radio' | 'richtext';
1315
+ /**
1316
+ * Form field definition
1317
+ */
1318
+ interface FormField {
1319
+ /** Field key in payload object */
1320
+ name: string;
1321
+ /** Field label for UI */
1322
+ label: string;
1323
+ /** Field type (determines UI component) */
1324
+ type: FormFieldType;
1325
+ /** Default value */
1326
+ defaultValue?: unknown;
1327
+ /** Required field */
1328
+ required?: boolean;
1329
+ /** Placeholder text */
1330
+ placeholder?: string;
1331
+ /** Help text */
1332
+ description?: string;
1333
+ /** Options for select/radio */
1334
+ options?: Array<{
1335
+ label: string;
1336
+ value: string | number;
1337
+ }>;
1338
+ /** Min/max for number */
1339
+ min?: number;
1340
+ max?: number;
1341
+ /** Path to context value for pre-filling (dot notation, e.g., 'proposal.summary') */
1342
+ defaultValueFromContext?: string;
1343
+ }
1344
+ /**
1345
+ * Form schema for action payload collection
1346
+ */
1347
+ interface FormSchema {
1348
+ /** Form title */
1349
+ title?: string;
1350
+ /** Form description */
1351
+ description?: string;
1352
+ /** Form fields */
1353
+ fields: FormField[];
1354
+ }
1369
1355
 
1370
1356
  /**
1371
- * Factory function for creating LLM adapters.
1372
- * Injected into the Agent class to decouple the engine from server-only provider SDKs.
1373
- * - API process: provides createLLMAdapter (real SDKs + process.env API keys)
1374
- * - SDK worker: provides PostMessageLLMAdapter (proxies via platform.call)
1375
- *
1376
- * Uses `any` for optional params so both the real createLLMAdapter (with typed
1377
- * AIUsageCollector/AICallContext) and the worker proxy (which ignores them) satisfy the type.
1357
+ * Execution interface configuration
1358
+ * Defines how a resource is executed via the UI (forms, scheduling, webhooks)
1359
+ * Applies to both agents and workflows
1378
1360
  */
1379
- type LLMAdapterFactory = (config: ModelConfig, ...args: any[]) => LLMAdapter;
1380
- type AgentKind = 'orchestrator' | 'specialist' | 'utility' | 'platform';
1381
- interface AgentConfig extends ResourceDefinition {
1382
- type: 'agent';
1383
- /** OM descriptor backing canonical identity and governance metadata. */
1384
- resource?: AgentResourceEntry;
1385
- kind: AgentKind;
1386
- systemPrompt: string;
1387
- constraints?: AgentConstraints;
1388
- /**
1389
- * Session capability declaration (opt-in)
1390
- * If true, agent is designed for multi-turn session interactions
1391
- * Controls whether agent can use message action and appears in Sessions UI
1392
- *
1393
- * Use for:
1394
- * - Conversational agents with multi-turn interactions
1395
- * - Agents requiring persistent context across turns
1396
- * - Agents that need human-in-the-loop communication
1397
- */
1398
- sessionCapable?: boolean;
1361
+ interface ExecutionInterface {
1362
+ /** Form configuration for execution inputs */
1363
+ form: ExecutionFormSchema;
1364
+ /** Optional: Schedule configuration */
1365
+ schedule?: ScheduleConfig;
1366
+ /** Optional: Webhook trigger configuration */
1367
+ webhook?: WebhookConfig;
1368
+ }
1369
+ /**
1370
+ * Execution form schema
1371
+ * Extends FormSchema with execution-specific fields
1372
+ */
1373
+ interface ExecutionFormSchema extends FormSchema {
1399
1374
  /**
1400
- * Security level for system prompt hardening (auto-derived if omitted)
1401
- *
1402
- * - 'standard': Lightweight defense (3 rules) - default for non-session agents
1403
- * - 'hardened': Comprehensive defense (5 rules) - default for session-capable agents
1404
- * - 'none': No security prompt - for pure internal agents with no external input
1405
- *
1406
- * If omitted, derived from sessionCapable:
1407
- * sessionCapable: true -> 'hardened'
1408
- * sessionCapable: false -> 'standard'
1375
+ * Field mappings to resource input schema
1376
+ * Maps form field names to contract input paths
1377
+ * If omitted, field names must match contract input keys exactly
1409
1378
  */
1410
- securityLevel?: 'standard' | 'hardened' | 'none';
1379
+ fieldMappings?: Record<string, string>;
1411
1380
  /**
1412
- * Memory management preferences (opt-in)
1413
- * If provided, agent can use memoryOps to manage session memory
1414
- * If omitted, agent has no memory management capabilities
1415
- *
1416
- * Agent-specific guidance on what to preserve, when to persist, and what to clean up.
1417
- * This guidance is injected into the system prompt when memory management is enabled.
1418
- *
1419
- * Use for:
1420
- * - Conversational agents needing cross-turn context
1421
- * - Agents managing complex user preferences
1422
- * - Agents tracking decisions over multiple iterations
1381
+ * Submit button configuration
1382
+ * Default: { label: 'Run', loadingLabel: 'Running...' }
1423
1383
  */
1424
- memoryPreferences?: string;
1384
+ submitButton?: {
1385
+ label?: string;
1386
+ loadingLabel?: string;
1387
+ confirmMessage?: string;
1388
+ };
1389
+ }
1390
+ /**
1391
+ * Schedule configuration for automated execution
1392
+ */
1393
+ interface ScheduleConfig {
1394
+ /** Whether scheduling is enabled for this resource */
1395
+ enabled: boolean;
1396
+ /** Default schedule (cron expression) */
1397
+ defaultSchedule?: string;
1398
+ /** Allowed schedule patterns (if restricted) */
1399
+ allowedPatterns?: string[];
1425
1400
  }
1426
- interface AgentConstraints {
1427
- maxIterations?: number;
1428
- timeout?: number;
1429
- maxSessionMemoryKeys?: number;
1430
- maxMemoryTokens?: number;
1401
+ /**
1402
+ * Webhook configuration for external triggers
1403
+ */
1404
+ interface WebhookConfig {
1405
+ /** Whether webhook trigger is enabled */
1406
+ enabled: boolean;
1407
+ /** Expected payload schema (for documentation) */
1408
+ payloadSchema?: unknown;
1431
1409
  }
1432
- interface AgentDefinition {
1433
- config: AgentConfig;
1410
+
1411
+ interface WorkflowConfig extends ResourceDefinition {
1412
+ type: 'workflow';
1413
+ /** OM descriptor backing canonical identity and governance metadata. */
1414
+ resource?: WorkflowResourceEntry;
1415
+ }
1416
+ interface WorkflowStepDefinition {
1417
+ id: string;
1418
+ name: string;
1419
+ description: string;
1420
+ }
1421
+ type StepHandler = (input: unknown, context: ExecutionContext) => Promise<unknown>;
1422
+ interface LinearNext {
1423
+ type: 'linear';
1424
+ target: string;
1425
+ }
1426
+ interface ConditionalNext {
1427
+ type: 'conditional';
1428
+ routes: Array<{
1429
+ condition: (data: unknown) => boolean;
1430
+ target: string;
1431
+ }>;
1432
+ default: string;
1433
+ }
1434
+ type NextConfig = LinearNext | ConditionalNext | null;
1435
+ interface WorkflowStep extends WorkflowStepDefinition {
1436
+ handler: StepHandler;
1437
+ inputSchema: z.ZodSchema;
1438
+ outputSchema: z.ZodSchema;
1439
+ next: NextConfig;
1440
+ }
1441
+ interface WorkflowDefinition {
1442
+ config: WorkflowConfig;
1434
1443
  contract: Contract;
1435
- tools: Tool[];
1436
- /**
1437
- * Model configuration for LLM execution
1438
- * Specifies provider, API key, and model-specific options
1439
- */
1440
- modelConfig: ModelConfig;
1441
- /**
1442
- * Preload memory before execution starts
1443
- * Handles BOTH context loading AND session restoration
1444
- *
1445
- * @param context - Execution context (includes sessionId if session turn)
1446
- * @returns Initial AgentMemory state (sessionMemory entries + optionally history)
1447
- */
1448
- preloadMemory?: (context: ExecutionContext) => Promise<AgentMemory> | AgentMemory;
1444
+ steps: Record<string, WorkflowStep>;
1445
+ entryPoint: string;
1449
1446
  /**
1450
1447
  * Metrics configuration for ROI calculations
1451
1448
  * Optional: Only needed if tracking automation savings
@@ -1453,30 +1450,19 @@ interface AgentDefinition {
1453
1450
  metricsConfig?: ResourceMetricsConfig;
1454
1451
  /**
1455
1452
  * Execution interface configuration (optional)
1456
- * If provided, agent appears in Execution Runner UI
1453
+ * If provided, workflow appears in Execution Runner UI
1457
1454
  */
1458
1455
  interface?: ExecutionInterface;
1459
- }
1460
- /**
1461
- * Agent execution context
1462
- * Groups all state needed for agent execution phases
1463
- */
1464
- interface IterationContext {
1465
- config: AgentConfig;
1466
- contract: Contract;
1467
- toolRegistry: Map<string, Tool>;
1468
- memoryManager: MemoryManager;
1469
- executionContext: ExecutionContext;
1470
- iteration: number;
1471
- logger: AgentScopedLogger;
1472
- modelConfig: ModelConfig;
1473
- adapterFactory: LLMAdapterFactory;
1474
1456
  /**
1475
- * The validated input for this execution, serialized. It travels here because the model gets
1476
- * it as its own `role:'user'` message; nothing else in this context carried it, so the input
1477
- * had to be read back out of memory history and shipped inside the memory block.
1457
+ * Lead-gen processing stage this workflow implements (optional).
1458
+ * Must match a key in the platform lead-gen stage catalog.
1459
+ * Used by org-os graph derivation to surface workflow→stage edges and
1460
+ * by pipeline_config validation to confirm each catalog stage has an
1461
+ * implementing workflow before a list is activated.
1462
+ *
1463
+ * Example: stageImplemented: 'verified' on the email-verification workflow.
1478
1464
  */
1479
- currentInput: string;
1465
+ stageImplemented?: string;
1480
1466
  }
1481
1467
 
1482
1468
  declare const OntologyScopeSchema: z.ZodDefault<z.ZodObject<{
@@ -2756,6 +2742,32 @@ declare const OrganizationModelSchema$1: z.ZodObject<{
2756
2742
 
2757
2743
  type OrganizationModel$1 = z.infer<typeof OrganizationModelSchema$1>;
2758
2744
 
2745
+ /**
2746
+ * AIUsageCollector
2747
+ * Centralized token tracking that aggregates usage across all LLM calls in an execution
2748
+ */
2749
+ declare class AIUsageCollector {
2750
+ private model;
2751
+ private calls;
2752
+ private callSequence;
2753
+ /**
2754
+ * Record a single AI call with usage metrics
2755
+ *
2756
+ * @param usage - Token usage and latency data from LLM adapter
2757
+ * @param callType - Type discriminator (agent-reasoning, tool, etc.)
2758
+ * @param context - Optional typed context specific to callType
2759
+ */
2760
+ record(usage: LLMUsageData, callType?: BaseAICall['callType'], context?: AICallContext): void;
2761
+ /**
2762
+ * Get aggregated summary of all AI calls
2763
+ */
2764
+ getSummary(): AIUsageSummary;
2765
+ /**
2766
+ * Check if any usage has been recorded
2767
+ */
2768
+ hasUsage(): boolean;
2769
+ }
2770
+
2759
2771
  /**
2760
2772
  * MetricsCollector
2761
2773
  * Tracks execution timing and ROI metrics
@@ -3014,29 +3026,147 @@ interface ResourceMetricsConfig {
3014
3026
  }
3015
3027
 
3016
3028
  /**
3017
- * AIUsageCollector
3018
- * Centralized token tracking that aggregates usage across all LLM calls in an execution
3029
+ * Agent-specific type definitions
3030
+ * Types for autonomous agents with tools, memory, and constraints
3019
3031
  */
3020
- declare class AIUsageCollector {
3021
- private model;
3022
- private calls;
3023
- private callSequence;
3032
+
3033
+ /**
3034
+ * Factory function for creating LLM adapters.
3035
+ * Injected into the Agent class to decouple the engine from server-only provider SDKs.
3036
+ * - API process: provides createLLMAdapter (real SDKs + process.env API keys)
3037
+ * - SDK worker: provides PostMessageLLMAdapter (proxies via platform.call)
3038
+ *
3039
+ * Uses `any` for optional params so both the real createLLMAdapter (with typed
3040
+ * AIUsageCollector/AICallContext) and the worker proxy (which ignores them) satisfy the type.
3041
+ */
3042
+ type LLMAdapterFactory = (config: ModelConfig, ...args: any[]) => LLMAdapter;
3043
+ type AgentKind = 'orchestrator' | 'specialist' | 'utility' | 'platform';
3044
+ interface AgentConfig extends ResourceDefinition {
3045
+ type: 'agent';
3046
+ /** OM descriptor backing canonical identity and governance metadata. */
3047
+ resource?: AgentResourceEntry;
3048
+ kind: AgentKind;
3049
+ systemPrompt: string;
3050
+ constraints?: AgentConstraints;
3024
3051
  /**
3025
- * Record a single AI call with usage metrics
3052
+ * Session capability declaration (opt-in)
3053
+ * If true, agent is designed for multi-turn session interactions
3054
+ * Controls whether agent can use message action and appears in Sessions UI
3026
3055
  *
3027
- * @param usage - Token usage and latency data from LLM adapter
3028
- * @param callType - Type discriminator (agent-reasoning, tool, etc.)
3029
- * @param context - Optional typed context specific to callType
3056
+ * Use for:
3057
+ * - Conversational agents with multi-turn interactions
3058
+ * - Agents requiring persistent context across turns
3059
+ * - Agents that need human-in-the-loop communication
3030
3060
  */
3031
- record(usage: LLMUsageData, callType?: BaseAICall['callType'], context?: AICallContext): void;
3061
+ sessionCapable?: boolean;
3032
3062
  /**
3033
- * Get aggregated summary of all AI calls
3063
+ * Overrides the default `message` requiredness for a session-capable agent (ignored for
3064
+ * non-session agents, which always get `AgentCapabilities.message: 'off'`). Defaults to
3065
+ * `'required'` -- see `AgentCapabilities.message`'s doc comment for why. Set `'optional'` only
3066
+ * when the agent legitimately needs tool-only turns with no reply, and the deploy target can
3067
+ * tolerate the blind-retry risk `validateResponseSchema` carries on any path where the schema is
3068
+ * not compiled into a sampling grammar.
3034
3069
  */
3035
- getSummary(): AIUsageSummary;
3070
+ messagePolicy?: 'optional' | 'required';
3036
3071
  /**
3037
- * Check if any usage has been recorded
3072
+ * Explicit opt-in to skip the iteration loop and produce `contract.outputSchema`-shaped output in
3073
+ * a single LLM call (round 3 decision B6: explicit opt-in, never inferred from `kind`,
3074
+ * `sessionCapable`, or tool count -- so no existing agent changes shape by default). Structurally
3075
+ * the normal path pays two calls minimum: `iterate()` always runs at least one, and `complete()`
3076
+ * runs a second whose prompt re-derives the answer from history rather than reading what the
3077
+ * iteration already decided. A single-shot classifier -- one input in, one structured output out,
3078
+ * no multi-step reasoning needed -- does not need that second derivation; `complete()` already
3079
+ * makes exactly the one call it needs, from `currentInput` directly.
3080
+ *
3081
+ * Requires `sessionCapable` to be falsy and `contract.outputSchema` to be present. `Agent`
3082
+ * validates both during initialization and throws `AgentInitializationError` if either is missing,
3083
+ * rather than silently falling back to the normal two-call path on a misconfigured opt-in. Tools
3084
+ * registered on the agent are never invoked in this path -- there is no iteration loop to call
3085
+ * them from, so an agent that needs tool calls before it can answer is not eligible regardless of
3086
+ * this flag.
3038
3087
  */
3039
- hasUsage(): boolean;
3088
+ singleShot?: boolean;
3089
+ /**
3090
+ * Security level for system prompt hardening (auto-derived if omitted)
3091
+ *
3092
+ * - 'standard': Lightweight defense (3 rules) - default for non-session agents
3093
+ * - 'hardened': Comprehensive defense (5 rules) - default for session-capable agents
3094
+ * - 'none': No security prompt - for pure internal agents with no external input
3095
+ *
3096
+ * If omitted, derived from sessionCapable:
3097
+ * sessionCapable: true -> 'hardened'
3098
+ * sessionCapable: false -> 'standard'
3099
+ */
3100
+ securityLevel?: 'standard' | 'hardened' | 'none';
3101
+ /**
3102
+ * Memory management preferences (opt-in)
3103
+ * If provided, agent can use memoryOps to manage session memory
3104
+ * If omitted, agent has no memory management capabilities
3105
+ *
3106
+ * Agent-specific guidance on what to preserve, when to persist, and what to clean up.
3107
+ * This guidance is injected into the system prompt when memory management is enabled.
3108
+ *
3109
+ * Use for:
3110
+ * - Conversational agents needing cross-turn context
3111
+ * - Agents managing complex user preferences
3112
+ * - Agents tracking decisions over multiple iterations
3113
+ */
3114
+ memoryPreferences?: string;
3115
+ }
3116
+ interface AgentConstraints {
3117
+ maxIterations?: number;
3118
+ timeout?: number;
3119
+ maxSessionMemoryKeys?: number;
3120
+ maxMemoryTokens?: number;
3121
+ }
3122
+ interface AgentDefinition {
3123
+ config: AgentConfig;
3124
+ contract: Contract;
3125
+ tools: Tool[];
3126
+ /**
3127
+ * Model configuration for LLM execution
3128
+ * Specifies provider, API key, and model-specific options
3129
+ */
3130
+ modelConfig: ModelConfig;
3131
+ /**
3132
+ * Preload memory before execution starts
3133
+ * Handles BOTH context loading AND session restoration
3134
+ *
3135
+ * @param context - Execution context (includes sessionId if session turn)
3136
+ * @returns Initial AgentMemory state (sessionMemory entries + optionally history)
3137
+ */
3138
+ preloadMemory?: (context: ExecutionContext) => Promise<AgentMemory> | AgentMemory;
3139
+ /**
3140
+ * Metrics configuration for ROI calculations
3141
+ * Optional: Only needed if tracking automation savings
3142
+ */
3143
+ metricsConfig?: ResourceMetricsConfig;
3144
+ /**
3145
+ * Execution interface configuration (optional)
3146
+ * If provided, agent appears in Execution Runner UI
3147
+ */
3148
+ interface?: ExecutionInterface;
3149
+ }
3150
+ /**
3151
+ * Agent execution context
3152
+ * Groups all state needed for agent execution phases
3153
+ */
3154
+ interface IterationContext {
3155
+ config: AgentConfig;
3156
+ contract: Contract;
3157
+ toolRegistry: Map<string, Tool>;
3158
+ memoryManager: MemoryManager;
3159
+ executionContext: ExecutionContext;
3160
+ iteration: number;
3161
+ logger: AgentScopedLogger;
3162
+ modelConfig: ModelConfig;
3163
+ adapterFactory: LLMAdapterFactory;
3164
+ /**
3165
+ * The validated input for this execution, serialized. It travels here because the model gets
3166
+ * it as its own `role:'user'` message; nothing else in this context carried it, so the input
3167
+ * had to be read back out of memory history and shipped inside the memory block.
3168
+ */
3169
+ currentInput: string;
3040
3170
  }
3041
3171
 
3042
3172
  /**
@@ -3184,6 +3314,19 @@ interface Tool {
3184
3314
  outputSchema: z.ZodSchema;
3185
3315
  execute: (options: ToolExecutionOptions) => Promise<unknown>;
3186
3316
  timeout?: number;
3317
+ /**
3318
+ * Optional per-tool output size bound, in approximate tokens. Today the ONLY size bound on tool
3319
+ * output is a 4,000-token truncation applied post-hoc at memory insert -- after the payload is
3320
+ * already fully materialized, validated against `outputSchema`, emitted as a session message, and
3321
+ * logged. This field exists so a tool can declare its own bound up front instead.
3322
+ *
3323
+ * Enforcement is NOT here. `executor.ts:executeToolCall` is the single call site that produces the
3324
+ * value handed to all four sinks (memory, the `agent:tool_result` event, the session message, and
3325
+ * the log line) -- enforcing there, before that value is emitted, is what makes the four sinks agree
3326
+ * instead of three of them seeing the untruncated payload. See that file's own comment for the
3327
+ * exact insertion point.
3328
+ */
3329
+ maxOutputTokens?: number;
3187
3330
  }
3188
3331
 
3189
3332
  /**