nestjs-agentic@1.6.0

Patch Changes

• Updated dependencies [22276c8]
• Updated dependencies [68218ea]
• Updated dependencies [f81f32e]
• Updated dependencies [cf18702]
  • @nestjs-agentic/core@1.6.0

@nestjs-agentic/rag@1.6.0

Patch Changes

• Updated dependencies [22276c8]
• Updated dependencies [68218ea]
• Updated dependencies [f81f32e]
• Updated dependencies [cf18702]
  • @nestjs-agentic/core@1.6.0
  • @nestjs-agentic/memory@1.6.0

@nestjs-agentic/orchestration@1.6.0

Patch Changes

• Updated dependencies [22276c8]
• Updated dependencies [68218ea]
• Updated dependencies [f81f32e]
• Updated dependencies [cf18702]
  • @nestjs-agentic/core@1.6.0

@nestjs-agentic/openai@1.6.0

Minor Changes

• cf18702: Add structured output: constrain an agent's final answer to a JSON Schema, with validation and bounded repair.

  • outputSchema (JSON Schema) on AgentConfig and RunInput, with optional structuredOutput options: name, description, strict (default false), maxRepairAttempts (default 2), and a pluggable validate for Ajv, zod, or any validator.
  • The built-in runtime sends the schema as the new optional ModelRequest.outputFormat. Adapters that set supportsStructuredOutput forward it natively; for others the schema is described in that request's system prompt only. Existing adapters keep working unchanged.
  • The final answer is parsed (tolerating code fences and surrounding prose), validated, and repaired by re-prompting with the specific issues. The parsed value is returned as AgentResult.structured, and runner.run<T>() types it. A turn that never conforms fails with StructuredOutputError. Repair rounds count toward usage and budgets and stay out of session history.
  • Streaming emits output_rejected before each repair round and structured on final_answer.
  • A dependency-free validateJsonSchema covering the subset providers' structured-output modes use, plus parseJsonAnswer.
  • A turn that would run through a RuntimeAdapter with an outputSchema throws StructuredOutputNotSupportedError instead of returning unvalidated text.
  • A run that sets its own outputSchema does not inherit the agent's options (its validate in particular). Schemas with an invalid pattern or an unresolvable $ref throw InvalidOutputSchemaError before any model call. A refusal (ModelResponse.refusal) fails the turn without repair. A failed turn's conversation is still saved, and an approval whose resumed answer misses its schema is recorded as settled. The schema and repair state survive approval resume and checkpoint recovery, and an agent schema that broke after a run started does not block resuming it.
  • The validator also covers patternProperties, OpenAPI nullable, draft-07 tuples, code-point string lengths, patterns that need the non-Unicode regex mode, and deep recursive $refs.
  • OpenAiModelAdapter sends outputFormat as response_format: { type: 'json_schema' } (non-object roots wrapped as { value }, with root-relative $refs kept pointing at the original root), sends schemas outside strict mode's subset non-strict instead of letting the request fail, reports refusals, and takes structuredOutput: 'prompt' for servers without json_schema support.

@nestjs-agentic/memory@1.6.0

@nestjs-agentic/memory@1.6.0

@nestjs-agentic/mcp@1.6.0

Patch Changes

• Updated dependencies [22276c8]
• Updated dependencies [68218ea]
• Updated dependencies [f81f32e]
• Updated dependencies [cf18702]
  • @nestjs-agentic/core@1.6.0

@nestjs-agentic/jev@1.6.0

Minor Changes

• 68218ea: New package: @nestjs-agentic/jev, Jev (TypeSafe System One) decisions for nestjs-agentic governance.

  • JevActionGate(options) creates an injectable tool policy that maps Jev's calibrated probability that a call is safe onto allow (at or above allowAt), require_approval (between the thresholds, with requiredApprovals as a number or a function of the probability for dual control), and deny (below denyBelow). Every decision reason, allow included, records the probability and thresholds for the audit trail. Unnamed gates get distinct names.
  • JevOutputGate(options) creates an output rail that withholds tool output, by default prompt injections, before the model sees it.
  • describe controls exactly what is sent to the TypeSafe API; an error it throws propagates instead of being treated as an outage. onError decides what happens when Jev is unreachable, slow (timeoutMs, enforced even for clients that ignore the abort signal), or answers malformed: human review by default for action gates, with onErrorRequiredApprovals approvers (required when requiredApprovals is a function), and deny for output gates. A circuit breaker, shared through JevModule, makes calls fail fast after repeated failures. A cancelled run throws ExecutionCancelledError rather than following onError.
  • JevModule.forRoot() / forRootAsync() share one client and defaults; each gate can also take its own client.
  • jevFaithfulnessJudge and jevTaskJudge plug into @nestjs-agentic/evaluation's FaithfulnessMetric and LLMAsAJudgeMetric.
  • Works with TypeSafeClient from @typesafe-ai/sdk (optional peer dependency) or any client with a compatible systemOne().

  Core: an allow policy result, before or after execution, may carry an optional reason, recorded on the audit trail (with audit.includeAllowDecisions for tool calls) and never shown to the model. When several policies require approval, a policy without its own ttlSeconds now counts with the module's approvalTtlSeconds when the shortest lifetime is chosen, so it is no longer outlived by another policy's longer explicit lifetime.

@nestjs-agentic/evaluation@1.6.0

Minor Changes

• 2482962: Add retrieval-quality metrics for RAG pipelines.

  • New RetrievalEvalMetric contract, parallel to EvalMetric, scored against retrieval output (RetrievalContext, which RAGContext satisfies structurally) rather than an AgentResult. Dataset items carry ground truth as relevantIds, with optional relevanceGrades.
  • Metrics: RecallAtKMetric, PrecisionAtKMetric, HitRateAtKMetric, ReciprocalRankMetric (MRR), NdcgAtKMetric (graded relevance), and FaithfulnessMetric with a pluggable judge that sees the passages plus any hydrated, compressed, or graph context the pipeline produced.
  • Relevance is matched per chunk (matchOn: 'id'), per document ('parentId'), or by any function of the chunk. Results are ranked by context.scores when present, or by chunks order with rankBy: 'order' for rerankers such as MMR.
  • RetrievalBenchmarkRunner runs a labeled query set through a RAGPipeline, a KnowledgeBase, or any retriever function, optionally generates answers for answer-level metrics, and reports per-metric averages for CI gates. Its defaults are Recall@k, Precision@k, MRR@k, and nDCG@k, with k separate from topK. A failing retrieval scores its item 0, a failing answer step is reported as answerError without discarding the retrieval scores, and a failing metric fails only itself.

Patch Changes

• Updated dependencies [22276c8]
• Updated dependencies [68218ea]
• Updated dependencies [f81f32e]
• Updated dependencies [cf18702]
  • @nestjs-agentic/core@1.6.0

@nestjs-agentic/core@1.6.0

Minor Changes

• 22276c8: Add dual control (N-of-M approvals) for high-risk actions.

  • A require_approval policy decision can set requiredApprovals. The withheld tool then runs only after that many distinct approvers have called ApprovalService.approve(). Earlier calls return a pending_approval result with signatures and requiredApprovals. Every signature runs through the ApprovalAuthorizer, tenant isolation, and separation-of-duties checks and must carry actor.userId. A repeated signature from the same userId is refused rather than counted twice, and a single authorized reject() vetoes.
  • New optional ApprovalStore.addSignature(id, signature) appends a signature atomically and claims the approval on the one that meets the threshold, so exactly one caller settles it. Implemented by InMemoryApprovalStore, PostgresApprovalStore, and RedisApprovalStore (when the client exposes eval; the approval and its signatures share one key). Stores without it keep working for single-approver approvals, and a multi-approver decision on such a store is denied instead of downgraded.
  • Behavior change: when a policy returns require_approval, the remaining policies on the tool still run. A later deny now wins, and the approval takes the strictest terms across policies (most approvers, shortest ttlSeconds). Previously the first approval requirement stopped evaluation, so later policies, including deny rules, never ran for that call, because an approved call skips policies.
  • Each signature records the version of the approval it was given for (approvalVersion), and completion is refused if any signature was given for a different version. Every settlement path checks the signature count before the tool runs. Cancellation is honored before store writes, and an approval consumed by a cancelled call is put back.
  • RedisApprovalStore accepts an evalFn adapter (for node-redis v4 and similar). claim() uses an atomic script when the client has eval but no getdel. The signature script reads the record without decoding JSON.
  • New approval_signed audit event per counted signature, requiredApprovals on approval_requested, and signatures on approval_settled.
  • runApprovalStoreContract gains capability-gated groups for addSignature: threshold accumulation, duplicate rejection, and atomicity under concurrent signers. Its argument round-trip check no longer depends on JSON key order, which PostgreSQL jsonb does not preserve.
  • New exports: ApprovalSignature, ApprovalSignatureResult, ApprovalSignedAuditEvent, ApprovalSignaturesUnsupportedError, requiredApprovalsOf.

• 68218ea: New package: @nestjs-agentic/jev, Jev (TypeSafe System One) decisions for nestjs-agentic governance.

  • JevActionGate(options) creates an injectable tool policy that maps Jev's calibrated probability that a call is safe onto allow (at or above allowAt), require_approval (between the thresholds, with requiredApprovals as a number or a function of the probability for dual control), and deny (below denyBelow). Every decision reason, allow included, records the probability and thresholds for the audit trail. Unnamed gates get distinct names.
  • JevOutputGate(options) creates an output rail that withholds tool output, by default prompt injections, before the model sees it.
  • describe controls exactly what is sent to the TypeSafe API; an error it throws propagates instead of being treated as an outage. onError decides what happens when Jev is unreachable, slow (timeoutMs, enforced even for clients that ignore the abort signal), or answers malformed: human review by default for action gates, with onErrorRequiredApprovals approvers (required when requiredApprovals is a function), and deny for output gates. A circuit breaker, shared through JevModule, makes calls fail fast after repeated failures. A cancelled run throws ExecutionCancelledError rather than following onError.
  • JevModule.forRoot() / forRootAsync() share one client and defaults; each gate can also take its own client.
  • jevFaithfulnessJudge and jevTaskJudge plug into @nestjs-agentic/evaluation's FaithfulnessMetric and LLMAsAJudgeMetric.
  • Works with TypeSafeClient from @typesafe-ai/sdk (optional peer dependency) or any client with a compatible systemOne().

  Core: an allow policy result, before or after execution, may carry an optional reason, recorded on the audit trail (with audit.includeAllowDecisions for tool calls) and never shown to the model. When several policies require approval, a policy without its own ttlSeconds now counts with the module's approvalTtlSeconds when the shortest lifetime is chosen, so it is no longer outlived by another policy's longer explicit lifetime.

• f81f32e: Redact errors before they reach observers.

  Observers usually forward events to external telemetry, and provider SDK errors routinely carry request headers, API keys, and prompt content. ObserverNotifier now scrubs every error-carrying field before dispatch: AgentErrorEvent.error, ModelRetryEvent.error, CircuitBreakerEvent.reason, and the error text of failed tool results in ToolResultEvent and AgentEndEvent.

  • Behavior change: by default observers now receive a RedactedError instead of the original error. It keeps name, a credential-masked message capped at 500 characters, numeric status/statusCode, a short code, stack frames, and a redacted cause chain (3 links deep). Request config, headers, response bodies, and every other property are dropped. The error thrown to the caller of run() is unchanged.
  • Configure it with observability.errorRedaction on AgenticModuleOptions: pass createErrorRedactor({ maxMessageLength, maxCauseDepth, patterns, mask }), any (error: unknown) => Error (optionally with redactText for free text), or 'none' to restore raw errors. Any other value is rejected. ObserverNotifier accepts the same setting as errorRedaction.
  • New exports: ErrorRedactor, ObservabilityOptions, ErrorRedactorOptions, RedactedError, createErrorRedactor, defaultErrorRedactor.

• cf18702: Add structured output: constrain an agent's final answer to a JSON Schema, with validation and bounded repair.

  • outputSchema (JSON Schema) on AgentConfig and RunInput, with optional structuredOutput options: name, description, strict (default false), maxRepairAttempts (default 2), and a pluggable validate for Ajv, zod, or any validator.
  • The built-in runtime sends the schema as the new optional ModelRequest.outputFormat. Adapters that set supportsStructuredOutput forward it natively; for others the schema is described in that request's system prompt only. Existing adapters keep working unchanged.
  • The final answer is parsed (tolerating code fences and surrounding prose), validated, and repaired by re-prompting with the specific issues. The parsed value is returned as AgentResult.structured, and runner.run<T>() types it. A turn that never conforms fails with StructuredOutputError. Repair rounds count toward usage and budgets and stay out of session history.
  • Streaming emits output_rejected before each repair round and structured on final_answer.
  • A dependency-free validateJsonSchema covering the subset providers' structured-output modes use, plus parseJsonAnswer.
  • A turn that would run through a RuntimeAdapter with an outputSchema throws StructuredOutputNotSupportedError instead of returning unvalidated text.
  • A run that sets its own outputSchema does not inherit the agent's options (its validate in particular). Schemas with an invalid pattern or an unresolvable $ref throw InvalidOutputSchemaError before any model call. A refusal (ModelResponse.refusal) fails the turn without repair. A failed turn's conversation is still saved, and an approval whose resumed answer misses its schema is recorded as settled. The schema and repair state survive approval resume and checkpoint recovery, and an agent schema that broke after a run started does not block resuming it.
  • The validator also covers patternProperties, OpenAPI nullable, draft-07 tuples, code-point string lengths, patterns that need the non-Unicode regex mode, and deep recursive $refs.
  • OpenAiModelAdapter sends outputFormat as response_format: { type: 'json_schema' } (non-object roots wrapped as { value }, with root-relative $refs kept pointing at the original root), sends schemas outside strict mode's subset non-strict instead of letting the request fail, reports refusals, and takes structuredOutput: 'prompt' for servers without json_schema support.

nestjs-agentic@1.5.0

Patch Changes

• c679715: Support NestJS 12 and tighten what gets published.

  • Peer dependencies on @nestjs/common and @nestjs/core now accept ^12.0.0, so installing into a NestJS 12 app no longer fails with ERESOLVE. Every package's test suite passes against NestJS 11.2 and 12.1, and CI now checks both. NestJS 12 is ESM only; with these CommonJS packages it needs Node 20.19+ or 22.12+.
  • Internal * ranges (@nestjs-agentic/memory in rag's dependencies, and the @nestjs-agentic/core / nestjs-agentic peers of memory, rag and openai) are now ^1.4.0, so a future major of a sibling package is never pulled in silently.
  • Compiled tests are written to test-dist/ instead of dist/test/, so they are no longer published. evaluation, mcp and orchestration now declare files like the other packages, so they stop shipping src/, test/ and tsconfig.json and exclude test-dist/ even when packed outside the repo. @nestjs-agentic/core drops from 712 files (3.4 MB) to about 315 files (1.4 MB) unpacked.

• Updated dependencies [9ee3e4a]

• Updated dependencies [c679715]

  • @nestjs-agentic/core@1.5.0