Add TEE attestation and data settlement metadata to LLM responses by adambalogh · Pull Request #7 · OpenGradient/typescript-sdk

adambalogh · 2026-05-05T15:31:00Z

Summary

This PR enhances the LLM client to include TEE (Trusted Execution Environment) attestation fields and data settlement metadata in all responses. It also updates the settlement mode enum to reflect the current protocol, adds support for response format control, and improves streaming reliability for tool-call requests.

Key Changes

TEE Attestation Fields: All responses now include teeSignature, teeTimestamp, teeId, teeEndpoint, and teePaymentAddress to cryptographically verify inference was performed inside a TEE enclave.
Data Settlement Metadata: Responses now carry dataSettlementTransactionHash and dataSettlementBlobId headers extracted from the server response, enabling auditability of inference data settlement.
Settlement Mode Enum Update: Renamed settlement modes to match current protocol:
- SETTLE → PRIVATE (payment-only, no on-chain data)
- SETTLE_BATCH → BATCH_HASHED (default, aggregated batch settlement)
- SETTLE_METADATA → INDIVIDUAL_FULL (full on-chain metadata)
Response Format Support: Added ResponseFormat interface and validation to support json_object and json_schema output formats, with provider-specific validation (e.g., rejecting json_object for Anthropic models).
Tool-Call Streaming Fallback: Implemented chatToolsAsStream() to handle tool-call requests by falling back to non-streaming mode and emitting a single final chunk, matching Python SDK behavior.
Streaming Improvements:
- Added retry logic with reconnection for stream setup failures
- Fixed handling of SSE events that use message key instead of delta
- Deferred final chunk emission until [DONE] marker to ensure all metadata is available
- Added TokenUsage type for consistent token reporting across streaming and non-streaming responses
Model Catalog Updates: Updated TEE_LLM enum with latest model versions across OpenAI, Anthropic, Google, xAI, and added ByteDance Seed models.
Connection Management: Enhanced RegistryTEEConnection with deduplication of concurrent reconnect attempts and proper cleanup on close.

Notable Implementation Details

Response format serialization validates that jsonSchema is provided when type: "json_schema" is specified
Stream chunks now properly accumulate TEE metadata on the final chunk only
Malformed SSE chunks are skipped gracefully rather than breaking the stream
Settlement mode defaults to BATCH_HASHED for cost efficiency
Content flattening for Anthropic responses that return content as typed blocks

https://claude.ai/code/session_01BE3dkcPw4jbUrrZN3cizCh

Brings the TS SDK's LLM response surface in line with the Python SDK. - TextGenerationOutput: replace `transactionHash` with the Python field set — `dataSettlementTransactionHash`, `dataSettlementBlobId`, `chatOutput`, `completionOutput`, `usage`, `paymentHash`, `teeSignature`, `teeTimestamp`, `teeId`, `teeEndpoint`, `teePaymentAddress`, `finishReason`. - StreamChunk: add TEE attestation and data settlement fields populated on the final chunk; rename `is_final` to `isFinal`. SSE parser now handles servers that send the non-streaming `message` key in a delta event (mirrors Python's `from_sse_data`). - X402SettlementMode: align enum values with the wire protocol used by the Python SDK (`private` / `batch` / `individual`). - TEE_LLM: refresh the model list (GPT-5 family, Claude Sonnet/Haiku/ Opus 4.5+, Gemini 3, Grok 4 family, ByteDance Seed). - Add `ResponseFormat` type with json_object/json_schema enforcement (and the Anthropic + json_object guard from the Python SDK). - llm.ts: read x-settlement-tx-hash and x-settlement-walrus-blob-id response headers; flatten Anthropic content-block arrays; fall back to non-streaming for tool-call requests and emit a single final StreamChunk; only retry on connection-level failures (preserve the HTTP-status no-retry rule). - teeConnection.ts: dedupe concurrent reconnect() calls and guard the refresh loop after close.

adambalogh · 2026-05-05T15:31:00Z

Summary

This PR enhances the LLM client to include TEE (Trusted Execution Environment) attestation fields and data settlement metadata in all responses. It also updates the settlement mode enum to reflect the current protocol, adds support for response format control, and improves streaming reliability for tool-call requests.

Key Changes

TEE Attestation Fields: All responses now include teeSignature, teeTimestamp, teeId, teeEndpoint, and teePaymentAddress to cryptographically verify inference was performed inside a TEE enclave.
Data Settlement Metadata: Responses now carry dataSettlementTransactionHash and dataSettlementBlobId headers extracted from the server response, enabling auditability of inference data settlement.
Settlement Mode Enum Update: Renamed settlement modes to match current protocol:
- SETTLE → PRIVATE (payment-only, no on-chain data)
- SETTLE_BATCH → BATCH_HASHED (default, aggregated batch settlement)
- SETTLE_METADATA → INDIVIDUAL_FULL (full on-chain metadata)
Response Format Support: Added ResponseFormat interface and validation to support json_object and json_schema output formats, with provider-specific validation (e.g., rejecting json_object for Anthropic models).
Tool-Call Streaming Fallback: Implemented chatToolsAsStream() to handle tool-call requests by falling back to non-streaming mode and emitting a single final chunk, matching Python SDK behavior.
Streaming Improvements:
- Added retry logic with reconnection for stream setup failures
- Fixed handling of SSE events that use message key instead of delta
- Deferred final chunk emission until [DONE] marker to ensure all metadata is available
- Added TokenUsage type for consistent token reporting across streaming and non-streaming responses
Model Catalog Updates: Updated TEE_LLM enum with latest model versions across OpenAI, Anthropic, Google, xAI, and added ByteDance Seed models.
Connection Management: Enhanced RegistryTEEConnection with deduplication of concurrent reconnect attempts and proper cleanup on close.

Notable Implementation Details

Response format serialization validates that jsonSchema is provided when type: "json_schema" is specified
Stream chunks now properly accumulate TEE metadata on the final chunk only
Malformed SSE chunks are skipped gracefully rather than breaking the stream
Settlement mode defaults to BATCH_HASHED for cost efficiency
Content flattening for Anthropic responses that return content as typed blocks

https://claude.ai/code/session_01BE3dkcPw4jbUrrZN3cizCh

Brings the TS SDK's LLM response surface in line with the Python SDK. - TextGenerationOutput: replace `transactionHash` with the Python field set — `dataSettlementTransactionHash`, `dataSettlementBlobId`, `chatOutput`, `completionOutput`, `usage`, `paymentHash`, `teeSignature`, `teeTimestamp`, `teeId`, `teeEndpoint`, `teePaymentAddress`, `finishReason`. - StreamChunk: add TEE attestation and data settlement fields populated on the final chunk; rename `is_final` to `isFinal`. SSE parser now handles servers that send the non-streaming `message` key in a delta event (mirrors Python's `from_sse_data`). - X402SettlementMode: align enum values with the wire protocol used by the Python SDK (`private` / `batch` / `individual`). - TEE_LLM: refresh the model list (GPT-5 family, Claude Sonnet/Haiku/ Opus 4.5+, Gemini 3, Grok 4 family, ByteDance Seed). - Add `ResponseFormat` type with json_object/json_schema enforcement (and the Anthropic + json_object guard from the Python SDK). - llm.ts: read x-settlement-tx-hash and x-settlement-walrus-blob-id response headers; flatten Anthropic content-block arrays; fall back to non-streaming for tool-call requests and emit a single final StreamChunk; only retry on connection-level failures (preserve the HTTP-status no-retry rule). - teeConnection.ts: dedupe concurrent reconnect() calls and guard the refresh loop after close.

adambalogh · 2026-05-05T15:31:00Z

Summary

This PR enhances the LLM client to include TEE (Trusted Execution Environment) attestation fields and data settlement metadata in all responses. It also updates the settlement mode enum to reflect the current protocol, adds support for response format control, and improves streaming reliability for tool-call requests.

Key Changes

TEE Attestation Fields: All responses now include teeSignature, teeTimestamp, teeId, teeEndpoint, and teePaymentAddress to cryptographically verify inference was performed inside a TEE enclave.
Data Settlement Metadata: Responses now carry dataSettlementTransactionHash and dataSettlementBlobId headers extracted from the server response, enabling auditability of inference data settlement.
Settlement Mode Enum Update: Renamed settlement modes to match current protocol:
- SETTLE → PRIVATE (payment-only, no on-chain data)
- SETTLE_BATCH → BATCH_HASHED (default, aggregated batch settlement)
- SETTLE_METADATA → INDIVIDUAL_FULL (full on-chain metadata)
Response Format Support: Added ResponseFormat interface and validation to support json_object and json_schema output formats, with provider-specific validation (e.g., rejecting json_object for Anthropic models).
Tool-Call Streaming Fallback: Implemented chatToolsAsStream() to handle tool-call requests by falling back to non-streaming mode and emitting a single final chunk, matching Python SDK behavior.
Streaming Improvements:
- Added retry logic with reconnection for stream setup failures
- Fixed handling of SSE events that use message key instead of delta
- Deferred final chunk emission until [DONE] marker to ensure all metadata is available
- Added TokenUsage type for consistent token reporting across streaming and non-streaming responses
Model Catalog Updates: Updated TEE_LLM enum with latest model versions across OpenAI, Anthropic, Google, xAI, and added ByteDance Seed models.
Connection Management: Enhanced RegistryTEEConnection with deduplication of concurrent reconnect attempts and proper cleanup on close.

Notable Implementation Details

Response format serialization validates that jsonSchema is provided when type: "json_schema" is specified
Stream chunks now properly accumulate TEE metadata on the final chunk only
Malformed SSE chunks are skipped gracefully rather than breaking the stream
Settlement mode defaults to BATCH_HASHED for cost efficiency
Content flattening for Anthropic responses that return content as typed blocks

https://claude.ai/code/session_01BE3dkcPw4jbUrrZN3cizCh

Brings the TS SDK's LLM response surface in line with the Python SDK. - TextGenerationOutput: replace `transactionHash` with the Python field set — `dataSettlementTransactionHash`, `dataSettlementBlobId`, `chatOutput`, `completionOutput`, `usage`, `paymentHash`, `teeSignature`, `teeTimestamp`, `teeId`, `teeEndpoint`, `teePaymentAddress`, `finishReason`. - StreamChunk: add TEE attestation and data settlement fields populated on the final chunk; rename `is_final` to `isFinal`. SSE parser now handles servers that send the non-streaming `message` key in a delta event (mirrors Python's `from_sse_data`). - X402SettlementMode: align enum values with the wire protocol used by the Python SDK (`private` / `batch` / `individual`). - TEE_LLM: refresh the model list (GPT-5 family, Claude Sonnet/Haiku/ Opus 4.5+, Gemini 3, Grok 4 family, ByteDance Seed). - Add `ResponseFormat` type with json_object/json_schema enforcement (and the Anthropic + json_object guard from the Python SDK). - llm.ts: read x-settlement-tx-hash and x-settlement-walrus-blob-id response headers; flatten Anthropic content-block arrays; fall back to non-streaming for tool-call requests and emit a single final StreamChunk; only retry on connection-level failures (preserve the HTTP-status no-retry rule). - teeConnection.ts: dedupe concurrent reconnect() calls and guard the refresh loop after close.

adambalogh · 2026-05-05T15:31:00Z

Summary

This PR enhances the LLM client to include TEE (Trusted Execution Environment) attestation fields and data settlement metadata in all responses. It also updates the settlement mode enum to reflect the current protocol, adds support for response format control, and improves streaming reliability for tool-call requests.

Key Changes

TEE Attestation Fields: All responses now include teeSignature, teeTimestamp, teeId, teeEndpoint, and teePaymentAddress to cryptographically verify inference was performed inside a TEE enclave.
Data Settlement Metadata: Responses now carry dataSettlementTransactionHash and dataSettlementBlobId headers extracted from the server response, enabling auditability of inference data settlement.
Settlement Mode Enum Update: Renamed settlement modes to match current protocol:
- SETTLE → PRIVATE (payment-only, no on-chain data)
- SETTLE_BATCH → BATCH_HASHED (default, aggregated batch settlement)
- SETTLE_METADATA → INDIVIDUAL_FULL (full on-chain metadata)
Response Format Support: Added ResponseFormat interface and validation to support json_object and json_schema output formats, with provider-specific validation (e.g., rejecting json_object for Anthropic models).
Tool-Call Streaming Fallback: Implemented chatToolsAsStream() to handle tool-call requests by falling back to non-streaming mode and emitting a single final chunk, matching Python SDK behavior.
Streaming Improvements:
- Added retry logic with reconnection for stream setup failures
- Fixed handling of SSE events that use message key instead of delta
- Deferred final chunk emission until [DONE] marker to ensure all metadata is available
- Added TokenUsage type for consistent token reporting across streaming and non-streaming responses
Model Catalog Updates: Updated TEE_LLM enum with latest model versions across OpenAI, Anthropic, Google, xAI, and added ByteDance Seed models.
Connection Management: Enhanced RegistryTEEConnection with deduplication of concurrent reconnect attempts and proper cleanup on close.

Notable Implementation Details

Response format serialization validates that jsonSchema is provided when type: "json_schema" is specified
Stream chunks now properly accumulate TEE metadata on the final chunk only
Malformed SSE chunks are skipped gracefully rather than breaking the stream
Settlement mode defaults to BATCH_HASHED for cost efficiency
Content flattening for Anthropic responses that return content as typed blocks

https://claude.ai/code/session_01BE3dkcPw4jbUrrZN3cizCh

Brings the TS SDK's LLM response surface in line with the Python SDK. - TextGenerationOutput: replace `transactionHash` with the Python field set — `dataSettlementTransactionHash`, `dataSettlementBlobId`, `chatOutput`, `completionOutput`, `usage`, `paymentHash`, `teeSignature`, `teeTimestamp`, `teeId`, `teeEndpoint`, `teePaymentAddress`, `finishReason`. - StreamChunk: add TEE attestation and data settlement fields populated on the final chunk; rename `is_final` to `isFinal`. SSE parser now handles servers that send the non-streaming `message` key in a delta event (mirrors Python's `from_sse_data`). - X402SettlementMode: align enum values with the wire protocol used by the Python SDK (`private` / `batch` / `individual`). - TEE_LLM: refresh the model list (GPT-5 family, Claude Sonnet/Haiku/ Opus 4.5+, Gemini 3, Grok 4 family, ByteDance Seed). - Add `ResponseFormat` type with json_object/json_schema enforcement (and the Anthropic + json_object guard from the Python SDK). - llm.ts: read x-settlement-tx-hash and x-settlement-walrus-blob-id response headers; flatten Anthropic content-block arrays; fall back to non-streaming for tool-call requests and emit a single final StreamChunk; only retry on connection-level failures (preserve the HTTP-status no-retry rule). - teeConnection.ts: dedupe concurrent reconnect() calls and guard the refresh loop after close.

adambalogh · 2026-05-05T15:31:00Z

Summary

This PR enhances the LLM client to include TEE (Trusted Execution Environment) attestation fields and data settlement metadata in all responses. It also updates the settlement mode enum to reflect the current protocol, adds support for response format control, and improves streaming reliability for tool-call requests.

Key Changes

TEE Attestation Fields: All responses now include teeSignature, teeTimestamp, teeId, teeEndpoint, and teePaymentAddress to cryptographically verify inference was performed inside a TEE enclave.
Data Settlement Metadata: Responses now carry dataSettlementTransactionHash and dataSettlementBlobId headers extracted from the server response, enabling auditability of inference data settlement.
Settlement Mode Enum Update: Renamed settlement modes to match current protocol:
- SETTLE → PRIVATE (payment-only, no on-chain data)
- SETTLE_BATCH → BATCH_HASHED (default, aggregated batch settlement)
- SETTLE_METADATA → INDIVIDUAL_FULL (full on-chain metadata)
Response Format Support: Added ResponseFormat interface and validation to support json_object and json_schema output formats, with provider-specific validation (e.g., rejecting json_object for Anthropic models).
Tool-Call Streaming Fallback: Implemented chatToolsAsStream() to handle tool-call requests by falling back to non-streaming mode and emitting a single final chunk, matching Python SDK behavior.
Streaming Improvements:
- Added retry logic with reconnection for stream setup failures
- Fixed handling of SSE events that use message key instead of delta
- Deferred final chunk emission until [DONE] marker to ensure all metadata is available
- Added TokenUsage type for consistent token reporting across streaming and non-streaming responses
Model Catalog Updates: Updated TEE_LLM enum with latest model versions across OpenAI, Anthropic, Google, xAI, and added ByteDance Seed models.
Connection Management: Enhanced RegistryTEEConnection with deduplication of concurrent reconnect attempts and proper cleanup on close.

Notable Implementation Details

Response format serialization validates that jsonSchema is provided when type: "json_schema" is specified
Stream chunks now properly accumulate TEE metadata on the final chunk only
Malformed SSE chunks are skipped gracefully rather than breaking the stream
Settlement mode defaults to BATCH_HASHED for cost efficiency
Content flattening for Anthropic responses that return content as typed blocks

https://claude.ai/code/session_01BE3dkcPw4jbUrrZN3cizCh

Brings the TS SDK's LLM response surface in line with the Python SDK. - TextGenerationOutput: replace `transactionHash` with the Python field set — `dataSettlementTransactionHash`, `dataSettlementBlobId`, `chatOutput`, `completionOutput`, `usage`, `paymentHash`, `teeSignature`, `teeTimestamp`, `teeId`, `teeEndpoint`, `teePaymentAddress`, `finishReason`. - StreamChunk: add TEE attestation and data settlement fields populated on the final chunk; rename `is_final` to `isFinal`. SSE parser now handles servers that send the non-streaming `message` key in a delta event (mirrors Python's `from_sse_data`). - X402SettlementMode: align enum values with the wire protocol used by the Python SDK (`private` / `batch` / `individual`). - TEE_LLM: refresh the model list (GPT-5 family, Claude Sonnet/Haiku/ Opus 4.5+, Gemini 3, Grok 4 family, ByteDance Seed). - Add `ResponseFormat` type with json_object/json_schema enforcement (and the Anthropic + json_object guard from the Python SDK). - llm.ts: read x-settlement-tx-hash and x-settlement-walrus-blob-id response headers; flatten Anthropic content-block arrays; fall back to non-streaming for tool-call requests and emit a single final StreamChunk; only retry on connection-level failures (preserve the HTTP-status no-retry rule). - teeConnection.ts: dedupe concurrent reconnect() calls and guard the refresh loop after close.

adambalogh · 2026-05-05T15:31:00Z

Summary

This PR enhances the LLM client to include TEE (Trusted Execution Environment) attestation fields and data settlement metadata in all responses. It also updates the settlement mode enum to reflect the current protocol, adds support for response format control, and improves streaming reliability for tool-call requests.

Key Changes

TEE Attestation Fields: All responses now include teeSignature, teeTimestamp, teeId, teeEndpoint, and teePaymentAddress to cryptographically verify inference was performed inside a TEE enclave.
Data Settlement Metadata: Responses now carry dataSettlementTransactionHash and dataSettlementBlobId headers extracted from the server response, enabling auditability of inference data settlement.
Settlement Mode Enum Update: Renamed settlement modes to match current protocol:
- SETTLE → PRIVATE (payment-only, no on-chain data)
- SETTLE_BATCH → BATCH_HASHED (default, aggregated batch settlement)
- SETTLE_METADATA → INDIVIDUAL_FULL (full on-chain metadata)
Response Format Support: Added ResponseFormat interface and validation to support json_object and json_schema output formats, with provider-specific validation (e.g., rejecting json_object for Anthropic models).
Tool-Call Streaming Fallback: Implemented chatToolsAsStream() to handle tool-call requests by falling back to non-streaming mode and emitting a single final chunk, matching Python SDK behavior.
Streaming Improvements:
- Added retry logic with reconnection for stream setup failures
- Fixed handling of SSE events that use message key instead of delta
- Deferred final chunk emission until [DONE] marker to ensure all metadata is available
- Added TokenUsage type for consistent token reporting across streaming and non-streaming responses
Model Catalog Updates: Updated TEE_LLM enum with latest model versions across OpenAI, Anthropic, Google, xAI, and added ByteDance Seed models.
Connection Management: Enhanced RegistryTEEConnection with deduplication of concurrent reconnect attempts and proper cleanup on close.

Notable Implementation Details

Response format serialization validates that jsonSchema is provided when type: "json_schema" is specified
Stream chunks now properly accumulate TEE metadata on the final chunk only
Malformed SSE chunks are skipped gracefully rather than breaking the stream
Settlement mode defaults to BATCH_HASHED for cost efficiency
Content flattening for Anthropic responses that return content as typed blocks

https://claude.ai/code/session_01BE3dkcPw4jbUrrZN3cizCh

Brings the TS SDK's LLM response surface in line with the Python SDK. - TextGenerationOutput: replace `transactionHash` with the Python field set — `dataSettlementTransactionHash`, `dataSettlementBlobId`, `chatOutput`, `completionOutput`, `usage`, `paymentHash`, `teeSignature`, `teeTimestamp`, `teeId`, `teeEndpoint`, `teePaymentAddress`, `finishReason`. - StreamChunk: add TEE attestation and data settlement fields populated on the final chunk; rename `is_final` to `isFinal`. SSE parser now handles servers that send the non-streaming `message` key in a delta event (mirrors Python's `from_sse_data`). - X402SettlementMode: align enum values with the wire protocol used by the Python SDK (`private` / `batch` / `individual`). - TEE_LLM: refresh the model list (GPT-5 family, Claude Sonnet/Haiku/ Opus 4.5+, Gemini 3, Grok 4 family, ByteDance Seed). - Add `ResponseFormat` type with json_object/json_schema enforcement (and the Anthropic + json_object guard from the Python SDK). - llm.ts: read x-settlement-tx-hash and x-settlement-walrus-blob-id response headers; flatten Anthropic content-block arrays; fall back to non-streaming for tool-call requests and emit a single final StreamChunk; only retry on connection-level failures (preserve the HTTP-status no-retry rule). - teeConnection.ts: dedupe concurrent reconnect() calls and guard the refresh loop after close.

adambalogh · 2026-05-05T15:31:00Z

Summary

This PR enhances the LLM client to include TEE (Trusted Execution Environment) attestation fields and data settlement metadata in all responses. It also updates the settlement mode enum to reflect the current protocol, adds support for response format control, and improves streaming reliability for tool-call requests.

Key Changes

TEE Attestation Fields: All responses now include teeSignature, teeTimestamp, teeId, teeEndpoint, and teePaymentAddress to cryptographically verify inference was performed inside a TEE enclave.
Data Settlement Metadata: Responses now carry dataSettlementTransactionHash and dataSettlementBlobId headers extracted from the server response, enabling auditability of inference data settlement.
Settlement Mode Enum Update: Renamed settlement modes to match current protocol:
- SETTLE → PRIVATE (payment-only, no on-chain data)
- SETTLE_BATCH → BATCH_HASHED (default, aggregated batch settlement)
- SETTLE_METADATA → INDIVIDUAL_FULL (full on-chain metadata)
Response Format Support: Added ResponseFormat interface and validation to support json_object and json_schema output formats, with provider-specific validation (e.g., rejecting json_object for Anthropic models).
Tool-Call Streaming Fallback: Implemented chatToolsAsStream() to handle tool-call requests by falling back to non-streaming mode and emitting a single final chunk, matching Python SDK behavior.
Streaming Improvements:
- Added retry logic with reconnection for stream setup failures
- Fixed handling of SSE events that use message key instead of delta
- Deferred final chunk emission until [DONE] marker to ensure all metadata is available
- Added TokenUsage type for consistent token reporting across streaming and non-streaming responses
Model Catalog Updates: Updated TEE_LLM enum with latest model versions across OpenAI, Anthropic, Google, xAI, and added ByteDance Seed models.
Connection Management: Enhanced RegistryTEEConnection with deduplication of concurrent reconnect attempts and proper cleanup on close.

Notable Implementation Details

Response format serialization validates that jsonSchema is provided when type: "json_schema" is specified
Stream chunks now properly accumulate TEE metadata on the final chunk only
Malformed SSE chunks are skipped gracefully rather than breaking the stream
Settlement mode defaults to BATCH_HASHED for cost efficiency
Content flattening for Anthropic responses that return content as typed blocks

https://claude.ai/code/session_01BE3dkcPw4jbUrrZN3cizCh

Brings the TS SDK's LLM response surface in line with the Python SDK. - TextGenerationOutput: replace `transactionHash` with the Python field set — `dataSettlementTransactionHash`, `dataSettlementBlobId`, `chatOutput`, `completionOutput`, `usage`, `paymentHash`, `teeSignature`, `teeTimestamp`, `teeId`, `teeEndpoint`, `teePaymentAddress`, `finishReason`. - StreamChunk: add TEE attestation and data settlement fields populated on the final chunk; rename `is_final` to `isFinal`. SSE parser now handles servers that send the non-streaming `message` key in a delta event (mirrors Python's `from_sse_data`). - X402SettlementMode: align enum values with the wire protocol used by the Python SDK (`private` / `batch` / `individual`). - TEE_LLM: refresh the model list (GPT-5 family, Claude Sonnet/Haiku/ Opus 4.5+, Gemini 3, Grok 4 family, ByteDance Seed). - Add `ResponseFormat` type with json_object/json_schema enforcement (and the Anthropic + json_object guard from the Python SDK). - llm.ts: read x-settlement-tx-hash and x-settlement-walrus-blob-id response headers; flatten Anthropic content-block arrays; fall back to non-streaming for tool-call requests and emit a single final StreamChunk; only retry on connection-level failures (preserve the HTTP-status no-retry rule). - teeConnection.ts: dedupe concurrent reconnect() calls and guard the refresh loop after close.

adambalogh · 2026-05-05T15:31:00Z

Summary

This PR enhances the LLM client to include TEE (Trusted Execution Environment) attestation fields and data settlement metadata in all responses. It also updates the settlement mode enum to reflect the current protocol, adds support for response format control, and improves streaming reliability for tool-call requests.

Key Changes

TEE Attestation Fields: All responses now include teeSignature, teeTimestamp, teeId, teeEndpoint, and teePaymentAddress to cryptographically verify inference was performed inside a TEE enclave.
Data Settlement Metadata: Responses now carry dataSettlementTransactionHash and dataSettlementBlobId headers extracted from the server response, enabling auditability of inference data settlement.
Settlement Mode Enum Update: Renamed settlement modes to match current protocol:
- SETTLE → PRIVATE (payment-only, no on-chain data)
- SETTLE_BATCH → BATCH_HASHED (default, aggregated batch settlement)
- SETTLE_METADATA → INDIVIDUAL_FULL (full on-chain metadata)
Response Format Support: Added ResponseFormat interface and validation to support json_object and json_schema output formats, with provider-specific validation (e.g., rejecting json_object for Anthropic models).
Tool-Call Streaming Fallback: Implemented chatToolsAsStream() to handle tool-call requests by falling back to non-streaming mode and emitting a single final chunk, matching Python SDK behavior.
Streaming Improvements:
- Added retry logic with reconnection for stream setup failures
- Fixed handling of SSE events that use message key instead of delta
- Deferred final chunk emission until [DONE] marker to ensure all metadata is available
- Added TokenUsage type for consistent token reporting across streaming and non-streaming responses
Model Catalog Updates: Updated TEE_LLM enum with latest model versions across OpenAI, Anthropic, Google, xAI, and added ByteDance Seed models.
Connection Management: Enhanced RegistryTEEConnection with deduplication of concurrent reconnect attempts and proper cleanup on close.

Notable Implementation Details

Response format serialization validates that jsonSchema is provided when type: "json_schema" is specified
Stream chunks now properly accumulate TEE metadata on the final chunk only
Malformed SSE chunks are skipped gracefully rather than breaking the stream
Settlement mode defaults to BATCH_HASHED for cost efficiency
Content flattening for Anthropic responses that return content as typed blocks

https://claude.ai/code/session_01BE3dkcPw4jbUrrZN3cizCh

Brings the TS SDK's LLM response surface in line with the Python SDK. - TextGenerationOutput: replace `transactionHash` with the Python field set — `dataSettlementTransactionHash`, `dataSettlementBlobId`, `chatOutput`, `completionOutput`, `usage`, `paymentHash`, `teeSignature`, `teeTimestamp`, `teeId`, `teeEndpoint`, `teePaymentAddress`, `finishReason`. - StreamChunk: add TEE attestation and data settlement fields populated on the final chunk; rename `is_final` to `isFinal`. SSE parser now handles servers that send the non-streaming `message` key in a delta event (mirrors Python's `from_sse_data`). - X402SettlementMode: align enum values with the wire protocol used by the Python SDK (`private` / `batch` / `individual`). - TEE_LLM: refresh the model list (GPT-5 family, Claude Sonnet/Haiku/ Opus 4.5+, Gemini 3, Grok 4 family, ByteDance Seed). - Add `ResponseFormat` type with json_object/json_schema enforcement (and the Anthropic + json_object guard from the Python SDK). - llm.ts: read x-settlement-tx-hash and x-settlement-walrus-blob-id response headers; flatten Anthropic content-block arrays; fall back to non-streaming for tool-call requests and emit a single final StreamChunk; only retry on connection-level failures (preserve the HTTP-status no-retry rule). - teeConnection.ts: dedupe concurrent reconnect() calls and guard the refresh loop after close.

adambalogh merged commit 949532a into mainMay 5, 2026
4 of 5 checks passed

Uh oh!

Conversation

adambalogh commented May 5, 2026

Summary

Key Changes

Notable Implementation Details

Uh oh!

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

2 participants

Uh oh!

Conversation

adambalogh commented May 5, 2026

Summary

Key Changes

Notable Implementation Details

Uh oh!

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

2 participants

Uh oh!

Conversation

adambalogh commented May 5, 2026

Summary

Key Changes

Notable Implementation Details

Uh oh!

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

2 participants

Uh oh!

Conversation

adambalogh commented May 5, 2026

Summary

Key Changes

Notable Implementation Details

Uh oh!

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

2 participants

Uh oh!

Conversation

adambalogh commented May 5, 2026

Summary

Key Changes

Notable Implementation Details

Uh oh!

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

2 participants