Bridge Kernel

The bridge kernel (src/bridge/) is the core translation engine. It sits between the responses orchestration layer and provider implementations, converting between the OpenAI Responses API protocol and Chat Completions formats in a provider-agnostic way.

Submodule Responsibilities

Submodule	Role
compatibility/	Plans supported, degraded, ignored, and rejected request features
request/	Normalizes Responses input and session history into Chat Completions messages
tools/	Plans tool declarations, tool_choice, degradation, identity mapping, and call restoration
output/	Plans structured-output contracts and validates strict downgraded JSON output
response/	Reconstructs sync ResponseObject results from provider responses
stream/	Maps provider deltas into Responses SSE events through a state machine
provider-spec/	Defines ProviderSpec, ProviderEdge, provider constants, and package shape checks
finish-reason/	Maps provider finish reasons to Responses terminal states

Request Building

The buildChatCompletionRequest() function in bridge/request/ is the main entry point. It composes four planning steps:

1. Compatibility planning — planBridgeCompatibility() checks each request parameter against the provider's capabilities and records supported, degraded, ignored, or rejected decisions.
2. Tool planning — planTools() maps tool definitions to provider tool declarations, handles tool type degradation (e.g., local_shell to function), and computes the effective tool_choice.
3. Output contract planning — planOutputContract() decides how structured output should be handled, including degrading json_schema to json_object when the provider does not support native schemas.
4. Message building — Normalizes the current input and session history into Chat Completions messages, injecting synthetic preamble instructions for downgraded output formats.

Compatibility Planning

Each parameter is checked against the provider's capability set and produces a CompatibilityDecision:

Action	Meaning
supported	Parameter is forwarded as-is
degraded	Parameter is transformed to a lower-fidelity equivalent
ignored	Parameter is consumed by GodeX and not forwarded
rejected	Parameter is not supported; the request fails with a BridgeError

Diagnostics are accumulated in ResponsesContext.diagnostics and logged after the response completes.

Provider Edge

The ProviderEdge interface is the bridge's contract with providers. Each provider implements a ProviderSpec that declares capabilities, accessors for reading responses and stream chunks, and optional hooks for request patching.

Provider-specific differences belong in each provider's spec.ts, hooks.ts, protocol types, and HTTP client. The bridge kernel itself never contains provider-specific logic.

Stream Pipeline