LLM Package Guide

Effect

Prefer HttpClient.HttpClient / HttpClientResponse.HttpClientResponse over web fetch / Response at package boundaries.
Use Stream.Stream for streaming data flow. Avoid ad hoc async generators or manual web reader loops unless an Effect Stream API cannot model the behavior.
Use Effect Schema codecs for JSON encode/decode (Schema.fromJsonString(...)) instead of direct JSON.parse / JSON.stringify in implementation code.
In Effect.gen, yield yieldable errors directly (return yield* new MyError(...)) instead of Effect.fail(new MyError(...)).
Use Effect.void instead of Effect.succeed(undefined) when the successful value is intentionally void.

Conventions

Per-type constructors live on the type, not as top-level re-exports. Use Message.system(...), Message.user(...), Message.assistant(...), Message.tool(...), Model.make(...), ToolDefinition.make(...), ToolCallPart.make(...), ToolResultPart.make(...), ToolChoice.make(...), ToolChoice.named(...), SystemPart.make(...), and GenerationOptions.make(...) directly. The top-level LLM namespace is reserved for request-shaped call APIs: LLM.request, LLM.generate, LLM.stream, LLM.updateRequest, and LLM.generateObject. Two ways to construct the same thing is one too many.

Tests

Use testEffect(...) from test/lib/effect.ts for tests requiring Effect layers.
Keep provider tests fixture-first. Live provider calls must stay behind RECORD=true and required API-key checks.

Architecture

This package is an Effect Schema-first LLM core. The Schema classes in src/schema/ are the canonical runtime data model. Convenience functions in src/llm.ts are thin constructors that return those same Schema class instances; they should improve callsites without creating a second model.

Primary in-repo integration point:

packages/opencode/src/session/llm.ts is the session-owned orchestration layer that decides whether a request uses AI SDK or this package's native route runtime.
packages/opencode/src/session/llm/native-request.ts is the lowering adapter from opencode's session/AI SDK-shaped data into this package's LLMRequest model.
packages/opencode/src/session/llm/native-runtime.ts is the execution adapter that calls raw LLMClient.stream(request) and bridges one provider turn of opencode tool calls through this package's typed dispatcher.
packages/opencode/src/session/llm/ai-sdk.ts keeps the default AI SDK path compatible by converting AI SDK stream parts into this package's shared LLMEvents.

Keep this package independent of session concerns. Session auth, permissions, plugins, telemetry headers, and runtime selection belong in packages/opencode/src/session/llm.ts and its local adapters.

Request Flow

The intended callsite is:

const request = LLM.request({
  model: OpenAI.configure({ apiKey }).responses("gpt-4o-mini"),
  system: "You are concise.",
  prompt: "Say hello.",
})

const response = yield * LLMClient.generate(request)

LLM.request(...) builds an LLMRequest. LLMClient.generate(...) reads the executable route carried by request.model.route, builds the provider-native body, asks the route's transport for a real HttpClientRequest.HttpClientRequest, sends it through RequestExecutor.Service, parses the provider stream into common LLMEvents, and finally returns an LLMResponse.

Use LLMClient.stream(request) when callers want incremental LLMEvents. Use LLMClient.generate(request) when callers want those same events collected into an LLMResponse. Use LLMClient.prepare<Body>(request) to compile a request through the route pipeline without sending it — the optional Body type argument narrows .body to the route's native shape (e.g. prepare<OpenAIChatBody>(...) returns a PreparedRequestOf<OpenAIChatBody>). The runtime body is identical; the generic is a type-level assertion.

Filter or narrow LLMEvent streams with LLMEvent.is.* (camelCase guards, e.g. events.filter(LLMEvent.is.toolCall)). The kebab-case LLMEvent.guards["tool-call"] form also works but prefer is.* in new code.

Routes

A route is the registered, runnable composition of four orthogonal pieces:

Protocol (src/route/protocol.ts) — semantic API contract. Owns request body construction (body.from), the body schema (body.schema), the streaming-event schema (stream.event), and the event-to-LLMEvent state machine (stream.step). Route.make(...) validates and JSON-encodes the body from body.schema and decodes frames with stream.event. Examples: OpenAIChat.protocol, OpenAIResponses.protocol, AnthropicMessages.protocol, Gemini.protocol, BedrockConverse.protocol.
Endpoint (src/route/endpoint.ts) — URL construction. The host, path, and route query live on the endpoint. Endpoint.path("/chat/completions", { baseURL }) is the common case; pass a function for paths that embed the model id or a body field (e.g. Endpoint.path(({ body }) =>/model/${body.modelId}/converse-stream)).
Auth (src/route/auth.ts) — per-request transport authentication. Provider facades configure credentials onto the route before model selection, usually via Auth.bearer(apiKey) or Auth.header(name, apiKey). Routes that need per-request signing (Bedrock SigV4, future Vertex IAM, Azure AAD) implement Auth as a function that signs the body and merges signed headers into the result.
Framing (src/route/framing.ts) — bytes → frames. SSE (Framing.sse) is shared; Bedrock keeps its AWS event-stream framing as a typed Framing<object> value alongside its protocol.

Compose them via Route.make(...):

export const route = Route.make({
  id: "openai-chat",
  provider: "openai",
  protocol: OpenAIChat.protocol,
  endpoint: Endpoint.path("/chat/completions", {
    baseURL: "https://api.openai.com/v1",
  }),
  auth: Auth.bearer(),
  framing: Framing.sse,
})

Route defaults are request-shaping defaults such as headers, limits, generation, providerOptions, and http. Endpoint host/query belongs on the route endpoint. Selected Model values carry only model id, provider id, and the configured route value. Model capability/catalog metadata lives outside this package; protocol support is enforced by request lowering and typed LLMErrors.

The four-axis decomposition is the reason DeepSeek, TogetherAI, Cerebras, Baseten, Fireworks, and DeepInfra all reuse OpenAIChat.protocol verbatim — each provider deployment is a 5-15 line Route.make(...) call instead of a 300-400 line route clone. Bug fixes in one protocol propagate to every consumer of that protocol in a single commit.

When a provider ships a non-HTTP transport (OpenAI's WebSocket Responses backend, hypothetical bidirectional streaming APIs), the seam is Transport — WebSocketTransport.jsonTransport.with(...) constructs an IO template whose prepare receives the route endpoint/auth at compile time, builds a WebSocket URL and message, and whose frames yields decoded text from the socket. Same protocol and endpoint source, different transport.

URL Construction

Endpoint owns { baseURL, path, query }. Each protocol route includes a canonical endpoint when the provider has one (e.g. https://api.openai.com/v1); provider helpers override endpoint fields by configuring the route before selecting a model. Routes that have no canonical URL (OpenAI-compatible Chat, GitHub Copilot) require configuration before execution.

For providers where the URL is derived from typed inputs (Azure resource name, Bedrock region), the provider helper configures the route endpoint before calling .model(...). Use AtLeastOne<T> from route/auth-options.ts for inputs that accept either of two derivation paths (Azure: resourceName or baseURL).

Provider Facades

Provider-facing APIs are configured facades over route values. Endpoint/auth/resource/API-version setup happens before model selection, and model selectors accept only a model or deployment id:

const openai = OpenAI.configure({ apiKey, baseURL })
const model = openai.responses("gpt-4o-mini")

const azure = Azure.configure({ resourceName, apiKey, apiVersion: "v1" })
const deployment = azure.responses("my-deployment")

const gateway = CloudflareAIGateway.configure({ accountId, gatewayId, gatewayApiKey, apiKey })
const proxied = gateway.model("openai/gpt-4o-mini")

Keep provider facades small and explicit:

Use branded ProviderID.make(...) and ModelID.make(...) where ids are constructed directly.
Use model for the default API path and named methods for provider-native alternatives such as OpenAI responses, responsesWebSocket, and chat.
Put provider-specific setup on .configure(...); do not add model(id, overrides) as a duplicate construction path.
Export lower-level routes arrays separately only when advanced internal wiring needs them.
Prefer apiKey as provider-specific sugar and auth as the explicit override; keep them mutually exclusive in provider option types with ProviderAuthOption.
Resolve apiKey → Auth with AuthOptions.bearer(options, "<PROVIDER>_API_KEY") (it honors an explicit auth override and falls back to Auth.config(envVar) so missing keys surface a typed Authentication error rather than a runtime crash).
Use separate top-level facades for products with different required setup, such as CloudflareAIGateway and CloudflareWorkersAI.

Provider.make(...) remains available for simple static provider definitions, but new built-in providers should prefer plain configured facades unless a helper removes real duplication without adding runtime behavior.

Provider Package Entrypoints

Catalog-selected native providers use package-like export paths from @opencode-ai/llm. They are internal entrypoints in one npm package, not separately published provider packages. Every entrypoint implements ProviderPackage.Definition and exposes model(modelID, settings), where settings are serializable provider configuration plus common headers, body, and limits overlays.

import { model } from "@opencode-ai/llm/providers/openai/responses"

const selected = model("gpt-5", {
  apiKey,
  transport: "websocket",
})

Keep semantic APIs as separate entrypoints, such as OpenAI chat and responses. Keep transport choices inside the semantic entrypoint settings, so OpenAI Responses HTTP and WebSocket share one entrypoint. Provider facades may still expose named selectors such as responsesWebSocket for direct typed call sites; the package-like contract maps its settings to those selectors before returning an executable Model.

Do not expose Route in provider package settings. Route composition stays an implementation detail behind model(...).

Folder layout

packages/llm/src/
  schema/                   canonical Schema model, split by concern
    ids.ts                  branded IDs, literal types, ProviderMetadata
    options.ts              Generation/Provider/Http options, Limits, Model, cache policy
    messages.ts             content parts, Message, ToolDefinition, LLMRequest
    events.ts               Usage, individual events, LLMEvent, PreparedRequest, LLMResponse
    errors.ts               error reasons, LLMError, ToolFailure
    index.ts                barrel
  llm.ts                    request constructors and convenience helpers
  route/
    index.ts                @opencode-ai/llm/route advanced barrel
    client.ts               Route.make + LLMClient.prepare/stream/generate
    executor.ts             RequestExecutor service + transport error mapping
    protocol.ts             Protocol type + Protocol.make
    endpoint.ts             Endpoint type + Endpoint.path
    auth.ts                 Auth type + Auth.bearer / Auth.apiKeyHeader / Auth.passthrough
    auth-options.ts         ProviderAuthOption shape, AuthOptions.bearer, AtLeastOne helper
    framing.ts              Framing type + Framing.sse
    transport/              transport implementations
      index.ts              Transport type + HttpTransport / WebSocketTransport namespaces
      http.ts               HttpTransport.httpJson — POST + framing
      websocket.ts          WebSocketTransport.json + WebSocketExecutor service
  protocols/
    shared.ts               ProviderShared toolkit used inside protocol impls
    openai-chat.ts          protocol + route (compose OpenAIChat.protocol)
    openai-responses.ts
    anthropic-messages.ts
    gemini.ts
    bedrock-converse.ts
    bedrock-event-stream.ts framing for AWS event-stream binary frames
    openai-compatible-chat.ts route that reuses OpenAIChat.protocol, no canonical URL
    openai-compatible-responses.ts route that reuses OpenAIResponses.protocol, no canonical URL
    utils/                  per-protocol helpers (auth, cache, media, tool-stream, ...)
  providers/
    openai-compatible.ts    generic Chat helper + family model helpers
    openai-compatible-responses.ts generic Responses helper
    openai-compatible-profile.ts family defaults (deepseek, togetherai, ...)
    azure.ts / amazon-bedrock.ts / cloudflare.ts / github-copilot.ts / google.ts / xai.ts / openai.ts / anthropic.ts / openrouter.ts
  tool.ts                   typed tool() helper
  tool-runtime.ts           narrow one-call typed tool dispatcher

The dependency arrow points down: providers/*.ts files import protocol routes and auth-option utilities; protocol modules import endpoint, auth, framing, and transport pieces. Protocols do not import provider facades. Lower-level modules know nothing about provider catalog metadata.

Shared protocol helpers

ProviderShared exports a small toolkit used inside protocol implementations to keep them focused on provider-native shapes:

joinText(parts) — joins an array of TextPart (or anything with a .text) with newlines. Use this anywhere a protocol flattens text content into a single string for a provider field.
parseToolInput(route, name, raw) — Schema-decodes a tool-call argument string with the canonical "Invalid JSON input for <route> tool call <name>" error message. Treats empty input as {}.
parseJson(route, raw, message) — generic JSON-via-Schema decode for non-tool bodies.
eventError(route, message, ...) — typed InvalidProviderOutput constructor for stream-time decode failures.
validateWith(decoder) — maps Schema decode errors to InvalidRequest. Route.make(...) uses this for body validation; lower-level routes can reuse it.
matchToolChoice(provider, choice, branches) — branches over LLMRequest["toolChoice"] for provider-specific lowering.

If you find yourself copying a 3-to-5-line snippet between two protocols, lift it into ProviderShared next to these helpers rather than duplicating.

Chronological System Updates

LLMRequest.system is the initial privileged prompt that applies ahead of the conversation. Message.system(...) is a separate, provider-neutral chronological operator update inside LLMRequest.messages; it applies only from its position in history onward and accepts text content only.

Native chronological system messages are route/model-specific. Anthropic Messages lowers them natively for Claude Opus 4.8 (claude-opus-4-8). Other routes and models intentionally lower the update in place into ordinary user-compatible text using this stable escaped representation:

<system-update>
...
</system-update>

The wrapped-user fallback preserves ordering while visibly lowering authority. Never silently pass a raw chronological role: "system" through a route that might reject it. Do not insert raw retrieved documents, tool output, or web content into privileged chronological system updates; keep untrusted content in ordinary user/tool channels.

Tools

Tool loops are represented in common messages and events:

const call = ToolCallPart.make({ id: "call_1", name: "lookup", input: { query: "weather" } })
const result = Message.tool({ id: "call_1", name: "lookup", result: { forecast: "sunny" } })

const followUp = LLM.request({
  model,
  messages: [Message.user("Weather?"), Message.assistant([call]), result],
})

Routes lower these into provider-native assistant tool-call messages and tool-result messages. Streaming providers should emit tool-input-delta events while arguments arrive, then a final tool-call event with parsed input.

Tool dispatch

LLM.stream(request) and LLM.generate(request) each run exactly one provider turn. Add tool schemas to request.tools with Tool.toDefinitions(tools). When a caller wants the package's typed one-call execution behavior, pass each canonical local tool-call event to ToolRuntime.dispatch(tools, call).

const get_weather = tool({
  description: "Get current weather for a city",
  parameters: Schema.Struct({ city: Schema.String }),
  success: Schema.Struct({ temperature: Schema.Number, condition: Schema.String }),
  execute: ({ city }) =>
    Effect.gen(function* () {
      // city: string  — typed from parameters Schema
      const data = yield* WeatherApi.fetch(city)
      return { temperature: data.temp, condition: data.cond }
      // return type checked against success Schema
    }),
})

const tools = { get_weather, get_time, ... }
const events = yield* LLM.stream(
  LLM.updateRequest(request, { tools: Tool.toDefinitions(tools) }),
).pipe(Stream.runCollect)

const call = Array.from(events).find(LLMEvent.is.toolCall)
if (call && !call.providerExecuted) {
  const dispatched = yield* ToolRuntime.dispatch(tools, call)
  // Persist call + dispatched.result, then construct the next request explicitly.
}

The dispatcher:

On tool-call: looks up the named tool, decodes input against parameters Schema, dispatches to the typed execute, encodes the result against success Schema, and returns canonical tool-result events.
Does not stream providers, construct Session events, schedule fibers, append history, count steps, or continue model rounds.
Leaves persistence and continuation to the enclosing product flow.

Handler dependencies (services, permissions, plugin hooks, abort handling) are closed over by the consumer at tool-construction time. Build the tools record inside an Effect.gen once and reuse it across many dispatches.

Errors must be expressed as ToolFailure. The runtime catches it and emits a tool-error event, then a tool-result of type: "error", so the model can self-correct on the next step. Anything that is not a ToolFailure is treated as a defect and fails the stream. Three recoverable error paths produce tool-error events:

The model called an unknown tool name.
Input failed the parameters Schema.
The handler returned a ToolFailure.

Provider-defined / hosted tools (Anthropic web_search / code_execution / web_fetch, OpenAI Responses web_search_call / file_search_call / code_interpreter_call / mcp_call / local_shell_call / image_generation_call / computer_use_call) pass through the runtime untouched:

Routes surface the model's call as a tool-call event with providerExecuted: true, and the provider's result as a matching tool-result event with providerExecuted: true.
Callers detect providerExecuted on tool-call and skip local dispatch — no handler is invoked and no tool-error is raised for "unknown tool". The provider already executed it.
Callers that continue should retain both events in explicit history when the protocol requires it. Anthropic encodes them back as server_tool_use + web_search_tool_result (or code_execution_tool_result / web_fetch_tool_result) blocks; OpenAI Responses callers typically use previous_response_id instead of resending hosted-tool items.

Add provider-defined tools to request.tools (no runtime entry needed). The matching route must know how to lower the tool definition into the provider-native shape; right now Anthropic accepts web_search / code_execution / web_fetch and OpenAI Responses accepts the hosted tool names listed above.

Protocol File Style

Protocol files should look self-similar. Provider quirks belong behind named helpers so a new route can be reviewed by comparing the same sections across files.

Section order

Use this order for every protocol module:

Public model input
Request body schema
Streaming event schema
Parser state
Request body construction (fromRequest)
Stream parsing (step and per-event handlers)
Protocol and route
Protocol route export

Rules

Keep protocol files focused on the protocol. Move provider-specific projection, signing, media normalization, or other bulky transformations into src/protocols/utils/*.
Use Effect.fn("Provider.fromRequest") for request body construction entrypoints. Use Effect.fn(...) for event handlers that yield effects; keep purely synchronous handlers as plain functions returning a StepResult that the dispatcher lifts via Effect.succeed(...).
Parser state owns terminal information. The state machine records finish reason, usage, and pending tool calls; emit one terminal finish event (or provider-error) for each completed response. If a provider splits reason and usage across events, merge them in parser state before flushing.
Emit exactly one terminal finish event for a completed response, normally after a matching step-finish. Use stream.terminal to stop reading when the provider has a completion sentinel; use stream.onHalt when the final event must be flushed after the framed stream ends.
Use shared helpers for repeated protocol policy such as text joining, usage totals, JSON parsing, and tool-call accumulation. ToolStream (protocols/utils/tool-stream.ts) accumulates streamed tool-call arguments uniformly.
Make intentional provider differences explicit in helper names or comments. If two protocol files differ visually, the reason should be obvious from the names.
Prefer dispatched per-event handlers (onMessageStart, onContentBlockDelta, ...) called from a small top-level step switch over a long if-chain. The dispatcher keeps the event surface visible at a glance.
Keep tests in the same conceptual order as the protocol: basic prepare, tools prepare, unsupported lowering, text/usage parsing, tool streaming, finish reasons, provider errors.

Review checklist

Can the file be skimmed side-by-side with openai-chat.ts without hunting for equivalent sections?
Are provider quirks named, isolated, and covered by focused tests?
Does request body construction validate unsupported common content at the protocol boundary?
Does stream parsing emit stable common events without leaking provider event order to callers?
Does toolChoice: "none" behavior read as intentional?

Recording Tests

Recorded tests use one cassette file per scenario. A cassette holds an ordered array of { request, response } interactions, so multi-step flows (tool loops, retries, polling) record into a single file. Use recordedTests({ prefix, requires }) and let the helper derive cassette names from test names:

const recorded = recordedTests({ prefix: "openai-chat", requires: ["OPENAI_API_KEY"] })

recorded.effect("streams text", () =>
  Effect.gen(function* () {
    // test body
  }),
)

Replay is the default. RECORD=true records fresh cassettes locally and requires the listed env vars; unset CI before recording because CI always forces replay. Cassettes are written as pretty-printed JSON so multi-interaction diffs stay reviewable.

Pass provider, protocol, and optional tags to recordedTests(...) / recorded.effect.with(...) so cassettes carry searchable metadata. Use recorded-test filters to replay or record a narrow subset without rewriting a whole file:

RECORDED_PROVIDER=openai matches tests tagged with provider:openai; comma-separated values are allowed.
RECORDED_PREFIX=openai-chat matches cassette groups by recordedTests({ prefix }); comma-separated values are allowed.
RECORDED_TAGS=tool requires all listed tags to be present, e.g. RECORDED_TAGS=provider:togetherai,tool.
RECORDED_TEST="streams text" matches by test name, kebab-case test id, or cassette path.

Filters apply in replay and record mode. Combine them with RECORD=true when refreshing only one provider or scenario.

Binary response bodies. Most providers stream text (SSE, JSON). The recorder treats known textual media types (text/*, JSON/XML structured types, JavaScript, forms, YAML, and SVG) as text and stores every other response as base64 with bodyEncoding: "base64". This preserves binary formats such as AWS event-stream frames without a lossy UTF-8 round trip.

Matching strategy. A runtime request atomically claims the first unused recorded interaction that matches its method, URL, allow-listed headers, and canonical JSON body. Distinct requests may replay in any order or concurrently. Repeated identical requests consume their matching responses in cassette order, preserving deterministic retry and polling behavior. scriptedResponses (in test/lib/http.ts) is the deterministic counterpart for tests that don't need a live provider; it scripts response bodies in order without reading from disk.

Do not blanket re-record an entire test file when adding one cassette. RECORD=true rewrites every recorded case that runs, and provider streams contain volatile IDs, timestamps, fingerprints, and obfuscation fields. Prefer deleting the one cassette you intend to refresh, or run a focused test pattern that only registers the scenario you want to record. Keep stable existing cassettes unchanged unless their request shape or expected behavior changed.

AGENTS.md 25 KB Пермалинк Историја Датотека