AI Library Design

Discussion draft. This document describes the intended replacement for the current private @opencode-ai/llm API. Names and exact TypeScript signatures are illustrative until implementation, but the domain boundaries and defaults are deliberate.

Status

• Proposed package: @opencode-ai/ai
• Initial stable domain: LLM
• Release posture: pre-1.0, with a stable-core intent
• Migration posture: clean break; do not preserve compatibility aliases
• Primary audience: general-purpose TypeScript developers using Effect
• Secondary audience: OpenCode and other durable agent runtimes

The package name leaves room for future domains such as embeddings, images, and speech. Those domains are not part of this design and should not be forced into the LLM run/turn model.

Goals

1. Make a useful model call require very little code.
2. Make the default behavior good enough that most callers do not configure it.
3. Let advanced callers inspect, transform, or replace every important stage.
4. Keep provider quirks behind provider and protocol boundaries.
5. Preserve one provider turn as an explicit primitive for durable runtimes.
6. Keep serializable request data separate from process-local execution behavior.
7. Make unsupported combinations fail locally with useful typed errors.
8. Stay Effect-native without making package-specific service provisioning part of every call site.

Non-goals

• A global provider or model registry
• Durable agent orchestration or persistence
• Session history ownership
• Permission handling
• Cost billing or accounting guarantees
• Runtime model-catalog network requests
• Compatibility with the current private API
• Designing embeddings, image generation, speech, or transcription now

Design Principles

Progressive disclosure

The API has four layers:

1. Run a model with LLM.generate or LLM.stream.
2. Control one provider turn with LLM.generateTurn or LLM.streamTurn.
3. Customize execution with model defaults, call options, hooks, and provider configuration.
4. Author providers with experimental provider definitions and protocols.

Normal documentation should teach only the first layer initially.

Values over registries

Provider definitions, configured providers, models, protocols, tools, and hooks are immutable values. Importing a provider does not register anything globally.

Portable data, local behavior

Requests, messages, tool definitions, events, usage, and result projections are plain immutable data with schemas. Configured models, executable tools, hooks, and provider definitions may contain functions and Effect requirements and are not serializable.

Strong defaults, explicit overrides

Defaults should make common calls correct without hiding where behavior comes from. Overrides compose in a documented order and never require patching installed dependencies.

Domain Model

Provider Definition

An immutable, declarative description of a provider integration. It owns model selection, option schemas, catalog corrections, protocols, and provider-wide hooks. It is an experimental provider-authoring API.

Configured Provider

A provider definition bound to deployment concerns such as credentials, endpoint, transport, and provider headers.

configure(...) is intentionally deployment-only. It does not establish hidden generation defaults.

Model

A process-local executable model value selected from a configured provider. It contains identity, capabilities, pricing metadata, provider-specific option types, reusable request-behavior defaults, and hidden execution behavior.

Normal users do not need to learn the current Route composite. Protocol, endpoint, auth, transport, and hooks are bound behind Model.

Request

Portable, model-independent input for a model call. It may contain system instructions, messages, tool definitions, generation controls, output intent, cache policy, and metadata. It does not contain a configured model, executable tool handlers, or hooks.

Provider Turn

Exactly one request to a model provider and its normalized response. It does not execute local tools or continue the conversation.

Model Run

A complete interaction consisting of one or more provider turns. A run executes local tools, appends their results, and continues until the model completes or a stopping condition matches.

TurnResult

The result of exactly one provider turn.

GenerateResult

The result of a complete model run. It preserves every turn, tool activity, aggregate usage, and estimated cost while exposing shortcuts to the final output.

Protocol

The provider-wire contract that lowers portable requests into provider-native bodies and raises provider-native stream events into normalized turn events. Protocols are public, reusable, fully inspectable, and immutably patchable, but the entire protocol-authoring API is experimental.

Happy Path

Effect

import { Effect } from "effect"
import { LLM } from "@opencode-ai/ai"
import { OpenAI } from "@opencode-ai/ai/providers/openai"

// Environment-based credentials are a provider default. No LLMClient layer is
// required: the Effect exposes standard runtime dependencies directly.
const model = OpenAI.model("gpt-4.1-mini")

const program = Effect.gen(function* () {
  const result = yield* LLM.generate({
    model,
    system: "You are concise.",
    prompt: "Explain Effect in one sentence.",
  })

  // `generate` always returns GenerateResult, even when the run has one turn.
  console.log(result.text)
  console.log(result.turns.length) // 1
  console.log(result.usage)
  console.log(result.cost) // Estimated cost, or undefined if any turn is unpriced.
})

The required Effect environment should contain standard services plus services required by tools and hooks. It should not contain an LLMClient wrapper service.

Current API

The current README appears similarly small but omits the package-specific service and layer required at runtime:

// Current API: this request contains an executable model/route value.
const request = LLM.request({
  model: OpenAI.configure({ apiKey }).responses("gpt-4o-mini"),
  prompt: "Say hello.",
})

// Current API: this performs one provider turn, despite the broad name.
const response = yield * LLM.generate(request)

// Current API: execution also needs LLMClient.layer and RequestExecutor services.

The proposal removes mandatory request construction, removes package-specific runtime provisioning, and makes generate mean a complete run.

Provider And Model Selection

Environment defaults

import { OpenAI } from "@opencode-ai/ai/providers/openai"

// Open strings receive autocomplete for IDs from the generated models.dev
// snapshot but continue to accept newly released and fine-tuned model IDs.
const model = OpenAI.model("gpt-4.1-mini")

Deployment configuration

const openai = OpenAI.configure({
  apiKey,
  baseURL: "https://gateway.example.com/openai/v1",
  headers: {
    "x-tenant": "acme",
  },
})

const model = openai.model("gpt-4.1-mini")

configure(...) owns deployment concerns only:

• Credentials and authentication
• Base URL and deployment location
• Transport selection
• Provider/deployment headers
• Other provider-specific connection setup

It does not own temperature, maximum output tokens, cache policy, retry policy, tools, output schema, or system instructions.

Reusable model defaults

const model = OpenAI.model("gpt-4.1-mini", {
  generation: {
    temperature: 0.2,
    maxTokens: 2_000,
  },
  cache: "auto",
  provider: {
    store: false,
  },
})

The second argument may default request behavior but not prompt/history or executable tools. Call-level values override model defaults.

Provider-specific options are inferred from the concrete model:

yield *
  LLM.generate({
    model: OpenAI.model("gpt-4.1-mini"),
    prompt: "Hello",
    provider: {
      store: false,
      // OpenAI-specific autocomplete here; no `{ openai: ... }` nesting.
    },
  })

Code choosing between providers dynamically must narrow the model before using provider-specific options. Portable generation controls remain available without narrowing.

Current API

// Current API mixes deployment configuration and reusable request behavior.
const model = OpenAI.configure({
  apiKey,
  generation: { maxTokens: 160 },
  providerOptions: {
    openai: { store: false },
  },
}).model("gpt-4o-mini")

The proposal separates deployment configuration from selected-model behavior and removes provider-keyed option bags when a concrete model already identifies the provider.

Requests

Inline input

const result =
  yield *
  LLM.generate({
    model,
    system: "You are concise.",
    prompt: "Summarize this pull request.",
    generation: { maxTokens: 500 },
  })

Reusable portable request

const request = LLM.request({
  system: "You are concise.",
  prompt: "Summarize this pull request.",
  generation: { maxTokens: 500 },
})

// Bind process-local execution behavior only when running.
const result = yield * LLM.generate({ model, request })

LLM.request(...) returns a plain immutable object. Use ordinary object spread to derive another request:

const longer = {
  ...request,
  generation: {
    ...request.generation,
    maxTokens: 1_000,
  },
}

There is no LLM.updateRequest(...) helper and no request Schema class.

Conversation history

import { Message } from "@opencode-ai/ai"

const request = LLM.request({
  system: "You are concise.",
  messages: [
    Message.user("What is Effect?"),
    Message.assistant("A TypeScript library for typed functional effects."),
    Message.user("Why would I use it?"),
  ],
})

Message helpers return plain immutable data. Object literals remain valid when they satisfy the same input type.

system stays separate from chronological messages because it is the initial privileged instruction. A chronological system message represents an instruction change at a specific point in history.

Complete Runs

Automatic local tool loop

import { Effect, Schema } from "effect"
import { LLM, Tool } from "@opencode-ai/ai"

const tools = {
  getWeather: Tool.make({
    description: "Get current weather for a city.",
    parameters: Schema.Struct({ city: Schema.String }),
    success: Schema.Struct({ forecast: Schema.String }),

    // Tool service requirements and typed errors flow into LLM.generate's
    // Effect environment/error model instead of being erased.
    execute: ({ city }) => Weather.get(city),

    // Expected domain failures need an explicit model-visible representation.
    formatError: (error) => ({
      type: "text",
      text: `Weather lookup failed: ${error.message}`,
    }),
  }),
}

const result =
  yield *
  LLM.generate({
    model,
    prompt: "What is the weather in London?",
    tools,
  })

// The runtime advertises definitions, dispatches calls, records results, and
// continues provider turns automatically.
console.log(result.text)
console.log(result.turns)
console.log(result.toolExecutions)

The default stopping condition is equivalent to:

stopWhen: StopWhen.turnCount(20)

This matches the Vercel AI SDK ToolLoopAgent default. Reaching the limit is a successful result with stopReason: "max-turns", not an Effect failure.

Custom stopping

const result =
  yield *
  LLM.generate({
    model,
    prompt,
    tools,
    stopWhen: StopWhen.any(StopWhen.turnCount(8), StopWhen.hasToolCall("finalize")),
  })

stopWhen accepts one predicate. Composition is explicit through combinators such as StopWhen.any, StopWhen.all, and StopWhen.not.

Successful run stop reasons are closed:

type RunStopReason = "completed" | "max-turns" | "stop-condition"

Tool concurrency

Independent tool calls emitted in one turn run concurrently with a bounded, configurable concurrency limit. Results are appended in deterministic emitted order. The runtime does not infer dependencies between tool calls; the model must request dependent calls in separate turns.

Tools may declare an optional timeout. The overall run timeout still applies.

Current API

Today callers must manually bridge every layer:

const request = LLM.request({
  model,
  prompt,
  tools: Tool.toDefinitions(tools),
})

const events = yield * LLM.stream(request).pipe(Stream.runCollect)
const call = Array.from(events).find(LLMEvent.is.toolCall)

if (call && !call.providerExecuted) {
  const dispatched = yield * ToolRuntime.dispatch(tools, call)
  const followUp = LLM.updateRequest(request, {
    messages: [...request.messages, Message.assistant([call]), Message.tool({ ...call, result: dispatched.result })],
  })
  // Caller must invoke the provider again and repeat the loop.
}

That explicit flow remains possible through turn APIs, but it is no longer the only tool experience.

One Provider Turn

OpenCode and other durable runtimes need to own persistence, tool settlement, and continuation. They use the explicit turn API:

const result =
  yield *
  LLM.generateTurn({
    model,
    request,
    // Definitions only. generateTurn never dispatches local handlers.
    tools: {
      getWeather: Tool.definition({
        description: "Get current weather for a city.",
        parameters: WeatherInput,
      }),
    },
  })

// Persist the TurnResult and settle calls durably before the next turn.
for (const call of result.toolCalls) {
  // Application-owned dispatch and persistence.
}

generateTurn and streamTurn make exactly one provider request. They never execute a local tool and never continue automatically.

This separation is load-bearing:

• generate / stream: complete Model Run
• generateTurn / streamTurn: one Provider Turn

Portable Tool Definitions

A portable request may declare serializable definitions, but executable handlers are bound at run time:

const request = LLM.request({
  prompt: "What is the weather in London?",
  tools: {
    getWeather: Tool.definition({
      description: "Get current weather for a city.",
      parameters: WeatherInput,
    }),
  },
})

const result =
  yield *
  LLM.generate({
    model,
    request,
    tools: {
      getWeather: Tool.make({
        description: "Get current weather for a city.",
        parameters: WeatherInput,
        success: WeatherOutput,
        execute: getWeather,
        formatError,
      }),
    },
  })

Definitions and handlers match by record key. Before the first provider call, the runtime validates that every local definition has a compatible executable binding. Missing or incompatible bindings fail with a typed tool-binding error.

Provider-hosted tools are distinct typed values:

const result =
  yield *
  LLM.generate({
    model: OpenAI.model("gpt-4.1"),
    prompt: "Find today's relevant announcements.",
    tools: {
      search: OpenAI.tool.webSearch({ searchContextSize: "medium" }),
    },
  })

Hosted tools do not pretend to have local handlers, and callers do not inspect a providerExecuted boolean to decide whether dispatch is safe.

Streaming

Run stream

LLM.stream returns an Effect Stream<RunEvent, LLMError, Requirements>. Run events explicitly expose orchestration boundaries:

const program = LLM.stream({ model, prompt, tools }).pipe(
  Stream.tap((event) =>
    Effect.sync(() => {
      switch (event.type) {
        case "run-start":
          break
        case "turn-start":
          break
        case "turn-event":
          // Normalized text, reasoning, tool-call, usage, and finish events.
          if (event.event.type === "text-delta") {
            process.stdout.write(event.event.text)
          }
          break
        case "tool-start":
          break
        case "tool-finish":
          break
        case "turn-finish":
          break
        case "run-finish":
          // Contains the same full GenerateResult returned by LLM.generate.
          console.log(event.result.usage)
          break
      }
    }),
  ),
  Stream.runDrain,
)

Exact event tag spelling remains an implementation detail to finalize, but the algebra is settled:

• A separate RunEvent union for run, turn, and tool lifecycle
• A focused TurnEvent union for normalized provider output
• streamTurn emits only TurnEvent
• The terminal run event contains the full GenerateResult

External cancellation remains Effect interruption. It does not fabricate a successful result with an interrupted stop reason.

Structured Output

Structured output is an option on generate, not a separate operation:

const Weather = Schema.Struct({
  city: Schema.String,
  forecast: Schema.String,
  highCelsius: Schema.Number,
})

const result =
  yield *
  LLM.generate({
    model,
    prompt: "Give me today's weather for London.",
    output: Weather,
  })

// Inferred from Weather.
result.output.city

The model declaration and protocol select the best reliable strategy:

1. Provider-native structured output when supported and reliable
2. Forced tool output when required as a compatibility fallback
3. Typed unsupported-capability failure before network execution when neither is available

Advanced callers may override the strategy when exact provider semantics matter.

Current API

// Current API is a separate operation and always forces a synthetic tool.
const result =
  yield *
  LLM.generateObject({
    model,
    prompt,
    schema: Weather,
  })

The proposal unifies generation and lets capabilities choose the strategy rather than permanently encoding one cross-provider workaround.

Model Catalog

models.dev is the release-time source for:

• Model ID suggestions
• Capabilities and modalities
• Context and output limits
• Pricing
• Other available model metadata

The package ships a generated, versioned snapshot. Normal execution performs no catalog network requests.

Provider definitions may correct generated metadata where protocol-specific knowledge is more accurate. Precedence is:

models.dev snapshot
  < provider-definition correction
  < provider configuration override
  < model-selection override
  < call override

Unknown model IDs inherit only capabilities guaranteed by the selected protocol. Unsupported request capabilities fail before network execution unless the caller explicitly overrides the model declaration.

Usage And Cost

GenerateResult aggregates normalized usage across every turn, including cache read/write usage where providers report it.

It also exposes estimated cost using the generated models.dev pricing snapshot:

result.usage.inputTokens
result.usage.outputTokens
result.usage.cacheReadInputTokens
result.usage.cacheWriteInputTokens

result.cost?.total
result.cost?.currency // e.g. "USD"

Cost is an estimate, not a billing guarantee. If reliable pricing is unavailable for any turn, aggregate run cost is unavailable rather than partial or silently zero. Per-turn metadata should retain the catalog/pricing identity used so an estimate can be explained.

Caching

Prompt caching remains "auto" by default. The library places protocol-aware cache boundaries where explicit caching is supported and does nothing on the wire where providers cache implicitly.

yield *
  LLM.generate({
    model,
    prompt,
    cache: "none", // Explicit opt-out.
  })

Granular cache policy remains available as an advanced request option.

Retries, Timeouts, And Cancellation

Retries

The default retry policy is deliberately conservative:

• Retry bounded transient transport and rate-limit failures
• Retry only before observable output
• Never silently retry after ambiguous tool execution or other side effects
• Allow each call to override or disable retry behavior

Retry configuration is call-scoped only. Provider and model configuration do not silently inherit custom retry policies.

Timeouts

yield *
  LLM.generate({
    model,
    prompt,
    timeout: "2 minutes", // Entire run, including tools.
    turnTimeout: "30 seconds", // Each provider turn.
    tools,
  })

Exact Duration input spelling follows Effect conventions. Individual tools may also declare optional timeouts.

Cancellation

• Effect API: fiber interruption
• Promise API: AbortSignal, rejecting with a recognizable abort error
• Cancellation is not a successful run stop reason

Hooks

Stable high-level hooks exist at five named stages:

1. Canonical request
2. Provider-native body
3. Prepared transport request
4. Normalized event
5. Error

Hooks are Effectful. They may transform the stage value or fail with a typed error. They may not secretly short-circuit execution, synthesize a response, retry, or redirect control flow.

const model = OpenAI.model("gpt-4.1", {
  hooks: {
    request: (request) =>
      Effect.succeed({
        ...request,
        metadata: { ...request.metadata, tenant: "acme" },
      }),
    body: (body, context) => auditBody(body, context),
    transport: (request) => signInternalGatewayRequest(request),
    event: (event) => redactProviderMetadata(event),
    error: (error) => classifyInternalError(error),
  },
})

Hook scopes compose in this order:

provider-definition hooks -> model hooks -> call hooks

Each hook sees the prior hook's output. Replacement requires an explicit definition-level patch, not accidental last-writer-wins semantics.

Provider-definition hooks are authored by provider integrations. They are not passed through Provider.configure(...), which remains deployment-only.

HTTP And Provider Escape Hatches

The request customization ladder is:

1. Portable generation controls
2. Model-typed provider options
3. Stable staged hooks
4. Serializable HTTP/body overlays
5. Experimental provider-definition or protocol patching

yield *
  LLM.generate({
    model,
    prompt,
    http: {
      headers: { "x-experimental": "1" },
      query: { debug: "true" },
      body: { newlyReleasedProviderField: true },
    },
  })

Raw overlays are intentional last-resort support for provider features that ship before the library has a typed option.

Provider-Native Metadata

Normalized message/content/event unions remain closed and exhaustive. Unknown or provider-required round-trip data lives in caller-writable providerMetadata.

const assistant = Message.assistant([
  {
    type: "reasoning",
    text: "...",
    providerMetadata: {
      openai: {
        // Opaque provider data needed for replay or continuation.
      },
    },
  },
])

Protocols validate metadata they consume. The field is an escape hatch, not a portable semantic guarantee.

Error Model

The Effect error channel is a tagged domain union rather than one LLMError wrapper with nested reasons. Illustrative categories:

type LLMError =
  | AuthenticationError
  | InvalidRequestError
  | UnsupportedCapabilityError
  | ToolBindingError
  | TransportError
  | ProviderResponseError
  | InvalidProviderOutputError
  | HookError

Each error retains relevant provider/model/turn/stage context and its underlying cause where available.

Expected tool errors keep their own typed error channel. Tool.make requires an explicit mapping before such errors become model-visible tool results. Expected mapped failures let the model recover; defects and interruption fail the run.

Observability

The core library emits Effect-native spans and metrics for:

• Model runs
• Provider turns
• Provider requests
• Retries
• Tool executions

Default telemetry records metadata only:

• Provider and model identity
• Timing
• Token/cache usage
• Estimated cost availability
• Finish and stop reasons
• Retry counts
• Tool names

Prompts, model output, tool arguments, and tool results are never recorded by default. Explicit hooks or telemetry configuration may opt into content capture.

Promise API

Promise wrappers live at a separate subpath so the root remains unambiguously Effect-first:

import { LLM } from "@opencode-ai/ai/promise"
import { OpenAI } from "@opencode-ai/ai/providers/openai"

const result = await LLM.generate({
  model: OpenAI.model("gpt-4.1-mini"),
  prompt: "Explain Effect in one sentence.",
  signal: abortController.signal,
})

Streaming returns an AsyncIterable<RunEvent>:

for await (const event of LLM.stream({ model, prompt, signal })) {
  if (event.type === "turn-event" && event.event.type === "text-delta") {
    process.stdout.write(event.event.text)
  }
}

Top-level Promise functions use a default runtime for built-in services. Custom Effect service requirements use a configured client:

const client = LLM.makeClient({
  layer: Layer.mergeAll(WeatherLive, AuditLive),
})

const result = await client.generate({ model, prompt, tools })

The Promise API mirrors Effect semantics. It does not invent different run, error, stopping, or cancellation behavior.

Schemas

Schemas live in a dedicated namespace/subpath instead of flooding root exports:

import { LLMSchema } from "@opencode-ai/ai/schema"

const request = yield * Schema.decodeUnknown(LLMSchema.Request)(input)

Schemas cover only serializable domain values:

• Requests and messages
• Portable tool definitions
• Turn and run events
• Serializable result projections
• Usage and cost estimates
• Tagged errors where serializable
• Provider metadata containers

Configured models, executable tools, hooks, provider definitions, and protocols are process-local behavior and do not receive fake serialization schemas.

Provider Authoring

Provider authoring is public but experimental.

Declarative provider definition

import { Provider, Protocol } from "@opencode-ai/ai/provider"

export const ExampleAI = Provider.define({
  id: "example",
  options: ExampleProviderOptions,
  configure: configureExampleDeployment,
  protocols: {
    responses: ExampleResponses,
  },
  models: ({ deployment, catalog }) => ({
    model: (id, defaults) =>
      Provider.model({
        id,
        deployment,
        protocol: ExampleResponses,
        metadata: catalog.model(id),
        defaults,
      }),
  }),
  catalog: generatedExampleCatalog,
  corrections: exampleCatalogCorrections,
  hooks: exampleProviderHooks,
})

The exact builder fields need implementation design, but it must remain one declarative immutable object, infer provider option types, and support .with(...) patching. It must not register globally.

Built-ins export their immutable definition for advanced forking:

import { OpenAI } from "@opencode-ai/ai/providers/openai"

const PatchedOpenAI = OpenAI.definition.with({
  protocols: {
    responses: OpenAI.protocols.responses.with({
      // Explicit immutable stage patch.
      body: {
        fromRequest: patchResponsesBody,
      },
    }),
  },
})

Protocols

A protocol exposes all native types and stages:

• Provider-native request body and schema
• Transport frame type
• Provider-native event and schema
• Parser state
• Request lowering
• Event stepping
• Terminal detection and final flushing

Every stage is immutably patchable. This is deliberately more open than the AI SDK integrations that motivated this package.

const PatchedResponses = OpenAIResponses.with({
  body: {
    fromRequest: (request) =>
      OpenAIResponses.body.fromRequest(request).pipe(Effect.map((body) => ({ ...body, custom_field: true }))),
  },
  stream: {
    step: patchResponsesStep,
  },
})

Protocol body, frame, native event, and parser-state types are exported. Because provider wire formats change often, these types and patch APIs are explicitly experimental and do not receive the high-level API's compatibility promise.

Package Surface

Illustrative export layout:

@opencode-ai/ai
  LLM
  Message
  Tool
  StopWhen
  stable domain types

@opencode-ai/ai/promise
  Promise/AsyncIterable LLM facade

@opencode-ai/ai/schema
  serializable domain schemas

@opencode-ai/ai/provider
  experimental Provider and Protocol authoring APIs

@opencode-ai/ai/providers/openai
@opencode-ai/ai/providers/anthropic
@opencode-ai/ai/providers/google
...

Providers are imported through individual subpaths. The root does not export all providers, and there is no preferred all-providers barrel.

Defaults

Clean-break Migration

The redesign intentionally removes or changes these current concepts:

OpenCode should migrate to generateTurn / streamTurn, preserving its durable prompt admission, persistence, permission, tool settlement, and continuation boundaries. It should not use the automatic run API for Session orchestration.

Remaining Implementation-level Questions

These do not reopen the main design:

1. Exact RunEvent and TurnEvent tag names and payloads
2. Exact GenerateResult shortcut fields for text, reasoning, output, and messages
3. Exact Provider definition TypeScript shape needed for strong inference
4. Exact protocol .with(...) patch syntax and replacement semantics
5. Exact Duration input fields and names
6. Exact models.dev generation pipeline and correction-file format
7. Exact cost representation and decimal arithmetic strategy
8. Exact default retry schedule and bounded tool concurrency number
9. Whether request-level serializable HTTP overlays belong in the stable schema
10. Which tagged errors are serializable versus process-local

These should be resolved with call-site sketches and implementation spikes rather than by changing the domain boundaries above.