[rekram1-node]

Problem

The Promise and Effect APIs had different shapes. Promise users created a client with AI.make(), called ai.llm.*, and disposed it. Effect users called LLM.*, Image.* and Decision.*, with compaction on LLMClient. The Promise root also re-exported the Effect LLM, Image and Decision values, while missing data types such as Message and Media.

Change

Both entrypoints now export one ai object with the same shape. Text calls sit directly on ai; other kinds keep a namespace. The Promise one returns Promise/AsyncIterable; the Effect one returns Effect/Stream.

// Before
import { AI } from "@opencode/ai"
const ai = AI.make()
await ai.llm.generate({ model, prompt: "Hi" })
await ai.dispose()

import { LLM, Image, LLMClient } from "@opencode/ai/effect"
yield* LLM.generate({ model, prompt: "Hi" })
yield* Image.generate({ model: imageModel, prompt: "A cat" })
yield* LLMClient.compact(request)

// After
import { ai } from "@opencode/ai" // or "@opencode/ai/effect" for Effect/Stream
await ai.generate({ model, prompt: "Hi" })
await ai.image.generate({ model: imageModel, prompt: "A cat" })
await ai.compact(request)

ai
├── request  generate  stream  loop  generateObject  compact  canCompact
├── image.{request, generate, stream, start, resume}
├── decision.{request, generate}
└── experimental.{speech, transcription, video}

• The default Promise ai builds its runtime on first use and has no dispose().
• The root @opencode/ai exports the data model a Promise user needs: Message, Media, AIError and its reasons, isRetryable, isContextOverflow, and the request, response and event types for text, image and decision. It exports no Effect services or layers (LLMClient, AIClient, ToolRuntime, Generation stay in /effect).
• /effect does not re-export Speech, Transcription or Video; they stay in /experimental and are reachable as ai.experimental.*.
• compact and canCompact are no longer on LLMClient; ai.compact and ai.canCompact replace them. LLMClient keeps Service, layer, stream and generate, and the service interface keeps its compact method.
• test/promise-parity.types.ts compares the two ai objects fully: same keys, same signatures for the top-level text calls and for every namespace.
• I updated every caller in core, the tests, the README and the tutorial.

AI.make({ layer, clients }) stays for now. It is still the only way to inject a custom executor or test layers. It is temporary until the fetch and test-double decisions land. Clients from AI.make keep dispose().

Renames (breaking, for release notes)

Before	After
ai.llm.request / generate / stream / loop / generateObject / compact / canCompact (Promise)	ai.request / generate / ...
ai.dispose() on the default client	removed (nothing to dispose)
LLM.request / generate / stream / loop / generateObject (/effect)	ai.request / generate / stream / loop / generateObject
LLMClient.compact / canCompact	ai.compact / canCompact
Image.request / generate / stream / start / resume (/effect)	ai.image.*
Decision.request / generate (/effect)	ai.decision.*
Speech.*, Transcription.*, Video.* (/experimental)	unchanged, also ai.experimental.speech / transcription / video
LLM, Image, Decision values exported from the root	removed (use ai from the root)

LLM is still exported from both entrypoints, but only as a type namespace, so LLM.RequestInput and LLM.LoopInput keep working. LLMRequest, LLMResponse and the other type names are unchanged.

Out of scope

• Type renames, string model ids, and where fetch, middleware and test doubles go (separate decisions).
• LLMClient.generate and LLMClient.stream stay; they take a built LLMRequest and core uses the service directly.

---

Added by the opencode bot (AI-generated).

Public API: @opencode/ai: replaces ai.llm.* with ai.*, removes LLM/Image/Decision values and LLMClient.compact/canCompact in favour of ai (breaking), adds root data-model exports

[opencodejr]

Looks good. Both entrypoints now expose the same ai shape. The parity type test, the packages/ai tests and type checks, and the touched packages/core tests and type check all pass. A quick script importing both entrypoints confirmed matching namespaces and operations. No other package in the repo still uses the removed values.

Public API impact on @opencode/ai. This is breaking, as the description's rename table says:

• Removed: the LLM, Image and Decision values from @opencode/ai and @opencode/ai/effect. Callers move to ai.text.*, ai.image.* and ai.decision.*. LLM remains, but only as a type namespace.
• Renamed: on the Promise client, ai.llm.* → ai.text.*. This also applies to clients from AI.make.
• Changed type: the default ai is now typed without dispose(). At runtime the method still exists, so untyped callers can still dispose the shared runtime.
• Added: ai from @opencode/ai/effect, including ai.text.compact/canCompact and ai.experimental.*.
• Added: many data-model exports on the root: Message, Media, AIError and its reasons, isRetryable, isContextOverflow, isContextOverflowFailure, and the text, image and decision request, response and event types.

The CLI, server HTTP API, plugin API and config are unaffected.

[opencodejr]

My earlier review raised no concerns, so there was nothing to address. The new commit looks good: text calls now sit directly on ai in both entrypoints, and compact/canCompact moved from LLMClient to ai. Nothing in the repo still uses LLMClient.compact, ai.text.* or ai.llm.*. Type checks and tests pass in packages/ai and packages/core, and a quick script importing both entrypoints shows the same keys on each ai.

Public API impact on @opencode/ai. This is breaking, as your rename table says:

• Renamed: on the Promise client, ai.llm.* → ai.*. This also applies to clients from AI.make.
• Removed: the LLM, Image and Decision values from @opencode/ai and @opencode/ai/effect. Use ai.*, ai.image.* and ai.decision.* instead; LLM remains, but only as a type namespace.
• Removed: LLMClient.compact and canCompact → ai.compact and ai.canCompact.
• Removed: the Speech, Transcription and Video re-exports from /effect. They stay in /experimental and are also reachable as ai.experimental.*.
• Changed type: the default ai is now typed without dispose().
• Added: ai from @opencode/ai/effect.
• Added: data-model exports on the root, such as Message, Media, AIError and its reasons, isRetryable and isContextOverflow.

The CLI, server, plugin API and config are unaffected.