Add abstract service interface packages

@deepseek-ai/dsh-llm: provider-neutral content-block vocabulary
(merge-extensible maps), raw StreamChunk protocol, ToolSchema,
abstract LlmAdapter, LlmService adapter registry, BlockAssembler.

@deepseek-ai/dsh-session: event-sourced Session (append-only log,
deriveMessages; context/steering render as tagged envelopes),
SessionStore, session/event + awaited session/flush durability seam.

@deepseek-ai/dsh-system-prompt: ordered sections + tool-schema
providers; assemble() through the system-prompt/assemble waterfall.
Tool schemas are part of the assembly by design.

@deepseek-ai/dsh-tools: tool registry feeding schemas into the
assembly; execute() through the tools/execute waterfall (the single
sandbox/permission/hook seam).

@deepseek-ai/dsh-agent: Agent interface (send/steer/inject/abort,
spawn/fork TODO seams), AgentRegistry, and the full agent/* event
taxonomy so plugins never depend on the concrete loop.

packages/llm/src/assembler.ts

@@ -0,0 +1,115 @@
+import type { ContentBlock, FinishReason, GenerateResult, Message, StreamChunk, TokenUsage } from './types.ts'
+
+/**
+ * Incrementally assembles raw {@link StreamChunk}s into complete
+ * {@link ContentBlock}s and a final assistant {@link Message}.
+ *
+ * This is the single shared assembly implementation: the agent loop feeds it
+ * while logging raw chunks for replay fidelity, and `LlmService.generate()` /
+ * `streamBlocks()` use it to offer assembled views of the same stream.
+ */
+export class BlockAssembler {
+  private partials = new Map<number, {
+    blockType: string
+    text: string
+    toolCallId?: string
+    toolCallName?: string
+    toolCallArguments: string
+    block?: ContentBlock
+  }>()
+
+  private order: number[] = []
+  private _usage: TokenUsage | undefined
+  private _finish: FinishReason | undefined
+
+  /**
+   * Feed one chunk. Returns the completed block when the chunk closes one
+   * (either an explicit `block-end` or an implicit close), otherwise undefined.
+   */
+  push(chunk: StreamChunk): ContentBlock | undefined {
+    switch (chunk.type) {
+      case 'block-start': {
+        if (!this.partials.has(chunk.index)) this.order.push(chunk.index)
+        this.partials.set(chunk.index, {
+          blockType: chunk.blockType,
+          text: '',
+          toolCallArguments: '',
+        })
+        return
+      }
+      case 'text-delta':
+      case 'reasoning-delta': {
+        const partial = this.ensure(chunk.index, chunk.type === 'text-delta' ? 'text' : 'reasoning')
+        partial.text += chunk.text
+        return
+      }
+      case 'tool-call-delta': {
+        const partial = this.ensure(chunk.index, 'tool-call')
+        partial.toolCallId = chunk.id
+        if (chunk.name) partial.toolCallName = chunk.name
+        partial.toolCallArguments += chunk.argumentsDelta
+        return
+      }
+      case 'block-end': {
+        const partial = this.ensure(chunk.index, chunk.block.type)
+        partial.block = chunk.block
+        return chunk.block
+      }
+      case 'usage': {
+        this._usage = chunk.usage
+        return
+      }
+      case 'finish': {
+        this._finish = chunk.reason
+        return
+      }
+    }
+  }
+
+  private ensure(index: number, blockType: string) {
+    let partial = this.partials.get(index)
+    if (!partial) {
+      partial = { blockType, text: '', toolCallArguments: '' }
+      this.partials.set(index, partial)
+      this.order.push(index)
+    }
+    return partial
+  }
+
+  /** Assemble all blocks seen so far, in stream order. */
+  blocks(): ContentBlock[] {
+    return this.order.map((index) => {
+      const partial = this.partials.get(index)!
+      if (partial.block) return partial.block
+      switch (partial.blockType) {
+        case 'text': return { type: 'text', text: partial.text }
+        case 'reasoning': return { type: 'reasoning', text: partial.text }
+        case 'tool-call': return {
+          type: 'tool-call',
+          id: partial.toolCallId ?? `call-${index}`,
+          name: partial.toolCallName ?? '',
+          arguments: partial.toolCallArguments,
+        }
+        default: throw new Error(`cannot assemble incomplete block of type "${partial.blockType}"`)
+      }
+    })
+  }
+
+  get usage(): TokenUsage | undefined {
+    return this._usage
+  }
+
+  get finish(): FinishReason {
+    return this._finish ?? { kind: 'stop' }
+  }
+
+  /** The assembled assistant message. */
+  message(): Message {
+    return { role: 'assistant', content: this.blocks() }
+  }
+
+  /** The assembled non-streaming result. */
+  result(): GenerateResult {
+    return { message: this.message(), usage: this._usage, finish: this.finish }
+  }
+}

packages/llm/src/index.ts

@@ -0,0 +1,108 @@
+import { Context, Service } from 'cordis'
+import type { ContentBlock, GenerateOptions, GenerateResult, StreamChunk } from './types.ts'
+import { BlockAssembler } from './assembler.ts'
+
+export * from './types.ts'
+export { BlockAssembler } from './assembler.ts'
+
+declare module 'cordis' {
+  interface Context {
+    llm: LlmService
+  }
+
+  interface Events {
+    /** Waterfall around every streaming model call (retry, caching, routing). */
+    'llm/stream'(this: LlmService, options: GenerateOptions, next: () => AsyncIterable<StreamChunk>): AsyncIterable<StreamChunk>
+    /** Waterfall around every non-streaming model call. */
+    'llm/generate'(this: LlmService, options: GenerateOptions, next: () => Promise<GenerateResult>): Promise<GenerateResult>
+    /** An adapter was registered or unregistered. */
+    'llm/adapter-change'(): void
+  }
+}
+
+export class LlmError extends Error {
+  constructor(message: string, public code: string) {
+    super(message)
+    this.name = 'LlmError'
+  }
+}
+
+/**
+ * Base class for LLM provider adapters.
+ *
+ * An adapter translates between the harness vocabulary (Message/ContentBlock/
+ * StreamChunk) and one provider's wire format. Adapters register themselves
+ * via `ctx.llm.registerAdapter(models, adapter)`.
+ *
+ * TODO: the first real adapter (DeepSeek V4) lands in a later phase; until
+ * then only mock adapters (tests, demo) exist.
+ */
+export abstract class LlmAdapter {
+  /** Stream one model call as raw chunks. The only required method. */
+  abstract stream(options: GenerateOptions): AsyncIterable<StreamChunk>
+}
+
+/**
+ * The abstract `llm` service: an adapter registry plus streaming /
+ * non-streaming call surfaces, both interceptable via waterfall events.
+ */
+export class LlmService extends Service {
+  private adapters = new Map<string, LlmAdapter>()
+
+  constructor(ctx: Context) {
+    super(ctx, 'llm')
+  }
+
+  /** Register an adapter for the given model names. Disposed with the fiber. */
+  registerAdapter(models: string[], adapter: LlmAdapter): () => void {
+    return this.ctx.effect(() => {
+      for (const model of models) this.adapters.set(model, adapter)
+      this.ctx.emit('llm/adapter-change')
+      return () => {
+        for (const model of models) this.adapters.delete(model)
+        this.ctx.emit('llm/adapter-change')
+      }
+    }, 'llm.registerAdapter()')
+  }
+
+  /** Model names with a registered adapter. */
+  models(): string[] {
+    return [...this.adapters.keys()]
+  }
+
+  private adapter(model: string): LlmAdapter {
+    const adapter = this.adapters.get(model)
+    if (!adapter) throw new LlmError(`no adapter registered for model "${model}"`, 'NO_ADAPTER')
+    return adapter
+  }
+
+  /** Stream one model call as raw chunks (token-level deltas). */
+  stream(options: GenerateOptions): AsyncIterable<StreamChunk> {
+    return this.ctx.waterfall(this, 'llm/stream', options, () => {
+      return this.adapter(options.model).stream(options)
+    })
+  }
+
+  /**
+   * Stream one model call as completed content blocks — a convenience view
+   * for consumers that don't care about token-level deltas.
+   */
+  async * streamBlocks(options: GenerateOptions): AsyncIterable<ContentBlock> {
+    const assembler = new BlockAssembler()
+    for await (const chunk of this.stream(options)) {
+      const block = assembler.push(chunk)
+      if (block) yield block
+    }
+  }
+
+  /** One model call, fully assembled (drains the chunk stream). */
+  generate(options: GenerateOptions): Promise<GenerateResult> {
+    return this.ctx.waterfall(this, 'llm/generate', options, async () => {
+      const assembler = new BlockAssembler()
+      for await (const chunk of this.stream(options)) assembler.push(chunk)
+      return assembler.result()
+    })
+  }
+}
+
+export default LlmService

packages/llm/src/types.ts

@@ -0,0 +1,178 @@
+/**
+ * Provider-neutral message and streaming vocabulary.
+ *
+ * This is the canonical language spoken by the agent loop, session logs, and
+ * every plugin. Adapters translate it to provider wire formats (DeepSeek V4
+ * first); nothing outside an adapter should ever see a provider-specific
+ * shape.
+ *
+ * Extensibility: the unions in this file are derived from interfaces
+ * (`ContentBlockMap`, `MessageSourceMap`, `FinishReasonMap`) so that plugins
+ * can extend them via declaration merging:
+ *
+ * ```ts
+ * declare module '@deepseek-ai/dsh-llm' {
+ *   interface ContentBlockMap {
+ *     video: { type: 'video'; url: string }
+ *   }
+ * }
+ * ```
+ */
+
+/** Cache hint attached to a content block (provider-interpreted). */
+export type CacheHint = 'ephemeral'
+
+/** Plain text visible to the end user. */
+export interface TextBlock {
+  type: 'text'
+  text: string
+  cache?: CacheHint
+}
+
+/** Reasoning / thinking content, distinct from visible text. */
+export interface ReasoningBlock {
+  type: 'reasoning'
+  text: string
+}
+
+/** A tool invocation requested by the model. */
+export interface ToolCallBlock {
+  type: 'tool-call'
+  /** Provider-issued call id; correlates with the matching tool result. */
+  id: string
+  name: string
+  /** Raw JSON string as produced by the model. */
+  arguments: string
+}
+
+/** The result of a tool invocation, sent back to the model. */
+export interface ToolResultBlock {
+  type: 'tool-result'
+  toolCallId: string
+  content: ContentBlock[]
+  isError?: boolean
+  cache?: CacheHint
+}
+
+/** An image, by URL or data URL. */
+export interface ImageBlock {
+  type: 'image'
+  url: string
+  mimeType?: string
+  cache?: CacheHint
+}
+
+/**
+ * All known content block shapes, keyed by their `type` tag.
+ * Merge-extensible: plugins add new block types via declaration merging.
+ */
+export interface ContentBlockMap {
+  'text': TextBlock
+  'reasoning': ReasoningBlock
+  'tool-call': ToolCallBlock
+  'tool-result': ToolResultBlock
+  'image': ImageBlock
+}
+
+export type ContentBlockType = keyof ContentBlockMap
+export type ContentBlock = ContentBlockMap[ContentBlockType]
+
+/** A single message in a conversation history. */
+export interface Message {
+  role: 'system' | 'user' | 'assistant'
+  content: ContentBlock[]
+}
+
+/**
+ * Where a message (or injected content) came from.
+ * Merge-extensible sum type — plugins add their own `kind`s.
+ */
+export interface MessageSourceMap {
+  user: { kind: 'user' }
+  plugin: { kind: 'plugin'; plugin: string }
+  agent: { kind: 'agent'; agentId: string }
+}
+
+export type MessageSource = MessageSourceMap[keyof MessageSourceMap]
+
+/**
+ * Why a model response stopped.
+ * Merge-extensible so adapters can surface provider-specific reasons.
+ */
+export interface FinishReasonMap {
+  'stop': { kind: 'stop' }
+  'tool-calls': { kind: 'tool-calls' }
+  'max-tokens': { kind: 'max-tokens' }
+  'aborted': { kind: 'aborted' }
+  'error': { kind: 'error'; message: string; code?: string }
+}
+
+export type FinishReason = FinishReasonMap[keyof FinishReasonMap]
+
+/** Token accounting for one model call (cache fields are optional). */
+export interface TokenUsage {
+  inputTokens: number
+  outputTokens: number
+  cacheReadTokens?: number
+  cacheWriteTokens?: number
+  reasoningTokens?: number
+}
+
+/**
+ * Raw streaming protocol emitted by adapters.
+ *
+ * A streaming response interleaves several typed blocks (text, reasoning,
+ * multiple tool calls); `index` ties each delta to its block, and `block-end`
+ * carries the fully-assembled ContentBlock so consumers don't have to
+ * re-assemble deltas themselves (use {@link BlockAssembler} when they do).
+ *
+ * TODO(review): this protocol needs careful review before the first real
+ * adapter lands (DeepSeek V4 wire format, partial JSON arguments, interleaved
+ * reasoning signatures, …).
+ */
+export type StreamChunk =
+  | { type: 'block-start'; index: number; blockType: ContentBlockType }
+  | { type: 'text-delta'; index: number; text: string }
+  | { type: 'reasoning-delta'; index: number; text: string }
+  | { type: 'tool-call-delta'; index: number; id: string; name?: string; argumentsDelta: string }
+  | { type: 'block-end'; index: number; block: ContentBlock }
+  | { type: 'usage'; usage: TokenUsage }
+  | { type: 'finish'; reason: FinishReason }
+
+/**
+ * JSON-schema description of a tool, as sent to the model.
+ *
+ * Declared here (not in dsh-tools) because it is part of {@link GenerateOptions};
+ * dsh-tools' ToolDefinition and dsh-system-prompt's PromptAssembly both import
+ * it from this package.
+ */
+export interface ToolSchema {
+  name: string
+  description: string
+  /** JSON Schema object for the arguments. */
+  parameters: Record<string, unknown>
+  strict?: boolean
+}
+
+/** A single model request, fully assembled. */
+export interface GenerateOptions {
+  model: string
+  messages: Message[]
+  /** System prompt text (adapters map to the provider's system slot). */
+  system?: string
+  /** Tool schemas (adapters map to the provider's `tools` field). */
+  tools?: ToolSchema[]
+  /** Assistant prefix continuation (prefill). */
+  prefill?: ContentBlock[]
+  temperature?: number
+  maxTokens?: number
+  stop?: string[]
+  signal?: AbortSignal
+}
+
+/** Non-streaming result, assembled from the chunk stream. */
+export interface GenerateResult {
+  message: Message
+  usage?: TokenUsage
+  finish: FinishReason
+}