path: root/.dispatch/wire.reference.md

# `@dispatch/wire` — in-repo reference (read THIS, not node_modules)

> MIRRORS the backend's `@dispatch/wire` package source so headless FE agents can read the wire
> types WITHOUT following the `file:` dep symlink out of this repo (which hangs on a permission
> prompt). Your CODE still imports `@dispatch/wire` normally — this file is for READING only.
>
> **Orchestrator:** SNAPSHOT of `[email protected]` (the metrics types below + the `user-message` turn event
> shipped + version-bumped). Regenerate whenever `@dispatch/wire` changes.
>
> **2026-06-12 delta (CR-3 user-message handoff — package bumped `0.5.0` → `0.6.0`, ADDITIVE):** adds a
> new `AgentEvent` union member `TurnInputEvent` (`{ type: "user-message"; conversationId; turnId; text }`)
> that surfaces the turn's USER prompt INTO the outward event stream. Emitted ONCE as the FIRST event of
> every turn (before `turn-start`), so it is buffered + replayed to every subscriber — live AND late-join
> — and rides `chat.delta`/NDJSON like any other event. Fixes CR-3 (a pure watcher couldn't see the prompt
> until seal). The sender still echoes its own prompt optimistically, so consumers DE-DUP against that
> (by text); a pure watcher renders it directly. Persistence/metrics unchanged. See `TurnInputEvent` below.
>
> **2026-06-12 delta (context-size handoff — package bumped `0.4.0` → `0.5.0`):** adds an OPTIONAL
> `contextSize?: number` to BOTH `TurnDoneEvent` (live `done`) and `TurnMetrics` (persisted) — the
> turn's FINAL step `inputTokens + outputTokens` (current context occupancy), NOT the aggregate
> `usage` (which overcounts multi-step turns). The two carriers are equal for the same turn. Current
> value = the LATEST turn's `contextSize`; `undefined` ⇒ render "unknown", never `0`. See the field
> doc-comments on `TurnMetrics`/`TurnDoneEvent` below.
>
> **0.3.0 changes (token + timing metrics):**
> - **Live per-step/per-turn telemetry on the event stream** (transient — NOT persisted):
>   `TurnUsageEvent` gained an OPTIONAL `stepId?` (attribute tokens per step). A NEW
>   `TurnStepCompleteEvent` (`type: "step-complete"`, REQUIRED `stepId`) carries the per-step
>   generation timing `ttftMs?` / `decodeMs?` / `genTotalMs?` (all optional — present only when the
>   runtime had a clock; `ttftMs`/`decodeMs` additionally require a first content token). `TurnDoneEvent`
>   gained an OPTIONAL `durationMs?` (total turn wall-clock) + OPTIONAL `usage?` (aggregate across
>   steps). `TurnToolResultEvent` gained an OPTIONAL `durationMs?` (tool execution time).
> - **Durable, replayable metrics** (persisted, keyed per turn): NEW `StepMetrics` + `TurnMetrics`
>   — the persisted counterparts of the live `usage` + `step-complete` + `done` packets. Served by
>   `GET /conversations/:id/metrics` (see `transport-contract.reference.md`). Build the SAME
>   `TurnMetrics` shape from the live events for the in-flight turn; the durable endpoint supplies it
>   for sealed turns. TPS is derived (`usage.outputTokens / (genTotalMs / 1000)`), not on the wire.
> - **0.2.0 (still current — step grouping):** `ToolCallChunk`/`ToolResultChunk` carry an OPTIONAL
>   `stepId?: StepId`; `TurnToolCallEvent`/`TurnToolResultEvent` carry a REQUIRED `stepId: StepId`.
>   Group batched/parallel tool calls by `stepId` equality. Live: read `event.stepId`. Replay: read
>   `storedChunk.chunk.stepId` (NOT the envelope; tolerate absence). `StoredChunk` envelope is
>   UNCHANGED (`{ seq, role, chunk }` — carries NO `turnId`).

```ts
/**
 * @dispatch/wire — pure wire types shared by the kernel, the transport
 * contract, and out-of-repo clients (the web frontend).
 *
 * Types ONLY: zero runtime, zero `@dispatch/*` dependencies, so a client can
 * depend on the wire without pulling the kernel runtime.
 */

// ─── Conversation model ─────────────────────────────────────────────────────

/** Who produced a message. */
export type Role = "system" | "user" | "assistant" | "tool";

/** Opaque identifier for a turn (one user→assistant cycle). */
export type TurnId = string & { readonly __brand: "TurnId" };

/**
 * Opaque identifier for a step (one LLM round-trip within a turn). It is the
 * authoritative grouping key for the tool calls a model batches together in a
 * single step (parallel/batched calls): every `tool-call`/`tool-result` event
 * and every persisted tool chunk (`ToolCallChunk`/`ToolResultChunk`) from the
 * same step carries the SAME `stepId`, so a client groups a batch purely by
 * equality — identically on the live stream and in replayed history. Per-turn
 * unique and gap-free in step order; treat it as opaque (do not parse it). The
 * runtime derives it deterministically from the turn id + 0-based step index.
 */
export type StepId = string & { readonly __brand: "StepId" };

/**
 * A chunk is one ordered piece of a message — the atomic unit of the
 * append-only conversation log. Discriminated by `type`.
 */
export type Chunk =
	| TextChunk
	| ThinkingChunk
	| ToolCallChunk
	| ToolResultChunk
	| ErrorChunk
	| SystemChunk;

/** A piece of plain text content from the assistant or user. */
export interface TextChunk {
	readonly type: "text";
	readonly text: string;
}

/** A piece of model reasoning / thinking content (e.g. extended thinking). */
export interface ThinkingChunk {
	readonly type: "thinking";
	readonly text: string;
}

/**
 * A model's request to run a tool. The kernel routes by `name`; the tool
 * implementation never sees this directly — it receives parsed `input` via
 * `ToolContract.execute`.
 */
export interface ToolCallChunk {
	readonly type: "tool-call";
	readonly toolCallId: string;
	readonly toolName: string;
	readonly input: unknown;
	/**
	 * The step that produced this call — generation provenance stamped by the
	 * runtime when the model emits the call (NOT storage metadata like `seq`,
	 * which is why it lives on the chunk and travels with it through persistence
	 * and replay). Tool calls a model batches together in one step share the same
	 * `stepId`: the grouping key for rendering a parallel batch as one unit, and
	 * equal to the `stepId` on the matching `tool-call` AgentEvent. Optional:
	 * absent on chunks reconstructed outside a turn and on rows persisted before
	 * this field existed, so a consumer must tolerate its absence (render
	 * ungrouped).
	 */
	readonly stepId?: StepId;
}

/**
 * The result of a tool execution, attributed to the originating tool-call id.
 * The kernel guarantees every tool-call chunk gets exactly one result chunk
 * (synthesized if interrupted — see reconcile).
 */
export interface ToolResultChunk {
	readonly type: "tool-result";
	readonly toolCallId: string;
	readonly toolName: string;
	readonly content: string;
	readonly isError: boolean;
	/**
	 * The step that produced the originating call — equal to the `stepId` on the
	 * matching `tool-call` chunk (same `toolCallId`) and on the `tool-result`
	 * AgentEvent, so a consumer groups a step's calls with their results.
	 * Generation provenance, not storage metadata (see `ToolCallChunk.stepId`).
	 * Optional for the same reasons; `reconcile` copies it from the originating
	 * call onto a synthesized (interrupted) result.
	 */
	readonly stepId?: StepId;
}

/** An error that occurred during generation or tool dispatch. */
export interface ErrorChunk {
	readonly type: "error";
	readonly message: string;
	readonly code?: string;
}

/**
 * A system-injected message (e.g. system prompt, context assembly output).
 * Kept distinct from text so the log records provenance.
 */
export interface SystemChunk {
	readonly type: "system";
	readonly text: string;
}

/**
 * A chat message: a role plus an ordered sequence of chunks. Messages are the
 * unit passed to and from the provider; chunks are the unit persisted and
 * rendered.
 */
export interface ChatMessage {
	readonly role: Role;
	readonly chunks: readonly Chunk[];
}

/**
 * A persisted chunk plus its sync metadata. The append-only conversation log
 * stamps every chunk with a monotonic, gap-free, per-conversation `seq` (the
 * sync cursor, assigned in append order) and records the `role` of the message
 * it belongs to. This makes a flat seq-ordered stream both incrementally
 * syncable ("give me chunks after seq N") and regroupable into messages by the
 * client. `chunk` is the content unit — `Chunk` carries no storage/sync cursor
 * (`seq` lives here on the envelope, not on the chunk, since it is assigned by
 * the store and the provider has no use for it). A chunk MAY still carry
 * generation provenance assigned at production time (e.g. a tool chunk's
 * `stepId`), which is intrinsic to the content and so travels with it.
 */
export interface StoredChunk {
	readonly seq: number;
	readonly role: Role;
	readonly chunk: Chunk;
}

// ─── Usage ──────────────────────────────────────────────────────────────────

/**
 * Token usage counters for a single step. All fields are counts of tokens.
 * Cache fields are optional because not all providers expose cache metrics.
 */
export interface Usage {
	readonly inputTokens: number;
	readonly outputTokens: number;
	readonly cacheReadTokens?: number;
	readonly cacheWriteTokens?: number;
}

// ─── Persisted metrics ───────────────────────────────────────────────────────

/**
 * Durable per-step metrics for a completed step — the persisted, replayable
 * counterpart of the live `usage` + `step-complete` events. Combines the step's
 * token usage with its generation timing so a client reopening a past
 * conversation renders the same per-step token/latency breakdown it would have
 * seen live. Built from the turn's events, stored by `conversation-store`, and
 * served by `GET /conversations/:id/metrics`.
 */
export interface StepMetrics {
	readonly stepId: StepId;
	/** The step's token usage (all four counters; cache fields optional per `Usage`). */
	readonly usage: Usage;
	/** Time to first token (stream start → first text/reasoning delta). Optional — see `TurnStepCompleteEvent.ttftMs`. */
	readonly ttftMs?: number;
	/** Decode time (first token → stream end). Optional — see `TurnStepCompleteEvent.decodeMs`. */
	readonly decodeMs?: number;
	/** Total generation time for the step (stream start → stream end). Optional: present only when a clock was available. */
	readonly genTotalMs?: number;
}

/**
 * Durable per-turn metrics for a completed (sealed) turn — the persisted,
 * replayable counterpart of the live `done` event's aggregate `usage` +
 * `durationMs`, plus the per-step breakdown. `usage` is the aggregate across all
 * steps; `steps` carries each step's `StepMetrics` in step order. Stored by
 * `conversation-store` keyed by `turnId` and served by
 * `GET /conversations/:id/metrics`. (`turnId` is the plain wire string carried
 * on every `AgentEvent`, the join key to the live stream.)
 */
export interface TurnMetrics {
	readonly turnId: string;
	/** Aggregate token usage across all steps in the turn. */
	readonly usage: Usage;
	/** Total wall-clock duration of the turn (turn start → turn end). Optional: present only when a clock was available. */
	readonly durationMs?: number;
	/** Per-step metrics in step order. */
	readonly steps: readonly StepMetrics[];
	/**
	 * **Context size** — tokens the conversation occupies as of this turn: the
	 * turn's FINAL step `inputTokens + outputTokens` (the last entry of `steps`),
	 * NOT the aggregate `usage` (which sums per-step prompts and overcounts a
	 * multi-step turn). The persisted, replayable counterpart of
	 * `TurnDoneEvent.contextSize` and equal to it for the same turn. A client
	 * reopening a past conversation reads the LAST turn's `contextSize` as the
	 * current context usage. Optional: absent when no per-step usage was available.
	 */
	readonly contextSize?: number;
}

// ─── Outward events ─────────────────────────────────────────────────────────

/**
 * The union of all events the runtime emits outward during a turn.
 * Consumers (transport, persistence, notifications) pattern-match on `type`.
 */
export type AgentEvent =
	| StatusEvent
	| TurnStartEvent
	| TurnInputEvent
	| TurnTextDeltaEvent
	| TurnReasoningDeltaEvent
	| TurnToolCallEvent
	| TurnToolResultEvent
	| TurnToolOutputEvent
	| TurnUsageEvent
	| TurnStepCompleteEvent
	| TurnErrorEvent
	| TurnDoneEvent
	| TurnSealedEvent;

/** Status change for a conversation (e.g. idle → running). */
export interface StatusEvent {
	readonly type: "status";
	readonly conversationId: string;
	readonly status: string;
}

/** A turn has begun. */
export interface TurnStartEvent {
	readonly type: "turn-start";
	readonly conversationId: string;
	readonly turnId: string;
}

/**
 * The user prompt that opened this turn, surfaced INTO the turn's outward event
 * stream so a WATCHER (subscribed but not the sender) can render the prompt
 * mid-turn — the user message is otherwise persisted only at seal. Emitted ONCE
 * as the FIRST event of the turn (before `turn-start`); buffered + replayed to
 * every subscriber (live + late-join). The sender echoes its own prompt
 * optimistically, so DE-DUP against that (by text); a pure watcher renders it
 * directly. Carries the raw `text` passed to the provider. (Turn-scoped: it
 * carries `turnId`, so a multi-turn transcript attributes each prompt to its turn.)
 */
export interface TurnInputEvent {
	readonly type: "user-message";
	readonly conversationId: string;
	readonly turnId: string;
	readonly text: string;
}

/** Incremental text content from the model during a turn. */
export interface TurnTextDeltaEvent {
	readonly type: "text-delta";
	readonly conversationId: string;
	readonly turnId: string;
	readonly delta: string;
}

/** Incremental reasoning / thinking content during a turn. */
export interface TurnReasoningDeltaEvent {
	readonly type: "reasoning-delta";
	readonly conversationId: string;
	readonly turnId: string;
	readonly delta: string;
}

/** The model has requested a tool to be run. */
export interface TurnToolCallEvent {
	readonly type: "tool-call";
	readonly conversationId: string;
	readonly turnId: string;
	/**
	 * The step that produced this call. Tool calls a model batches together in
	 * one step share the same `stepId` — the grouping key for rendering a
	 * parallel batch as one unit. Matches the `stepId` on the matching
	 * `tool-result` event and on the persisted tool chunk
	 * (`StoredChunk.chunk.stepId`).
	 */
	readonly stepId: StepId;
	readonly toolCallId: string;
	readonly toolName: string;
	readonly input: unknown;
}

/** A tool has completed execution. */
export interface TurnToolResultEvent {
	readonly type: "tool-result";
	readonly conversationId: string;
	readonly turnId: string;
	/**
	 * The step that produced the originating call. Equal to the `stepId` on the
	 * matching `tool-call` event (same `toolCallId`) and on the persisted tool
	 * chunk (`StoredChunk.chunk.stepId`), so a client groups a step's calls with
	 * their results.
	 */
	readonly stepId: StepId;
	readonly toolCallId: string;
	readonly toolName: string;
	readonly content: string;
	readonly isError: boolean;
	/**
	 * How long the tool took to execute (dispatch → result), in milliseconds —
	 * the backend's authoritative execution time, distinct from any client-side
	 * wall-clock. Optional: present only when the runtime was given a clock.
	 */
	readonly durationMs?: number;
}

/** Streaming output from a tool execution (e.g. shell stdout/stderr). */
export interface TurnToolOutputEvent {
	readonly type: "tool-output";
	readonly conversationId: string;
	readonly turnId: string;
	readonly toolCallId: string;
	readonly data: string;
	readonly stream: "stdout" | "stderr";
}

/** Token usage for the current step or turn. */
export interface TurnUsageEvent {
	readonly type: "usage";
	readonly conversationId: string;
	readonly turnId: string;
	/**
	 * The step this usage report belongs to, so a consumer can attribute tokens
	 * per step (and join with the matching `step-complete` timing by `stepId`).
	 * Optional: absent when the runtime had no step context, and on usage emitted
	 * before this field existed.
	 */
	readonly stepId?: StepId;
	readonly usage: Usage;
}

/**
 * A step (one LLM round-trip) has completed — the authoritative per-step metrics
 * packet, emitted once at the step's end (after the generation stream finishes),
 * so its timing is final (unlike `usage`, which may arrive mid-stream). Carries
 * the step's generation timing; join to the step's tokens via `stepId` on the
 * `usage` event. All timing fields are optional: present only when the runtime
 * was given a clock, and `ttftMs`/`decodeMs` additionally require that a first
 * content token (text or reasoning) was observed this step.
 */
export interface TurnStepCompleteEvent {
	readonly type: "step-complete";
	readonly conversationId: string;
	readonly turnId: string;
	readonly stepId: StepId;
	/** Time to first token: stream start → first text/reasoning delta. */
	readonly ttftMs?: number;
	/** Decode time: first token → stream end (generation total − TTFT). */
	readonly decodeMs?: number;
	/**
	 * Total generation time for the step: stream start → stream end. Present
	 * whenever a clock was available, even if no first token was seen (in which
	 * case `ttftMs`/`decodeMs` are absent). When a first token was seen,
	 * `genTotalMs === ttftMs + decodeMs`.
	 */
	readonly genTotalMs?: number;
}

/** An error occurred during the turn. */
export interface TurnErrorEvent {
	readonly type: "error";
	readonly conversationId: string;
	readonly turnId: string;
	readonly message: string;
	readonly code?: string;
}

/** The turn has completed (model finished generating). */
export interface TurnDoneEvent {
	readonly type: "done";
	readonly conversationId: string;
	readonly turnId: string;
	readonly reason: string;
	/**
	 * Total wall-clock duration of the turn (turn start → turn end), in
	 * milliseconds. Optional: present only when the runtime was given a clock.
	 */
	readonly durationMs?: number;
	/**
	 * Aggregate token usage across all steps in the turn — a convenience total so
	 * a consumer need not sum the per-step `usage` events. Optional (absent if the
	 * provider reported no usage).
	 */
	readonly usage?: Usage;
	/**
	 * **Context size** — tokens the conversation occupies right now: the turn's
	 * FINAL step `inputTokens + outputTokens` (the prompt sent into the last LLM
	 * round-trip plus that round-trip's output). This is the "tokens in context"
	 * figure a client renders as the chat's current context usage, and a client
	 * treats the LATEST turn's value as the live total.
	 *
	 * Deliberately NOT the aggregate `usage` above: `usage` SUMS each step's
	 * `inputTokens`, which overcounts a multi-step / tool-calling turn because every
	 * step re-prefills the growing prompt — the final step's input already includes
	 * all prior context, so its input+output is the true occupancy. Optional: absent
	 * when no per-step usage was observed this turn (mirrors `usage`). A later field
	 * will carry the model's max context-window LIMIT; this is only the current size.
	 */
	readonly contextSize?: number;
}

/**
 * The turn has been sealed — all chunks persisted, history is final.
 * This is the hook point for post-turn extensions (compaction, cache-warm).
 */
export interface TurnSealedEvent {
	readonly type: "turn-sealed";
	readonly conversationId: string;
	readonly turnId: string;
}
```