@leaflow/sdk 0.106.0 → 0.108.0

package/dist/assistant/v1/generated/models/uploadAttachmentParams.d.ts

@@ -0,0 +1,55 @@
+/**
+ * Generated by orval 🍺
+ * Do not edit manually.
+ * Leaflow Assistant API
+ * The AI assistant inside the console: conversations, messages, tool calls and their approval, attachments, and connecting the assistant to instant messaging platforms such as WeChat, Feishu and Telegram.
+ *
+ * ## A conversation
+ *
+ * The path one conversation takes, end to end. It names the operations that path needs, not all of them — the full list is below, and it also covers folders, attachments, memories, skills and the instant messaging side.
+ *
+ * 1. `POST /api/v1/threads` creates one;
+ * 2. `POST /api/v1/threads/{thread}/messages` sends a message and returns a `turnId` at once; the assistant runs in the background;
+ * 3. `GET /api/v1/threads/{thread}` returns the conversation document to render, whose `stream` is where live output arrives;
+ * 4. When the assistant needs the user, the document's `wait` says which kind: approving a batch of tool calls (`POST /api/v1/threads/{thread}/approvals/{batch}`), or answering a few questions (`POST /api/v1/threads/{thread}/questions/{item}`). It continues on its own once that is handled.
+ *
+ * Sending a file takes two steps: `POST /api/v1/attachments` for an id, then put that id in `attachmentIds` when sending the message. Images, text and arbitrary binaries all go through the same two steps; what differs is how the assistant reads them afterwards, which the upload response's `kind` says.
+ *
+ * ## Models
+ *
+ * Canopy owns the model catalogue and capabilities. Model ids in this API, conversation preferences and historical entries are Canopy model ids, with no separate assistant model registration. GET /api/v1/models returns the allowed, usable choices and allowModelSelection. Hide the selector when selection is disabled; the service enforces the same rule.
+ *
+ * preferredModelId is the user's choice. Null follows the platform default, including later changes. Disabled selection preserves but ignores preferences for new turns. A model removed from the allowlist, retired in Canopy or otherwise unusable falls back to the default without clearing the preference.
+ *
+ * Users can switch an existing conversation through PATCH /api/v1/threads/{thread} with preferredModelId. This preserves the transcript, starts no turn and makes no inference request. Null returns to following the default. Changes apply from the next turn; a running or waiting turn and messages queued into it retain their captured choices and fallback policy.
+ *
+ * If the chosen model cannot serve a request, the service may safely fall back to the default without confirmation or an additional switching message. If the default also fails, the turn fails normally; no third model is selected for ordinary answers. Fallback preserves executed tool results and never splices answers from different models. Invalid input, cancellation and a declined tool call do not trigger model fallback. Each entry's model records its actual Canopy author, not the saved preference.
+ *
+ * Users cannot choose a compaction model. Normal compaction reuses the effective model and reasoning settings for context-cache reuse. Only if that model cannot handle the compaction request or produce a summary fitting the target window may the service automatically use another allowed Canopy model with suitable capabilities. This does not change the preference or ordinary answer model, and is permitted even when user model selection is disabled.
+ *
+ * After compaction, the full next request, including instructions, tools and retained messages, must fit the target's usable prompt budget with room left for an answer, respecting its output limit. If still too large, further compaction attempts are bounded; exhaustion fails with CANNOT_FOLD instead of sending an oversized request or silently discarding history or images. The same checks apply when switching models or falling back. Each compaction entry records its actual author.
+ *
+ * An empty allowlist or unavailable default does not prevent reading conversations or saving permitted preference changes. A failed Canopy catalogue read is not evidence that a model was removed: use a valid cached snapshot or return MODEL_CATALOG_UNAVAILABLE, without deleting references or inventing capabilities. Catalogue presence is configuration information, not a live provider health check.
+ *
+ * ## Live output
+ *
+ * ```
+ * GET /api/v1/turns/{turn}/stream?ticket=<ticket>&cursor=<cursor>
+ * ```
+ *
+ * This is a Server-Sent Events stream, which is why it is not in the operation list below and why generated clients have no method for it. Open it with `EventSource`, or any SSE client. It is how live output arrives; polling the document on a timer is not the intended way to follow a conversation.
+ *
+ * - `path` and `ticket` both come from the conversation document's `stream`. `path` is relative, resolved against the address the document was fetched from;
+ * - `cursor` is where to start, from the document's `cursor`. Omit it to start at the beginning of what is still available;
+ * - every event is one incremental edit to that same document, so the same rendering logic applies;
+ * - the connection has a lifetime and closes cleanly when it ends. With `EventSource` the browser reconnects with `Last-Event-ID` and resumes by itself, so nothing needs handling;
+ * - three terminal events: `done` this turn finished (carrying its final status), `paused` the assistant is waiting on the user (fetch the document again and read `wait`), `stalled` execution was interrupted and will not resume — do not reconnect.
+ * OpenAPI spec version: 1.0.0
+ */
+export type UploadAttachmentParams = {
+    /**
+     * The name to show and to give the assistant. Images do not need one; anything else does, because the name is most of what says what the file is. Falls back to a generated name.
+     * @maxLength 255
+     */
+    filename?: string;
+};

package/dist/assistant/v1/generated/models/uploadAttachmentParams.js

@@ -0,0 +1,49 @@
+/**
+ * Generated by orval 🍺
+ * Do not edit manually.
+ * Leaflow Assistant API
+ * The AI assistant inside the console: conversations, messages, tool calls and their approval, attachments, and connecting the assistant to instant messaging platforms such as WeChat, Feishu and Telegram.
+ *
+ * ## A conversation
+ *
+ * The path one conversation takes, end to end. It names the operations that path needs, not all of them — the full list is below, and it also covers folders, attachments, memories, skills and the instant messaging side.
+ *
+ * 1. `POST /api/v1/threads` creates one;
+ * 2. `POST /api/v1/threads/{thread}/messages` sends a message and returns a `turnId` at once; the assistant runs in the background;
+ * 3. `GET /api/v1/threads/{thread}` returns the conversation document to render, whose `stream` is where live output arrives;
+ * 4. When the assistant needs the user, the document's `wait` says which kind: approving a batch of tool calls (`POST /api/v1/threads/{thread}/approvals/{batch}`), or answering a few questions (`POST /api/v1/threads/{thread}/questions/{item}`). It continues on its own once that is handled.
+ *
+ * Sending a file takes two steps: `POST /api/v1/attachments` for an id, then put that id in `attachmentIds` when sending the message. Images, text and arbitrary binaries all go through the same two steps; what differs is how the assistant reads them afterwards, which the upload response's `kind` says.
+ *
+ * ## Models
+ *
+ * Canopy owns the model catalogue and capabilities. Model ids in this API, conversation preferences and historical entries are Canopy model ids, with no separate assistant model registration. GET /api/v1/models returns the allowed, usable choices and allowModelSelection. Hide the selector when selection is disabled; the service enforces the same rule.
+ *
+ * preferredModelId is the user's choice. Null follows the platform default, including later changes. Disabled selection preserves but ignores preferences for new turns. A model removed from the allowlist, retired in Canopy or otherwise unusable falls back to the default without clearing the preference.
+ *
+ * Users can switch an existing conversation through PATCH /api/v1/threads/{thread} with preferredModelId. This preserves the transcript, starts no turn and makes no inference request. Null returns to following the default. Changes apply from the next turn; a running or waiting turn and messages queued into it retain their captured choices and fallback policy.
+ *
+ * If the chosen model cannot serve a request, the service may safely fall back to the default without confirmation or an additional switching message. If the default also fails, the turn fails normally; no third model is selected for ordinary answers. Fallback preserves executed tool results and never splices answers from different models. Invalid input, cancellation and a declined tool call do not trigger model fallback. Each entry's model records its actual Canopy author, not the saved preference.
+ *
+ * Users cannot choose a compaction model. Normal compaction reuses the effective model and reasoning settings for context-cache reuse. Only if that model cannot handle the compaction request or produce a summary fitting the target window may the service automatically use another allowed Canopy model with suitable capabilities. This does not change the preference or ordinary answer model, and is permitted even when user model selection is disabled.
+ *
+ * After compaction, the full next request, including instructions, tools and retained messages, must fit the target's usable prompt budget with room left for an answer, respecting its output limit. If still too large, further compaction attempts are bounded; exhaustion fails with CANNOT_FOLD instead of sending an oversized request or silently discarding history or images. The same checks apply when switching models or falling back. Each compaction entry records its actual author.
+ *
+ * An empty allowlist or unavailable default does not prevent reading conversations or saving permitted preference changes. A failed Canopy catalogue read is not evidence that a model was removed: use a valid cached snapshot or return MODEL_CATALOG_UNAVAILABLE, without deleting references or inventing capabilities. Catalogue presence is configuration information, not a live provider health check.
+ *
+ * ## Live output
+ *
+ * ```
+ * GET /api/v1/turns/{turn}/stream?ticket=<ticket>&cursor=<cursor>
+ * ```
+ *
+ * This is a Server-Sent Events stream, which is why it is not in the operation list below and why generated clients have no method for it. Open it with `EventSource`, or any SSE client. It is how live output arrives; polling the document on a timer is not the intended way to follow a conversation.
+ *
+ * - `path` and `ticket` both come from the conversation document's `stream`. `path` is relative, resolved against the address the document was fetched from;
+ * - `cursor` is where to start, from the document's `cursor`. Omit it to start at the beginning of what is still available;
+ * - every event is one incremental edit to that same document, so the same rendering logic applies;
+ * - the connection has a lifetime and closes cleanly when it ends. With `EventSource` the browser reconnects with `Last-Event-ID` and resumes by itself, so nothing needs handling;
+ * - three terminal events: `done` this turn finished (carrying its final status), `paused` the assistant is waiting on the user (fetch the document again and read `wait`), `stalled` execution was interrupted and will not resume — do not reconnect.
+ * OpenAPI spec version: 1.0.0
+ */
+export {};

package/dist/assistant/v1/generated/models/uploadedResource.d.ts

@@ -0,0 +1,79 @@
+/**
+ * Generated by orval 🍺
+ * Do not edit manually.
+ * Leaflow Assistant API
+ * The AI assistant inside the console: conversations, messages, tool calls and their approval, attachments, and connecting the assistant to instant messaging platforms such as WeChat, Feishu and Telegram.
+ *
+ * ## A conversation
+ *
+ * The path one conversation takes, end to end. It names the operations that path needs, not all of them — the full list is below, and it also covers folders, attachments, memories, skills and the instant messaging side.
+ *
+ * 1. `POST /api/v1/threads` creates one;
+ * 2. `POST /api/v1/threads/{thread}/messages` sends a message and returns a `turnId` at once; the assistant runs in the background;
+ * 3. `GET /api/v1/threads/{thread}` returns the conversation document to render, whose `stream` is where live output arrives;
+ * 4. When the assistant needs the user, the document's `wait` says which kind: approving a batch of tool calls (`POST /api/v1/threads/{thread}/approvals/{batch}`), or answering a few questions (`POST /api/v1/threads/{thread}/questions/{item}`). It continues on its own once that is handled.
+ *
+ * Sending a file takes two steps: `POST /api/v1/attachments` for an id, then put that id in `attachmentIds` when sending the message. Images, text and arbitrary binaries all go through the same two steps; what differs is how the assistant reads them afterwards, which the upload response's `kind` says.
+ *
+ * ## Models
+ *
+ * Canopy owns the model catalogue and capabilities. Model ids in this API, conversation preferences and historical entries are Canopy model ids, with no separate assistant model registration. GET /api/v1/models returns the allowed, usable choices and allowModelSelection. Hide the selector when selection is disabled; the service enforces the same rule.
+ *
+ * preferredModelId is the user's choice. Null follows the platform default, including later changes. Disabled selection preserves but ignores preferences for new turns. A model removed from the allowlist, retired in Canopy or otherwise unusable falls back to the default without clearing the preference.
+ *
+ * Users can switch an existing conversation through PATCH /api/v1/threads/{thread} with preferredModelId. This preserves the transcript, starts no turn and makes no inference request. Null returns to following the default. Changes apply from the next turn; a running or waiting turn and messages queued into it retain their captured choices and fallback policy.
+ *
+ * If the chosen model cannot serve a request, the service may safely fall back to the default without confirmation or an additional switching message. If the default also fails, the turn fails normally; no third model is selected for ordinary answers. Fallback preserves executed tool results and never splices answers from different models. Invalid input, cancellation and a declined tool call do not trigger model fallback. Each entry's model records its actual Canopy author, not the saved preference.
+ *
+ * Users cannot choose a compaction model. Normal compaction reuses the effective model and reasoning settings for context-cache reuse. Only if that model cannot handle the compaction request or produce a summary fitting the target window may the service automatically use another allowed Canopy model with suitable capabilities. This does not change the preference or ordinary answer model, and is permitted even when user model selection is disabled.
+ *
+ * After compaction, the full next request, including instructions, tools and retained messages, must fit the target's usable prompt budget with room left for an answer, respecting its output limit. If still too large, further compaction attempts are bounded; exhaustion fails with CANNOT_FOLD instead of sending an oversized request or silently discarding history or images. The same checks apply when switching models or falling back. Each compaction entry records its actual author.
+ *
+ * An empty allowlist or unavailable default does not prevent reading conversations or saving permitted preference changes. A failed Canopy catalogue read is not evidence that a model was removed: use a valid cached snapshot or return MODEL_CATALOG_UNAVAILABLE, without deleting references or inventing capabilities. Catalogue presence is configuration information, not a live provider health check.
+ *
+ * ## Live output
+ *
+ * ```
+ * GET /api/v1/turns/{turn}/stream?ticket=<ticket>&cursor=<cursor>
+ * ```
+ *
+ * This is a Server-Sent Events stream, which is why it is not in the operation list below and why generated clients have no method for it. Open it with `EventSource`, or any SSE client. It is how live output arrives; polling the document on a timer is not the intended way to follow a conversation.
+ *
+ * - `path` and `ticket` both come from the conversation document's `stream`. `path` is relative, resolved against the address the document was fetched from;
+ * - `cursor` is where to start, from the document's `cursor`. Omit it to start at the beginning of what is still available;
+ * - every event is one incremental edit to that same document, so the same rendering logic applies;
+ * - the connection has a lifetime and closes cleanly when it ends. With `EventSource` the browser reconnects with `Last-Event-ID` and resumes by itself, so nothing needs handling;
+ * - three terminal events: `done` this turn finished (carrying its final status), `paused` the assistant is waiting on the user (fetch the document again and read `wait`), `stalled` execution was interrupted and will not resume — do not reconnect.
+ * OpenAPI spec version: 1.0.0
+ */
+import type { UploadedResourceKind } from "./uploadedResourceKind.js";
+export interface UploadedResource {
+    /** Size of what was stored. For an image that has been resized, this is the resized size, not what was uploaded. */
+    byteSize: number;
+    /**
+     * When this upload gets cleared if no message ever references it. Sending a message with this id makes it permanent and this stops applying — a file that belongs to a conversation is kept as long as the conversation is.
+     *
+     * It is here so the editor can say so before it happens. An attachment chip that quietly stops working a week later reads as a bug, and the person who hits it has no way to tell that what they are seeing is a draft being collected.
+     */
+    draftExpiresAt: string;
+    filename: string;
+    /**
+     * Null unless kind is image.
+     * @nullable
+     */
+    height: number | null;
+    id: string;
+    /**
+     * How the assistant will see this file.
+     *
+     * - `image` — read directly, and only by models that accept image input
+     * - `text` — placed inline in the message when small enough, otherwise read on demand
+     * - `binary` — never read directly; the assistant downloads it onto a cloud instance to work with it
+     */
+    kind: UploadedResourceKind;
+    /**
+     * Null unless kind is image.
+     * @nullable
+     */
+    width: number | null;
+}

package/dist/assistant/v1/generated/models/uploadedResource.js

@@ -0,0 +1 @@
+export {};

package/dist/assistant/v1/generated/models/uploadedResourceKind.d.ts

@@ -0,0 +1,61 @@
+/**
+ * Generated by orval 🍺
+ * Do not edit manually.
+ * Leaflow Assistant API
+ * The AI assistant inside the console: conversations, messages, tool calls and their approval, attachments, and connecting the assistant to instant messaging platforms such as WeChat, Feishu and Telegram.
+ *
+ * ## A conversation
+ *
+ * The path one conversation takes, end to end. It names the operations that path needs, not all of them — the full list is below, and it also covers folders, attachments, memories, skills and the instant messaging side.
+ *
+ * 1. `POST /api/v1/threads` creates one;
+ * 2. `POST /api/v1/threads/{thread}/messages` sends a message and returns a `turnId` at once; the assistant runs in the background;
+ * 3. `GET /api/v1/threads/{thread}` returns the conversation document to render, whose `stream` is where live output arrives;
+ * 4. When the assistant needs the user, the document's `wait` says which kind: approving a batch of tool calls (`POST /api/v1/threads/{thread}/approvals/{batch}`), or answering a few questions (`POST /api/v1/threads/{thread}/questions/{item}`). It continues on its own once that is handled.
+ *
+ * Sending a file takes two steps: `POST /api/v1/attachments` for an id, then put that id in `attachmentIds` when sending the message. Images, text and arbitrary binaries all go through the same two steps; what differs is how the assistant reads them afterwards, which the upload response's `kind` says.
+ *
+ * ## Models
+ *
+ * Canopy owns the model catalogue and capabilities. Model ids in this API, conversation preferences and historical entries are Canopy model ids, with no separate assistant model registration. GET /api/v1/models returns the allowed, usable choices and allowModelSelection. Hide the selector when selection is disabled; the service enforces the same rule.
+ *
+ * preferredModelId is the user's choice. Null follows the platform default, including later changes. Disabled selection preserves but ignores preferences for new turns. A model removed from the allowlist, retired in Canopy or otherwise unusable falls back to the default without clearing the preference.
+ *
+ * Users can switch an existing conversation through PATCH /api/v1/threads/{thread} with preferredModelId. This preserves the transcript, starts no turn and makes no inference request. Null returns to following the default. Changes apply from the next turn; a running or waiting turn and messages queued into it retain their captured choices and fallback policy.
+ *
+ * If the chosen model cannot serve a request, the service may safely fall back to the default without confirmation or an additional switching message. If the default also fails, the turn fails normally; no third model is selected for ordinary answers. Fallback preserves executed tool results and never splices answers from different models. Invalid input, cancellation and a declined tool call do not trigger model fallback. Each entry's model records its actual Canopy author, not the saved preference.
+ *
+ * Users cannot choose a compaction model. Normal compaction reuses the effective model and reasoning settings for context-cache reuse. Only if that model cannot handle the compaction request or produce a summary fitting the target window may the service automatically use another allowed Canopy model with suitable capabilities. This does not change the preference or ordinary answer model, and is permitted even when user model selection is disabled.
+ *
+ * After compaction, the full next request, including instructions, tools and retained messages, must fit the target's usable prompt budget with room left for an answer, respecting its output limit. If still too large, further compaction attempts are bounded; exhaustion fails with CANNOT_FOLD instead of sending an oversized request or silently discarding history or images. The same checks apply when switching models or falling back. Each compaction entry records its actual author.
+ *
+ * An empty allowlist or unavailable default does not prevent reading conversations or saving permitted preference changes. A failed Canopy catalogue read is not evidence that a model was removed: use a valid cached snapshot or return MODEL_CATALOG_UNAVAILABLE, without deleting references or inventing capabilities. Catalogue presence is configuration information, not a live provider health check.
+ *
+ * ## Live output
+ *
+ * ```
+ * GET /api/v1/turns/{turn}/stream?ticket=<ticket>&cursor=<cursor>
+ * ```
+ *
+ * This is a Server-Sent Events stream, which is why it is not in the operation list below and why generated clients have no method for it. Open it with `EventSource`, or any SSE client. It is how live output arrives; polling the document on a timer is not the intended way to follow a conversation.
+ *
+ * - `path` and `ticket` both come from the conversation document's `stream`. `path` is relative, resolved against the address the document was fetched from;
+ * - `cursor` is where to start, from the document's `cursor`. Omit it to start at the beginning of what is still available;
+ * - every event is one incremental edit to that same document, so the same rendering logic applies;
+ * - the connection has a lifetime and closes cleanly when it ends. With `EventSource` the browser reconnects with `Last-Event-ID` and resumes by itself, so nothing needs handling;
+ * - three terminal events: `done` this turn finished (carrying its final status), `paused` the assistant is waiting on the user (fetch the document again and read `wait`), `stalled` execution was interrupted and will not resume — do not reconnect.
+ * OpenAPI spec version: 1.0.0
+ */
+/**
+ * How the assistant will see this file.
+ *
+ * - `image` — read directly, and only by models that accept image input
+ * - `text` — placed inline in the message when small enough, otherwise read on demand
+ * - `binary` — never read directly; the assistant downloads it onto a cloud instance to work with it
+ */
+export type UploadedResourceKind = (typeof UploadedResourceKind)[keyof typeof UploadedResourceKind];
+export declare const UploadedResourceKind: {
+    readonly image: "image";
+    readonly text: "text";
+    readonly binary: "binary";
+};

package/dist/assistant/v1/generated/models/uploadedResourceKind.js

@@ -0,0 +1,53 @@
+/**
+ * Generated by orval 🍺
+ * Do not edit manually.
+ * Leaflow Assistant API
+ * The AI assistant inside the console: conversations, messages, tool calls and their approval, attachments, and connecting the assistant to instant messaging platforms such as WeChat, Feishu and Telegram.
+ *
+ * ## A conversation
+ *
+ * The path one conversation takes, end to end. It names the operations that path needs, not all of them — the full list is below, and it also covers folders, attachments, memories, skills and the instant messaging side.
+ *
+ * 1. `POST /api/v1/threads` creates one;
+ * 2. `POST /api/v1/threads/{thread}/messages` sends a message and returns a `turnId` at once; the assistant runs in the background;
+ * 3. `GET /api/v1/threads/{thread}` returns the conversation document to render, whose `stream` is where live output arrives;
+ * 4. When the assistant needs the user, the document's `wait` says which kind: approving a batch of tool calls (`POST /api/v1/threads/{thread}/approvals/{batch}`), or answering a few questions (`POST /api/v1/threads/{thread}/questions/{item}`). It continues on its own once that is handled.
+ *
+ * Sending a file takes two steps: `POST /api/v1/attachments` for an id, then put that id in `attachmentIds` when sending the message. Images, text and arbitrary binaries all go through the same two steps; what differs is how the assistant reads them afterwards, which the upload response's `kind` says.
+ *
+ * ## Models
+ *
+ * Canopy owns the model catalogue and capabilities. Model ids in this API, conversation preferences and historical entries are Canopy model ids, with no separate assistant model registration. GET /api/v1/models returns the allowed, usable choices and allowModelSelection. Hide the selector when selection is disabled; the service enforces the same rule.
+ *
+ * preferredModelId is the user's choice. Null follows the platform default, including later changes. Disabled selection preserves but ignores preferences for new turns. A model removed from the allowlist, retired in Canopy or otherwise unusable falls back to the default without clearing the preference.
+ *
+ * Users can switch an existing conversation through PATCH /api/v1/threads/{thread} with preferredModelId. This preserves the transcript, starts no turn and makes no inference request. Null returns to following the default. Changes apply from the next turn; a running or waiting turn and messages queued into it retain their captured choices and fallback policy.
+ *
+ * If the chosen model cannot serve a request, the service may safely fall back to the default without confirmation or an additional switching message. If the default also fails, the turn fails normally; no third model is selected for ordinary answers. Fallback preserves executed tool results and never splices answers from different models. Invalid input, cancellation and a declined tool call do not trigger model fallback. Each entry's model records its actual Canopy author, not the saved preference.
+ *
+ * Users cannot choose a compaction model. Normal compaction reuses the effective model and reasoning settings for context-cache reuse. Only if that model cannot handle the compaction request or produce a summary fitting the target window may the service automatically use another allowed Canopy model with suitable capabilities. This does not change the preference or ordinary answer model, and is permitted even when user model selection is disabled.
+ *
+ * After compaction, the full next request, including instructions, tools and retained messages, must fit the target's usable prompt budget with room left for an answer, respecting its output limit. If still too large, further compaction attempts are bounded; exhaustion fails with CANNOT_FOLD instead of sending an oversized request or silently discarding history or images. The same checks apply when switching models or falling back. Each compaction entry records its actual author.
+ *
+ * An empty allowlist or unavailable default does not prevent reading conversations or saving permitted preference changes. A failed Canopy catalogue read is not evidence that a model was removed: use a valid cached snapshot or return MODEL_CATALOG_UNAVAILABLE, without deleting references or inventing capabilities. Catalogue presence is configuration information, not a live provider health check.
+ *
+ * ## Live output
+ *
+ * ```
+ * GET /api/v1/turns/{turn}/stream?ticket=<ticket>&cursor=<cursor>
+ * ```
+ *
+ * This is a Server-Sent Events stream, which is why it is not in the operation list below and why generated clients have no method for it. Open it with `EventSource`, or any SSE client. It is how live output arrives; polling the document on a timer is not the intended way to follow a conversation.
+ *
+ * - `path` and `ticket` both come from the conversation document's `stream`. `path` is relative, resolved against the address the document was fetched from;
+ * - `cursor` is where to start, from the document's `cursor`. Omit it to start at the beginning of what is still available;
+ * - every event is one incremental edit to that same document, so the same rendering logic applies;
+ * - the connection has a lifetime and closes cleanly when it ends. With `EventSource` the browser reconnects with `Last-Event-ID` and resumes by itself, so nothing needs handling;
+ * - three terminal events: `done` this turn finished (carrying its final status), `paused` the assistant is waiting on the user (fetch the document again and read `wait`), `stalled` execution was interrupted and will not resume — do not reconnect.
+ * OpenAPI spec version: 1.0.0
+ */
+export const UploadedResourceKind = {
+    image: "image",
+    text: "text",
+    binary: "binary",
+};

package/dist/assistant/v1/generated/models/verifyCodeRequestBody.d.ts

@@ -0,0 +1,52 @@
+/**
+ * Generated by orval 🍺
+ * Do not edit manually.
+ * Leaflow Assistant API
+ * The AI assistant inside the console: conversations, messages, tool calls and their approval, attachments, and connecting the assistant to instant messaging platforms such as WeChat, Feishu and Telegram.
+ *
+ * ## A conversation
+ *
+ * The path one conversation takes, end to end. It names the operations that path needs, not all of them — the full list is below, and it also covers folders, attachments, memories, skills and the instant messaging side.
+ *
+ * 1. `POST /api/v1/threads` creates one;
+ * 2. `POST /api/v1/threads/{thread}/messages` sends a message and returns a `turnId` at once; the assistant runs in the background;
+ * 3. `GET /api/v1/threads/{thread}` returns the conversation document to render, whose `stream` is where live output arrives;
+ * 4. When the assistant needs the user, the document's `wait` says which kind: approving a batch of tool calls (`POST /api/v1/threads/{thread}/approvals/{batch}`), or answering a few questions (`POST /api/v1/threads/{thread}/questions/{item}`). It continues on its own once that is handled.
+ *
+ * Sending a file takes two steps: `POST /api/v1/attachments` for an id, then put that id in `attachmentIds` when sending the message. Images, text and arbitrary binaries all go through the same two steps; what differs is how the assistant reads them afterwards, which the upload response's `kind` says.
+ *
+ * ## Models
+ *
+ * Canopy owns the model catalogue and capabilities. Model ids in this API, conversation preferences and historical entries are Canopy model ids, with no separate assistant model registration. GET /api/v1/models returns the allowed, usable choices and allowModelSelection. Hide the selector when selection is disabled; the service enforces the same rule.
+ *
+ * preferredModelId is the user's choice. Null follows the platform default, including later changes. Disabled selection preserves but ignores preferences for new turns. A model removed from the allowlist, retired in Canopy or otherwise unusable falls back to the default without clearing the preference.
+ *
+ * Users can switch an existing conversation through PATCH /api/v1/threads/{thread} with preferredModelId. This preserves the transcript, starts no turn and makes no inference request. Null returns to following the default. Changes apply from the next turn; a running or waiting turn and messages queued into it retain their captured choices and fallback policy.
+ *
+ * If the chosen model cannot serve a request, the service may safely fall back to the default without confirmation or an additional switching message. If the default also fails, the turn fails normally; no third model is selected for ordinary answers. Fallback preserves executed tool results and never splices answers from different models. Invalid input, cancellation and a declined tool call do not trigger model fallback. Each entry's model records its actual Canopy author, not the saved preference.
+ *
+ * Users cannot choose a compaction model. Normal compaction reuses the effective model and reasoning settings for context-cache reuse. Only if that model cannot handle the compaction request or produce a summary fitting the target window may the service automatically use another allowed Canopy model with suitable capabilities. This does not change the preference or ordinary answer model, and is permitted even when user model selection is disabled.
+ *
+ * After compaction, the full next request, including instructions, tools and retained messages, must fit the target's usable prompt budget with room left for an answer, respecting its output limit. If still too large, further compaction attempts are bounded; exhaustion fails with CANNOT_FOLD instead of sending an oversized request or silently discarding history or images. The same checks apply when switching models or falling back. Each compaction entry records its actual author.
+ *
+ * An empty allowlist or unavailable default does not prevent reading conversations or saving permitted preference changes. A failed Canopy catalogue read is not evidence that a model was removed: use a valid cached snapshot or return MODEL_CATALOG_UNAVAILABLE, without deleting references or inventing capabilities. Catalogue presence is configuration information, not a live provider health check.
+ *
+ * ## Live output
+ *
+ * ```
+ * GET /api/v1/turns/{turn}/stream?ticket=<ticket>&cursor=<cursor>
+ * ```
+ *
+ * This is a Server-Sent Events stream, which is why it is not in the operation list below and why generated clients have no method for it. Open it with `EventSource`, or any SSE client. It is how live output arrives; polling the document on a timer is not the intended way to follow a conversation.
+ *
+ * - `path` and `ticket` both come from the conversation document's `stream`. `path` is relative, resolved against the address the document was fetched from;
+ * - `cursor` is where to start, from the document's `cursor`. Omit it to start at the beginning of what is still available;
+ * - every event is one incremental edit to that same document, so the same rendering logic applies;
+ * - the connection has a lifetime and closes cleanly when it ends. With `EventSource` the browser reconnects with `Last-Event-ID` and resumes by itself, so nothing needs handling;
+ * - three terminal events: `done` this turn finished (carrying its final status), `paused` the assistant is waiting on the user (fetch the document again and read `wait`), `stalled` execution was interrupted and will not resume — do not reconnect.
+ * OpenAPI spec version: 1.0.0
+ */
+export interface VerifyCodeRequestBody {
+    /** @minLength 1 */
+    code: string;
+}

package/dist/assistant/v1/generated/models/verifyCodeRequestBody.js

@@ -0,0 +1,49 @@
+/**
+ * Generated by orval 🍺
+ * Do not edit manually.
+ * Leaflow Assistant API
+ * The AI assistant inside the console: conversations, messages, tool calls and their approval, attachments, and connecting the assistant to instant messaging platforms such as WeChat, Feishu and Telegram.
+ *
+ * ## A conversation
+ *
+ * The path one conversation takes, end to end. It names the operations that path needs, not all of them — the full list is below, and it also covers folders, attachments, memories, skills and the instant messaging side.
+ *
+ * 1. `POST /api/v1/threads` creates one;
+ * 2. `POST /api/v1/threads/{thread}/messages` sends a message and returns a `turnId` at once; the assistant runs in the background;
+ * 3. `GET /api/v1/threads/{thread}` returns the conversation document to render, whose `stream` is where live output arrives;
+ * 4. When the assistant needs the user, the document's `wait` says which kind: approving a batch of tool calls (`POST /api/v1/threads/{thread}/approvals/{batch}`), or answering a few questions (`POST /api/v1/threads/{thread}/questions/{item}`). It continues on its own once that is handled.
+ *
+ * Sending a file takes two steps: `POST /api/v1/attachments` for an id, then put that id in `attachmentIds` when sending the message. Images, text and arbitrary binaries all go through the same two steps; what differs is how the assistant reads them afterwards, which the upload response's `kind` says.
+ *
+ * ## Models
+ *
+ * Canopy owns the model catalogue and capabilities. Model ids in this API, conversation preferences and historical entries are Canopy model ids, with no separate assistant model registration. GET /api/v1/models returns the allowed, usable choices and allowModelSelection. Hide the selector when selection is disabled; the service enforces the same rule.
+ *
+ * preferredModelId is the user's choice. Null follows the platform default, including later changes. Disabled selection preserves but ignores preferences for new turns. A model removed from the allowlist, retired in Canopy or otherwise unusable falls back to the default without clearing the preference.
+ *
+ * Users can switch an existing conversation through PATCH /api/v1/threads/{thread} with preferredModelId. This preserves the transcript, starts no turn and makes no inference request. Null returns to following the default. Changes apply from the next turn; a running or waiting turn and messages queued into it retain their captured choices and fallback policy.
+ *
+ * If the chosen model cannot serve a request, the service may safely fall back to the default without confirmation or an additional switching message. If the default also fails, the turn fails normally; no third model is selected for ordinary answers. Fallback preserves executed tool results and never splices answers from different models. Invalid input, cancellation and a declined tool call do not trigger model fallback. Each entry's model records its actual Canopy author, not the saved preference.
+ *
+ * Users cannot choose a compaction model. Normal compaction reuses the effective model and reasoning settings for context-cache reuse. Only if that model cannot handle the compaction request or produce a summary fitting the target window may the service automatically use another allowed Canopy model with suitable capabilities. This does not change the preference or ordinary answer model, and is permitted even when user model selection is disabled.
+ *
+ * After compaction, the full next request, including instructions, tools and retained messages, must fit the target's usable prompt budget with room left for an answer, respecting its output limit. If still too large, further compaction attempts are bounded; exhaustion fails with CANNOT_FOLD instead of sending an oversized request or silently discarding history or images. The same checks apply when switching models or falling back. Each compaction entry records its actual author.
+ *
+ * An empty allowlist or unavailable default does not prevent reading conversations or saving permitted preference changes. A failed Canopy catalogue read is not evidence that a model was removed: use a valid cached snapshot or return MODEL_CATALOG_UNAVAILABLE, without deleting references or inventing capabilities. Catalogue presence is configuration information, not a live provider health check.
+ *
+ * ## Live output
+ *
+ * ```
+ * GET /api/v1/turns/{turn}/stream?ticket=<ticket>&cursor=<cursor>
+ * ```
+ *
+ * This is a Server-Sent Events stream, which is why it is not in the operation list below and why generated clients have no method for it. Open it with `EventSource`, or any SSE client. It is how live output arrives; polling the document on a timer is not the intended way to follow a conversation.
+ *
+ * - `path` and `ticket` both come from the conversation document's `stream`. `path` is relative, resolved against the address the document was fetched from;
+ * - `cursor` is where to start, from the document's `cursor`. Omit it to start at the beginning of what is still available;
+ * - every event is one incremental edit to that same document, so the same rendering logic applies;
+ * - the connection has a lifetime and closes cleanly when it ends. With `EventSource` the browser reconnects with `Last-Event-ID` and resumes by itself, so nothing needs handling;
+ * - three terminal events: `done` this turn finished (carrying its final status), `paused` the assistant is waiting on the user (fetch the document again and read `wait`), `stalled` execution was interrupted and will not resume — do not reconnect.
+ * OpenAPI spec version: 1.0.0
+ */
+export {};

package/dist/assistant/v1/generated/models/waitResource.d.ts

@@ -0,0 +1,61 @@
+/**
+ * Generated by orval 🍺
+ * Do not edit manually.
+ * Leaflow Assistant API
+ * The AI assistant inside the console: conversations, messages, tool calls and their approval, attachments, and connecting the assistant to instant messaging platforms such as WeChat, Feishu and Telegram.
+ *
+ * ## A conversation
+ *
+ * The path one conversation takes, end to end. It names the operations that path needs, not all of them — the full list is below, and it also covers folders, attachments, memories, skills and the instant messaging side.
+ *
+ * 1. `POST /api/v1/threads` creates one;
+ * 2. `POST /api/v1/threads/{thread}/messages` sends a message and returns a `turnId` at once; the assistant runs in the background;
+ * 3. `GET /api/v1/threads/{thread}` returns the conversation document to render, whose `stream` is where live output arrives;
+ * 4. When the assistant needs the user, the document's `wait` says which kind: approving a batch of tool calls (`POST /api/v1/threads/{thread}/approvals/{batch}`), or answering a few questions (`POST /api/v1/threads/{thread}/questions/{item}`). It continues on its own once that is handled.
+ *
+ * Sending a file takes two steps: `POST /api/v1/attachments` for an id, then put that id in `attachmentIds` when sending the message. Images, text and arbitrary binaries all go through the same two steps; what differs is how the assistant reads them afterwards, which the upload response's `kind` says.
+ *
+ * ## Models
+ *
+ * Canopy owns the model catalogue and capabilities. Model ids in this API, conversation preferences and historical entries are Canopy model ids, with no separate assistant model registration. GET /api/v1/models returns the allowed, usable choices and allowModelSelection. Hide the selector when selection is disabled; the service enforces the same rule.
+ *
+ * preferredModelId is the user's choice. Null follows the platform default, including later changes. Disabled selection preserves but ignores preferences for new turns. A model removed from the allowlist, retired in Canopy or otherwise unusable falls back to the default without clearing the preference.
+ *
+ * Users can switch an existing conversation through PATCH /api/v1/threads/{thread} with preferredModelId. This preserves the transcript, starts no turn and makes no inference request. Null returns to following the default. Changes apply from the next turn; a running or waiting turn and messages queued into it retain their captured choices and fallback policy.
+ *
+ * If the chosen model cannot serve a request, the service may safely fall back to the default without confirmation or an additional switching message. If the default also fails, the turn fails normally; no third model is selected for ordinary answers. Fallback preserves executed tool results and never splices answers from different models. Invalid input, cancellation and a declined tool call do not trigger model fallback. Each entry's model records its actual Canopy author, not the saved preference.
+ *
+ * Users cannot choose a compaction model. Normal compaction reuses the effective model and reasoning settings for context-cache reuse. Only if that model cannot handle the compaction request or produce a summary fitting the target window may the service automatically use another allowed Canopy model with suitable capabilities. This does not change the preference or ordinary answer model, and is permitted even when user model selection is disabled.
+ *
+ * After compaction, the full next request, including instructions, tools and retained messages, must fit the target's usable prompt budget with room left for an answer, respecting its output limit. If still too large, further compaction attempts are bounded; exhaustion fails with CANNOT_FOLD instead of sending an oversized request or silently discarding history or images. The same checks apply when switching models or falling back. Each compaction entry records its actual author.
+ *
+ * An empty allowlist or unavailable default does not prevent reading conversations or saving permitted preference changes. A failed Canopy catalogue read is not evidence that a model was removed: use a valid cached snapshot or return MODEL_CATALOG_UNAVAILABLE, without deleting references or inventing capabilities. Catalogue presence is configuration information, not a live provider health check.
+ *
+ * ## Live output
+ *
+ * ```
+ * GET /api/v1/turns/{turn}/stream?ticket=<ticket>&cursor=<cursor>
+ * ```
+ *
+ * This is a Server-Sent Events stream, which is why it is not in the operation list below and why generated clients have no method for it. Open it with `EventSource`, or any SSE client. It is how live output arrives; polling the document on a timer is not the intended way to follow a conversation.
+ *
+ * - `path` and `ticket` both come from the conversation document's `stream`. `path` is relative, resolved against the address the document was fetched from;
+ * - `cursor` is where to start, from the document's `cursor`. Omit it to start at the beginning of what is still available;
+ * - every event is one incremental edit to that same document, so the same rendering logic applies;
+ * - the connection has a lifetime and closes cleanly when it ends. With `EventSource` the browser reconnects with `Last-Event-ID` and resumes by itself, so nothing needs handling;
+ * - three terminal events: `done` this turn finished (carrying its final status), `paused` the assistant is waiting on the user (fetch the document again and read `wait`), `stalled` execution was interrupted and will not resume — do not reconnect.
+ * OpenAPI spec version: 1.0.0
+ */
+import type { Question } from "./question.js";
+import type { WaitResourceKind } from "./waitResourceKind.js";
+export interface WaitResource {
+    autoResolutionMs?: number;
+    batchId?: string;
+    itemId?: string;
+    /** @nullable */
+    itemIds?: string[] | null;
+    kind: WaitResourceKind;
+    /** @nullable */
+    questions?: Question[] | null;
+    turnId: string;
+}

package/dist/assistant/v1/generated/models/waitResource.js

@@ -0,0 +1 @@
+export {};

package/dist/assistant/v1/generated/models/waitResourceKind.d.ts

@@ -0,0 +1,53 @@
+/**
+ * Generated by orval 🍺
+ * Do not edit manually.
+ * Leaflow Assistant API
+ * The AI assistant inside the console: conversations, messages, tool calls and their approval, attachments, and connecting the assistant to instant messaging platforms such as WeChat, Feishu and Telegram.
+ *
+ * ## A conversation
+ *
+ * The path one conversation takes, end to end. It names the operations that path needs, not all of them — the full list is below, and it also covers folders, attachments, memories, skills and the instant messaging side.
+ *
+ * 1. `POST /api/v1/threads` creates one;
+ * 2. `POST /api/v1/threads/{thread}/messages` sends a message and returns a `turnId` at once; the assistant runs in the background;
+ * 3. `GET /api/v1/threads/{thread}` returns the conversation document to render, whose `stream` is where live output arrives;
+ * 4. When the assistant needs the user, the document's `wait` says which kind: approving a batch of tool calls (`POST /api/v1/threads/{thread}/approvals/{batch}`), or answering a few questions (`POST /api/v1/threads/{thread}/questions/{item}`). It continues on its own once that is handled.
+ *
+ * Sending a file takes two steps: `POST /api/v1/attachments` for an id, then put that id in `attachmentIds` when sending the message. Images, text and arbitrary binaries all go through the same two steps; what differs is how the assistant reads them afterwards, which the upload response's `kind` says.
+ *
+ * ## Models
+ *
+ * Canopy owns the model catalogue and capabilities. Model ids in this API, conversation preferences and historical entries are Canopy model ids, with no separate assistant model registration. GET /api/v1/models returns the allowed, usable choices and allowModelSelection. Hide the selector when selection is disabled; the service enforces the same rule.
+ *
+ * preferredModelId is the user's choice. Null follows the platform default, including later changes. Disabled selection preserves but ignores preferences for new turns. A model removed from the allowlist, retired in Canopy or otherwise unusable falls back to the default without clearing the preference.
+ *
+ * Users can switch an existing conversation through PATCH /api/v1/threads/{thread} with preferredModelId. This preserves the transcript, starts no turn and makes no inference request. Null returns to following the default. Changes apply from the next turn; a running or waiting turn and messages queued into it retain their captured choices and fallback policy.
+ *
+ * If the chosen model cannot serve a request, the service may safely fall back to the default without confirmation or an additional switching message. If the default also fails, the turn fails normally; no third model is selected for ordinary answers. Fallback preserves executed tool results and never splices answers from different models. Invalid input, cancellation and a declined tool call do not trigger model fallback. Each entry's model records its actual Canopy author, not the saved preference.
+ *
+ * Users cannot choose a compaction model. Normal compaction reuses the effective model and reasoning settings for context-cache reuse. Only if that model cannot handle the compaction request or produce a summary fitting the target window may the service automatically use another allowed Canopy model with suitable capabilities. This does not change the preference or ordinary answer model, and is permitted even when user model selection is disabled.
+ *
+ * After compaction, the full next request, including instructions, tools and retained messages, must fit the target's usable prompt budget with room left for an answer, respecting its output limit. If still too large, further compaction attempts are bounded; exhaustion fails with CANNOT_FOLD instead of sending an oversized request or silently discarding history or images. The same checks apply when switching models or falling back. Each compaction entry records its actual author.
+ *
+ * An empty allowlist or unavailable default does not prevent reading conversations or saving permitted preference changes. A failed Canopy catalogue read is not evidence that a model was removed: use a valid cached snapshot or return MODEL_CATALOG_UNAVAILABLE, without deleting references or inventing capabilities. Catalogue presence is configuration information, not a live provider health check.
+ *
+ * ## Live output
+ *
+ * ```
+ * GET /api/v1/turns/{turn}/stream?ticket=<ticket>&cursor=<cursor>
+ * ```
+ *
+ * This is a Server-Sent Events stream, which is why it is not in the operation list below and why generated clients have no method for it. Open it with `EventSource`, or any SSE client. It is how live output arrives; polling the document on a timer is not the intended way to follow a conversation.
+ *
+ * - `path` and `ticket` both come from the conversation document's `stream`. `path` is relative, resolved against the address the document was fetched from;
+ * - `cursor` is where to start, from the document's `cursor`. Omit it to start at the beginning of what is still available;
+ * - every event is one incremental edit to that same document, so the same rendering logic applies;
+ * - the connection has a lifetime and closes cleanly when it ends. With `EventSource` the browser reconnects with `Last-Event-ID` and resumes by itself, so nothing needs handling;
+ * - three terminal events: `done` this turn finished (carrying its final status), `paused` the assistant is waiting on the user (fetch the document again and read `wait`), `stalled` execution was interrupted and will not resume — do not reconnect.
+ * OpenAPI spec version: 1.0.0
+ */
+export type WaitResourceKind = (typeof WaitResourceKind)[keyof typeof WaitResourceKind];
+export declare const WaitResourceKind: {
+    readonly tool_approval: "tool_approval";
+    readonly user_input: "user_input";
+};

package/dist/assistant/v1/generated/models/waitResourceKind.js

@@ -0,0 +1,52 @@
+/**
+ * Generated by orval 🍺
+ * Do not edit manually.
+ * Leaflow Assistant API
+ * The AI assistant inside the console: conversations, messages, tool calls and their approval, attachments, and connecting the assistant to instant messaging platforms such as WeChat, Feishu and Telegram.
+ *
+ * ## A conversation
+ *
+ * The path one conversation takes, end to end. It names the operations that path needs, not all of them — the full list is below, and it also covers folders, attachments, memories, skills and the instant messaging side.
+ *
+ * 1. `POST /api/v1/threads` creates one;
+ * 2. `POST /api/v1/threads/{thread}/messages` sends a message and returns a `turnId` at once; the assistant runs in the background;
+ * 3. `GET /api/v1/threads/{thread}` returns the conversation document to render, whose `stream` is where live output arrives;
+ * 4. When the assistant needs the user, the document's `wait` says which kind: approving a batch of tool calls (`POST /api/v1/threads/{thread}/approvals/{batch}`), or answering a few questions (`POST /api/v1/threads/{thread}/questions/{item}`). It continues on its own once that is handled.
+ *
+ * Sending a file takes two steps: `POST /api/v1/attachments` for an id, then put that id in `attachmentIds` when sending the message. Images, text and arbitrary binaries all go through the same two steps; what differs is how the assistant reads them afterwards, which the upload response's `kind` says.
+ *
+ * ## Models
+ *
+ * Canopy owns the model catalogue and capabilities. Model ids in this API, conversation preferences and historical entries are Canopy model ids, with no separate assistant model registration. GET /api/v1/models returns the allowed, usable choices and allowModelSelection. Hide the selector when selection is disabled; the service enforces the same rule.
+ *
+ * preferredModelId is the user's choice. Null follows the platform default, including later changes. Disabled selection preserves but ignores preferences for new turns. A model removed from the allowlist, retired in Canopy or otherwise unusable falls back to the default without clearing the preference.
+ *
+ * Users can switch an existing conversation through PATCH /api/v1/threads/{thread} with preferredModelId. This preserves the transcript, starts no turn and makes no inference request. Null returns to following the default. Changes apply from the next turn; a running or waiting turn and messages queued into it retain their captured choices and fallback policy.
+ *
+ * If the chosen model cannot serve a request, the service may safely fall back to the default without confirmation or an additional switching message. If the default also fails, the turn fails normally; no third model is selected for ordinary answers. Fallback preserves executed tool results and never splices answers from different models. Invalid input, cancellation and a declined tool call do not trigger model fallback. Each entry's model records its actual Canopy author, not the saved preference.
+ *
+ * Users cannot choose a compaction model. Normal compaction reuses the effective model and reasoning settings for context-cache reuse. Only if that model cannot handle the compaction request or produce a summary fitting the target window may the service automatically use another allowed Canopy model with suitable capabilities. This does not change the preference or ordinary answer model, and is permitted even when user model selection is disabled.
+ *
+ * After compaction, the full next request, including instructions, tools and retained messages, must fit the target's usable prompt budget with room left for an answer, respecting its output limit. If still too large, further compaction attempts are bounded; exhaustion fails with CANNOT_FOLD instead of sending an oversized request or silently discarding history or images. The same checks apply when switching models or falling back. Each compaction entry records its actual author.
+ *
+ * An empty allowlist or unavailable default does not prevent reading conversations or saving permitted preference changes. A failed Canopy catalogue read is not evidence that a model was removed: use a valid cached snapshot or return MODEL_CATALOG_UNAVAILABLE, without deleting references or inventing capabilities. Catalogue presence is configuration information, not a live provider health check.
+ *
+ * ## Live output
+ *
+ * ```
+ * GET /api/v1/turns/{turn}/stream?ticket=<ticket>&cursor=<cursor>
+ * ```
+ *
+ * This is a Server-Sent Events stream, which is why it is not in the operation list below and why generated clients have no method for it. Open it with `EventSource`, or any SSE client. It is how live output arrives; polling the document on a timer is not the intended way to follow a conversation.
+ *
+ * - `path` and `ticket` both come from the conversation document's `stream`. `path` is relative, resolved against the address the document was fetched from;
+ * - `cursor` is where to start, from the document's `cursor`. Omit it to start at the beginning of what is still available;
+ * - every event is one incremental edit to that same document, so the same rendering logic applies;
+ * - the connection has a lifetime and closes cleanly when it ends. With `EventSource` the browser reconnects with `Last-Event-ID` and resumes by itself, so nothing needs handling;
+ * - three terminal events: `done` this turn finished (carrying its final status), `paused` the assistant is waiting on the user (fetch the document again and read `wait`), `stalled` execution was interrupted and will not resume — do not reconnect.
+ * OpenAPI spec version: 1.0.0
+ */
+export const WaitResourceKind = {
+    tool_approval: "tool_approval",
+    user_input: "user_input",
+};