@coasys/ad4m 0.13.1-staging.12 → 0.13.1-staging.13

This diff represents the content of publicly available package versions that have been released to one of the supported registries. The information contained in this diff is provided for informational purposes only and reflects changes between package versions as they appear in their respective public registries.
@@ -14,9 +14,10 @@ export type AISessionMessage = {
14
14
  /**
15
15
  * The transcript entry: `{"text": ..}` for user and assistant turns;
16
16
  * `{"name", "arguments", "target"?, "writes", "result"? | "error"?}` for a
17
- * tool step. `writes` is the executor's own read of the call's side effect
18
- * (`agent_session::tool_writes`), so a client marking the steps that
19
- * changed something does not have to guess from the name.
17
+ * tool step. `writes` is the tool's declared intent
18
+ * (`agent_session::tool_writes`), recorded before the call runs. It does
19
+ * not say the call changed anything: a refused or failed write still has
20
+ * `writes: true`.
20
21
  */
21
22
  content: unknown;
22
23
  /**
@@ -14,8 +14,9 @@ export type AiEndpoint = {
14
14
  maxNumCtx?: number | null;
15
15
  /**
16
16
  * How long a request to this endpoint may take, queueing included. Unset: 300 s for
17
- * chat `/v1` forwards, 600 s for chat the executor answers (Anthropic, Ollama, tools as
18
- * text; not bounded here when streamed), 60 s for embeddings, 120 s for speech.
17
+ * chat `/v1` forwards, 600 s for chat the executor answers, 60 s for embeddings, 120 s
18
+ * for speech. An OpenAI-protocol chat client uses this value for a stream too. A value
19
+ * above 600 s is kept. A stream has no second timer in `within_endpoint_limits`.
19
20
  */
20
21
  requestTimeoutSecs?: number | null;
21
22
  /**
package/package.json CHANGED
@@ -53,8 +53,8 @@
53
53
  "ws": "8.13.0"
54
54
  },
55
55
  "patchedDependencies": {},
56
- "version": "0.13.1-staging.12",
57
- "gitHead": "1cdb231b3221940744943222f07f7f75a7dc0711",
56
+ "version": "0.13.1-staging.13",
57
+ "gitHead": "3f800fc56a6b19a44a53b60b38b1ba2a7b809cf8",
58
58
  "scripts": {
59
59
  "build": "patch-package && tsc && pnpm run bundle",
60
60
  "bundle": "rollup -c rollup.config.js",
@@ -13,9 +13,10 @@ seq: number, role: AISessionMessageRole,
13
13
  /**
14
14
  * The transcript entry: `{"text": ..}` for user and assistant turns;
15
15
  * `{"name", "arguments", "target"?, "writes", "result"? | "error"?}` for a
16
- * tool step. `writes` is the executor's own read of the call's side effect
17
- * (`agent_session::tool_writes`), so a client marking the steps that
18
- * changed something does not have to guess from the name.
16
+ * tool step. `writes` is the tool's declared intent
17
+ * (`agent_session::tool_writes`), recorded before the call runs. It does
18
+ * not say the call changed anything: a refused or failed write still has
19
+ * `writes: true`.
19
20
  */
20
21
  content: unknown,
21
22
  /**
@@ -9,8 +9,9 @@ import type { EndpointKind } from "./EndpointKind";
9
9
  export type AiEndpoint = { id: string, name: string, kind: EndpointKind, baseUrl: string, hasApiKey: boolean, capabilities: Array<AiCapability>, maxNumCtx?: number | null,
10
10
  /**
11
11
  * How long a request to this endpoint may take, queueing included. Unset: 300 s for
12
- * chat `/v1` forwards, 600 s for chat the executor answers (Anthropic, Ollama, tools as
13
- * text; not bounded here when streamed), 60 s for embeddings, 120 s for speech.
12
+ * chat `/v1` forwards, 600 s for chat the executor answers, 60 s for embeddings, 120 s
13
+ * for speech. An OpenAI-protocol chat client uses this value for a stream too. A value
14
+ * above 600 s is kept. A stream has no second timer in `within_endpoint_limits`.
14
15
  */
15
16
  requestTimeoutSecs?: number | null,
16
17
  /**