@vincemakes/kiso-runtime 0.38.0 → 0.39.1

package/dist/summarize.d.ts

@@ -124,9 +124,21 @@ export interface SummarizeConversationResult {
     readonly usage: RawUsage | null;
 }
 /**
- * The one-shot summary call. Collects the adapter's text deltas into the
- * summary; usage/stop pass through untouched. Throws when the model
- * produced no text — the caller reports it and nothing is persisted.
+ * The summary call, with ONE retry.
+ *
+ * 0.39.1: `/compact` had no retry at all. The kernel owns retries
+ * (ADR-0005) and this path does not go through the kernel — it calls the
+ * adapter directly — so a gateway that drops a 95k-token summary stream
+ * failed `/compact` every single time, with an honest message and no way
+ * forward. The adapter classifies that as a retryable `network` failure
+ * (the transport-failure-after-headers class); here is the only place
+ * that can act on it.
+ *
+ * ONE retry, not a loop, and not the kernel's budget: a summary re-pays
+ * its whole input, which on the sessions that need compacting is the
+ * most expensive request the session makes. A permanent error still
+ * throws on the first attempt — the retry is keyed on the classification
+ * the adapter already made, never on the fact that something failed.
  */
 export declare function summarizeConversation(options: SummarizeConversationOptions): Promise<SummarizeConversationResult>;
 /** E6 — the crux-experiment drop arm: the covered turns are replaced by

package/dist/summarize.js

@@ -188,12 +188,69 @@ export function validateSummary(text) {
     }
     return null;
 }
+/** A thrown value the adapter CLASSIFIED — `{ code, retryable, message }`
+ *  (ADR-0005 rule 1). The core's own guard is private, and this asks a
+ *  narrower question than that one: is this a failure someone already
+ *  decided is worth trying again. A validation rejection below is a plain
+ *  `Error` and can never match, which is the separation E6 (b) wants —
+ *  "the model wrote a bad summary" is not retried, ever. */
+function retryableStructured(err) {
+    if (err === null || typeof err !== "object")
+        return false;
+    const e = err;
+    return typeof e.code === "string" && typeof e.message === "string" && e.retryable === true;
+}
+/** How long to wait before the one retry. The provider's own ask when it
+ *  made one and it is short enough to be worth honouring; otherwise a
+ *  token pause — the failure this exists for is a dropped stream, not a
+ *  rate limit, and an immediate second attempt is the right response to
+ *  a socket that died. */
+function retryWaitMs(err) {
+    const asked = err.retryAfterMs;
+    return typeof asked === "number" && Number.isFinite(asked) && asked >= 0 && asked <= 60_000 ? asked : 250;
+}
 /**
- * The one-shot summary call. Collects the adapter's text deltas into the
- * summary; usage/stop pass through untouched. Throws when the model
- * produced no text — the caller reports it and nothing is persisted.
+ * The summary call, with ONE retry.
+ *
+ * 0.39.1: `/compact` had no retry at all. The kernel owns retries
+ * (ADR-0005) and this path does not go through the kernel — it calls the
+ * adapter directly — so a gateway that drops a 95k-token summary stream
+ * failed `/compact` every single time, with an honest message and no way
+ * forward. The adapter classifies that as a retryable `network` failure
+ * (the transport-failure-after-headers class); here is the only place
+ * that can act on it.
+ *
+ * ONE retry, not a loop, and not the kernel's budget: a summary re-pays
+ * its whole input, which on the sessions that need compacting is the
+ * most expensive request the session makes. A permanent error still
+ * throws on the first attempt — the retry is keyed on the classification
+ * the adapter already made, never on the fact that something failed.
  */
 export async function summarizeConversation(options) {
+    // Asked as a CALL, twice, because `aborted` is a readonly property and
+    // the narrowing from the first read would otherwise be carried across
+    // the await — where the whole point is that it can have changed.
+    const aborted = () => options.signal?.aborted === true;
+    try {
+        return await summaryAttempt(options);
+    }
+    catch (err) {
+        // The caller's cancel is the caller's: an aborted `/compact` must
+        // not spend a second call on its way out.
+        if (!retryableStructured(err) || aborted())
+            throw err;
+        await new Promise((r) => setTimeout(r, retryWaitMs(err)));
+        if (aborted())
+            throw err;
+        return await summaryAttempt(options);
+    }
+}
+/**
+ * ONE attempt. Collects the adapter's text deltas into the summary;
+ * usage/stop pass through untouched. Throws when the model produced no
+ * text — the caller reports it and nothing is persisted.
+ */
+async function summaryAttempt(options) {
     const { adapter, model, messages } = options;
     let text = "";
     let usage = null;

package/package.json

@@ -1,6 +1,6 @@
 {
 	"name": "@vincemakes/kiso-runtime",
-	"version": "0.38.0",
+	"version": "0.39.1",
 	"description": "kiso runtime \u2014 durable multi-turn agent sessions: AgentDefinition, AgentRuntime, AgentSession, Run, append-only JSONL store.",
 	"type": "module",
 	"license": "MIT",
@@ -25,12 +25,12 @@
 		"test": "vitest run"
 	},
 	"dependencies": {
-		"@vincemakes/kiso-core": "0.38.0"
+		"@vincemakes/kiso-core": "0.39.1"
 	},
 	"peerDependencies": {
-		"@vincemakes/kiso-provider-anthropic": "0.38.0",
-		"@vincemakes/kiso-provider-openai": "0.38.0",
-		"@vincemakes/kiso-provider-openai-responses": "0.38.0"
+		"@vincemakes/kiso-provider-anthropic": "0.39.1",
+		"@vincemakes/kiso-provider-openai": "0.39.1",
+		"@vincemakes/kiso-provider-openai-responses": "0.39.1"
 	},
 	"peerDependenciesMeta": {
 		"@vincemakes/kiso-provider-anthropic": {
@@ -44,7 +44,7 @@
 		}
 	},
 	"devDependencies": {
-		"@vincemakes/kiso-evals": "0.38.0",
+		"@vincemakes/kiso-evals": "0.39.1",
 		"@types/node": "^26.1.2",
 		"typescript": "^5.7.2",
 		"vitest": "^3.0.0"