@boardwalk-labs/runner 0.2.1 → 0.2.4

package/dist/runtime/leaf_executor.js

@@ -238,7 +238,13 @@ export class EngineLeafExecutor {
                 providerIo.onDelta?.(frame.text);
             }
             else if (frame.kind === "result") {
-                result = { turn: frame.turn, modelRef: frame.modelRef };
+                // contextTokens (when the broker knows the served model's window) lets the engine's loop
+                // size compaction against the real window instead of a conservative default.
+                result = {
+                    turn: frame.turn,
+                    modelRef: frame.modelRef,
+                    ...(frame.contextTokens !== undefined ? { contextTokens: frame.contextTokens } : {}),
+                };
                 // The broker stamps the turn's exact upstream cost on the result frame (0 ⇒ none / BYO); hand
                 // it to the budget guardrail via the caller's stash so max_usd caps on real spend.
                 if (frame.costMicros > 0)

package/dist/runtime/wire/inference_proxy.d.ts

@@ -41,6 +41,7 @@ export type InferenceFrame = {
     turn: ChatTurn;
     modelRef: string;
     costMicros: number;
+    contextTokens?: number;
 } | {
     kind: "error";
     error: ProxyError;
@@ -60,7 +61,7 @@ export declare function serializeDeltaFrame(text: string): string;
 /** Serialize the single terminal turn result as one NDJSON line. `costMicros` is the turn's EXACT
  *  upstream cost (the managed provider's per-request cost × 1e6) on the managed lane — 0 for BYO or when unavailable.
  *  The worker feeds it to the budget guardrail so `max_usd` tracks real spend, not a token estimate. */
-export declare function serializeResultFrame(turn: ChatTurn, modelRef: string, costMicros?: number): string;
+export declare function serializeResultFrame(turn: ChatTurn, modelRef: string, costMicros?: number, contextTokens?: number): string;
 /** Serialize a terminal error as a single NDJSON line. */
 export declare function serializeErrorFrame(error: ProxyError): string;
 /** Serialize a heartbeat (no payload) as a single NDJSON line. The broker emits these on an interval

package/dist/runtime/wire/inference_proxy.js

@@ -15,7 +15,8 @@
 // many text deltas; NDJSON lets the runner consume them incrementally over the raw socket (the
 // buffered REST path can't stream). Each frame is one of:
 //   { "t": "delta",  "text": <string> }                  — a streamed assistant-text chunk
-//   { "t": "result", "turn": <ChatTurn>, "modelRef": <string>, "costMicros": <number> }  — terminal turn
+//   { "t": "result", "turn": <ChatTurn>, "modelRef": <string>, "costMicros": <number>,
+//     "contextTokens"?: <number> }                                                       — terminal turn
 //   { "t": "error",  "error": { code, message } }         — a terminal model/broker error
 //
 // A well-formed stream is zero-or-more `delta` frames followed by EXACTLY ONE `result` frame, OR a
@@ -80,8 +81,15 @@ export function serializeDeltaFrame(text) {
 /** Serialize the single terminal turn result as one NDJSON line. `costMicros` is the turn's EXACT
  *  upstream cost (the managed provider's per-request cost × 1e6) on the managed lane — 0 for BYO or when unavailable.
  *  The worker feeds it to the budget guardrail so `max_usd` tracks real spend, not a token estimate. */
-export function serializeResultFrame(turn, modelRef, costMicros = 0) {
-    return `${JSON.stringify({ t: "result", turn, modelRef, costMicros })}\n`;
+export function serializeResultFrame(turn, modelRef, costMicros = 0, contextTokens) {
+    const frame = {
+        t: "result",
+        turn,
+        modelRef,
+        costMicros,
+        ...(contextTokens !== undefined ? { contextTokens } : {}),
+    };
+    return `${JSON.stringify(frame)}\n`;
 }
 /** Serialize a terminal error as a single NDJSON line. */
 export function serializeErrorFrame(error) {
@@ -110,6 +118,11 @@ export function parseInferenceFrame(line) {
                 // A pre-cost-forwarding broker (rolling deploy) omits this → 0 → the worker falls back to the
                 // representative-rate estimate, exactly as before. Non-number / negative also clamps to 0.
                 costMicros: typeof rec.costMicros === "number" && rec.costMicros > 0 ? rec.costMicros : 0,
+                // The served model's context window; absent from a pre-window broker (rolling deploy) or
+                // when its catalog can't say ⇒ the agent loop keeps its conservative default budget.
+                ...(typeof rec.contextTokens === "number" && rec.contextTokens > 0
+                    ? { contextTokens: rec.contextTokens }
+                    : {}),
             };
         case "error":
             return { kind: "error", error: toProxyError(rec.error) };

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@boardwalk-labs/runner",
-  "version": "0.2.1",
+  "version": "0.2.4",
   "description": "Boardwalk self-hosted runner: execute cloud-scheduled runs on your own machines. Currently: the canonical runner contract types.",
   "license": "Apache-2.0",
   "repository": {
@@ -53,8 +53,8 @@
     "prebuild": "prebuildify --napi --strip -t 24.0.0"
   },
   "dependencies": {
-    "@boardwalk-labs/engine": "^0.2.0",
-    "@boardwalk-labs/workflow": "^0.2.0",
+    "@boardwalk-labs/engine": "0.2.3",
+    "@boardwalk-labs/workflow": "0.2.1",
     "@modelcontextprotocol/sdk": "^1.29.0",
     "node-gyp-build": "^4.8.4",
     "tar": "^7.5.16",