agent-relay-server 0.121.3 → 0.121.4

package/docs/openapi.json

@@ -2,7 +2,7 @@
   "openapi": "3.1.0",
   "info": {
     "title": "Agent Relay API",
-    "version": "0.121.3",
+    "version": "0.121.4",
     "description": "Real-time message bus for inter-agent communication. Agent-first: this spec is designed for machine consumption — agents can self-discover the full API surface via GET /api/spec.",
     "license": {
       "name": "MIT",

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "agent-relay-server",
-  "version": "0.121.3",
+  "version": "0.121.4",
   "description": "Lightweight HTTP message relay for inter-agent communication across machines",
   "module": "src/index.ts",
   "type": "module",
@@ -37,7 +37,7 @@
     "CONTRIBUTING.md"
   ],
   "dependencies": {
-    "agent-relay-channels-host": "0.121.3",
+    "agent-relay-channels-host": "0.121.4",
     "agent-relay-providers": "0.104.3",
     "agent-relay-sdk": "0.2.113",
     "ajv": "^8.20.0"

package/runner/plugins/claude/.claude-plugin/plugin.json

@@ -1,7 +1,7 @@
 {
   "name": "agent-relay-runner",
   "description": "Thin Agent Relay runner bridge for Claude Code",
-  "version": "0.121.3",
+  "version": "0.121.4",
   "agentRelayContracts": {
     "providerPluginProtocol": 1
   }

package/src/db/messages.ts

@@ -354,17 +354,33 @@ function childIsBusy(child: AgentCard): boolean {
 // verbatim floods the coordinator with a wall of mid-work text (the exact bug #1037 keeps
 // re-shipping). Idle-gating alone can't help: this is a genuine turn-final at a genuine idle, so
 // the payload itself must be distilled. Short bodies pass through untouched (concise finals are
-// unaffected); a long body is trimmed to its trailing result block(s) — a turn's conclusion lives
-// at its end — with an elision marker so the coordinator knows narration was dropped. The actual
-// result is never lost. Pure function of the body → deterministic, so re-promotion (the idle flush)
-// yields the identical distilled body + idempotency key and dedups cleanly.
+// unaffected). Pure function of the body → deterministic, so re-promotion (the idle flush) yields
+// the identical distilled body + idempotency key and dedups cleanly.
+//
+// #1056 — cut on a SEMANTIC result boundary, not a block count. The old heuristic kept the last N
+// `\n\n`-separated blocks by a size budget; that is fragile to how a model CHUNKS its answer. A
+// chatty Claude emitted its `## SUMMARY` result as one block (kept whole), but a chatty Codex
+// paragraph-broke the heading away from its bullets, so the tail-block budget clipped the header +
+// the first bullet — distillation ate the TOP of the result, delivering it mid-sentence. Fix: keep
+// everything from the FIRST markdown heading onward. Workers conclude under a heading (`## SUMMARY`,
+// `## FINAL REPORT`), narration is prose chatter emitted before it — so the result is the trailing
+// region a heading marks the start of, and cutting at the heading drops the leading narration while
+// keeping the WHOLE result, however the model split it. We cut at the FIRST heading (not the last):
+// the result may itself contain sub-headings (`### Details`) or multiple `##` sections, and "last
+// heading" would slice into it — the very clip #1056 is about; "first heading" never can, since it
+// sits at or before the result's start. Fallback (no heading anywhere — rare; workers are told to
+// emit one): there is no reliable boundary, so keep the trailing answer and trim only the leading
+// narration flood by the legacy tail budget, never slicing the block we keep.
 const REPORT_UP_DISTILL_THRESHOLD = 800;
 const REPORT_UP_TAIL_MAX = 800;
 const REPORT_UP_ELISION = "[…mid-work narration trimmed by report-up…]";
+const REPORT_UP_HEADING = /^#{1,6}\s/m;
 
 function distillLineageReportBody(body: string): string {
   const trimmed = body.trim();
   if (trimmed.length <= REPORT_UP_DISTILL_THRESHOLD) return body;
+  const heading = REPORT_UP_HEADING.exec(trimmed);
+  if (heading) return heading.index === 0 ? body : `${REPORT_UP_ELISION}\n\n${trimmed.slice(heading.index).trim()}`;
   const blocks = trimmed.split(/\n{2,}/).map((block) => block.trim()).filter(Boolean);
   const kept: string[] = [];
   let size = 0;