@melaya/runner 1.0.108 → 1.0.110

package/dist/assistantHost.py

@@ -142,7 +142,7 @@ _CANCELLED = object()
 class _RedactingStdout:
     """Wrap stdout to strip long base64 runs before they reach the log.
 
-    agentscope prints every agent message verbatim (the `MelayaAssistant: {...}`
+    agentscope prints every agent message verbatim (the `Assistant: {...}`
     lines), which for phone_screenshot tool results embeds a full base64 JPEG —
     tens of KB per turn that floods and explodes the runner logs. This filter is
     line-buffered and only replaces base64-looking runs (>=200 chars), so the
@@ -407,7 +407,13 @@ def _build_agent():
     )
 
     agent = make_agent(
-        name="MelayaAssistant",
+        # This name IS the agent identity that agentscope stamps onto the traced
+        # `invoke_agent <name>` span + gen_ai.agent.name — which is EXACTLY the
+        # dimension the Overview "by agent" token breakdown groups on
+        # (agentStudio.ts:4173/4258). So naming it "Assistant" makes every
+        # runner-assistant turn roll up under an "Assistant" agent in the dashboard
+        # automatically, no dashboard code change.
+        name="Assistant",
         sys_prompt=sys_prompt,
         toolkit=toolkit,
         model_name=model,
@@ -421,8 +427,11 @@ def _build_agent():
         # the task is finished. Give phone runs real headroom (they are HITL-gated,
         # so each write still needs the user's tap — cost stays bounded). Connector
         # WRITE fan-outs (1 list + N sends, each its own reasoning+HITL round) keep
-        # the 40 that fits them; read-only Q&A needs very few.
-        max_iters=100 if phone_enabled else (40 if connector_services else 8),
+        # the 40 that fits them; read-only Q&A needs very few. Phone raised + env-tunable
+        # for long-horizon tasks (100s of steps) now that per-step tokens are pruned +
+        # cached; each write is still HITL-gated so cost stays bounded.
+        max_iters=(int(os.environ.get("MEL_ASSISTANT_MAX_ITERS_PHONE", "300") or "300") if phone_enabled
+                   else (40 if connector_services else 8)),
         reliability=True,
         bounded_memory=True,
     )
@@ -611,6 +620,19 @@ def _stdin_reader(q: "queue.Queue[str]") -> None:
 
 def main() -> int:
     _log(f"booting (provider={os.environ.get('MEL_ASSISTANT_PROVIDER')})")
+    # Wire OpenTelemetry → Melaya event relay so agentscope's @trace_llm actually
+    # exports token spans to agents.spans — the SAME thing pipelines get via
+    # events.setup_studio_forwarder(). Without this the trace gate stays off AND
+    # the exporter is disabled, so every claude_code/codex assistant turn recorded
+    # ZERO tokens (no Overview cost row, no per-turn token count). The exporter
+    # self-gates on MEL_BUILDER_URL + MEL_RUN_ID (both set by the runner spawn);
+    # idempotent + best-effort — a tracing failure NEVER blocks a turn.
+    try:
+        from shared.runtime.tracing_exporter import setup_melaya_tracing
+        setup_melaya_tracing()
+        _log("tracing wired (spans → agents.spans)")
+    except Exception:
+        _log("tracing setup unavailable (non-fatal)")
     # Mirror the spawn-time assistant autonomy mode into MEL_HITL_MODE so the
     # in-process phone tools (phone.py._cmd) see it from turn one.
     _apply_hitl_mode(None)

package/dist/connection.js

@@ -1011,6 +1011,15 @@ export async function connect(opts) {
                 // consistent across surfaces. assistantHost mirrors MEL_ASSISTANT_HITL_MODE
                 // → MEL_HITL_MODE so phone.py stamps the mode on the command body.
                 MEL_RUN_ID: `assistant:${sid.split(":")[0]}`,
+                // Span tracing → agents.spans. tracing_exporter.MelayaSpanExporter only
+                // exports when BOTH MEL_RUN_ID and MEL_BUILDER_URL are present (it POSTs
+                // batches through this same local relay the pipeline spawns use, lines
+                // ~463/853). The assistant host was missing these two, so — even with
+                // @trace_llm wired — every claude_code/codex assistant turn recorded ZERO
+                // token spans. Spans land under conversation_id = MEL_RUN_ID
+                // (`assistant:<userId>`); the server trigger resolves user_id from it.
+                MEL_BUILDER_URL: `http://127.0.0.1:${relay.port}`,
+                MEL_RELAY_NONCE: relay.nonce,
                 MEL_PIPELINE_NAME: "Melaya · Assistant",
                 // Luma browser bridge (same injection as pipeline runs, line ~431): lets
                 // shared/tools/luma.py route /event/register through Playwright and bypass

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@melaya/runner",
-  "version": "1.0.108",
+  "version": "1.0.110",
   "description": "Run Melaya AI pipelines locally with your own LM Studio or Ollama models",
   "license": "UNLICENSED",
   "private": false,