fix(attribution): demote agentguild-skill UA to propagation_client (HTTP+MCP), never auto-qualified genuine external; retro read-time reclassification; invert discovery-reach census test

live/guild/app/attribution.py

@@ -11,8 +11,8 @@
 ITSELF AS AN AGENT in a way our own traffic does not:
   * an MCP client that named itself in the handshake (`mcp:<client>`) and is not one
     of ours, OR
-  * a recognised agent-framework user-agent (httpx/langchain/openai/... ) that is not
-    bare tooling.
+  * a recognised independent agent-framework user-agent
+    (httpx/langchain/openai/... ) that is not bare tooling.
 
 Deliberately NOT sufficient (all indistinguishable from our own traffic, so counting
 them would fool us):
@@ -33,10 +33,18 @@
 
 FRAMEWORK_RE = re.compile(
     r"(httpx|aiohttp|langchain|langgraph|openai|anthropic|claude|llamaindex|"
-    r"crewai|autogen|agentguild-skill|"
+    r"crewai|autogen|"
     r"semantic-kernel|node-fetch|undici|axios|okhttp|go-http-client|reqwest|"
     r"cursor|cline|continue|windsurf|cody|dify|n8n|flowise)", re.I)
 
+# Agent Guild publishes this User-Agent in its own installable skill and tells
+# downstream runtimes to preserve it. It proves propagation of our client
+# instructions, not an independent counterparty: the same identifier can be
+# emitted by our own worker or copied by any caller. Keep it distinct from
+# AG_TEST (a real third party may use the skill) and from EXTERNAL_* (the UA
+# alone is never enough to claim external demand or revenue).
+PROPAGATION_UA_RE = re.compile(r"\bagentguild-skill(?:/|\b)", re.I)
+
 # Bare tooling — indistinguishable from our own verification calls. NOT genuine.
 # `guild-ops-check` is our own scheduled ops probe and is named here explicitly
 # so advertised telemetry (discovery_stats) can never count our own heartbeat
@@ -233,6 +241,12 @@ def is_genuine_external(event: dict[str, Any]) -> bool:
         return False
     ua = (event.get("ua", event.get("user_agent")) or "").strip()
 
+    # This AG-authored identity may appear as either an HTTP User-Agent or an
+    # MCP clientInfo name (`mcp:agentguild-skill/...`). Guard it before the MCP
+    # short-circuit: neither transport proves an independent counterparty.
+    if PROPAGATION_UA_RE.search(ua):
+        return False
+
     # A self-identified MCP client that isn't one of ours.
     client = _mcp_client(ua)
     if client is not None:
@@ -260,6 +274,8 @@ def attribution_class(event: dict[str, Any]) -> str:
         return "ag_test"              # our own self-identified harnesses
     if CRAWLER_UA_RE.search(ua):
         return "registry_crawler"     # indexes manifests, never an agent
+    if PROPAGATION_UA_RE.search(ua):
+        return "propagation_client"   # AG-authored identifier, no external proof
     if ua == "mcp/remote":
         return "unattributable_mcp"
     if not ua or TOOLING_UA_RE.search(ua):
@@ -279,7 +295,8 @@ def attribution_class(event: dict[str, Any]) -> str:
 
 CALLER_CLASSES = (
     "AG_INTERNAL", "AG_TEST", "REGISTRY_CRAWLER",
-    "EXTERNAL_UNKNOWN", "EXTERNAL_VERIFIED", "EXTERNAL_MEMBER", "OPERATOR",
+    "PROPAGATION_CLIENT", "EXTERNAL_UNKNOWN", "EXTERNAL_VERIFIED",
+    "EXTERNAL_MEMBER", "OPERATOR",
 )
 
 # Registry / search-engine / uptime crawlers: they index manifests, they do
@@ -378,6 +395,8 @@ def caller_class(event: Mapping[str, Any], *,
         return "AG_TEST"
     if CRAWLER_UA_RE.search(ua):
         return "REGISTRY_CRAWLER"
+    if PROPAGATION_UA_RE.search(ua):
+        return "PROPAGATION_CLIENT"
     if member and verified:
         return "EXTERNAL_VERIFIED"
     if member:

live/guild/app/store.py

@@ -3033,7 +3033,8 @@ def discovery_reach(self, target: int = 25_000, *,
 
         * T1: a key-proved Guild member;
         * T2: a named third-party MCP client;
-        * T3: a recognised agent-framework UA (Sybil-unresolvable).
+        * T3: a recognised independent agent-framework UA
+          (Sybil-unresolvable).
 
         First-party traffic, Guild tests, generic tooling, registry crawlers,
         anonymous/unlinkable calls and repeated fetches are structurally
@@ -3236,7 +3237,8 @@ def discovery_reach(self, target: int = 25_000, *,
                 "One durable privacy-safe actor key counts once, across every "
                 "origin surface and repeat call, only when it is a key-proved "
                 "Guild member (T1), a named third-party MCP client (T2), or a "
-                "recognised agent-framework caller (T3). First-party traffic, Guild "
+                "recognised independent agent-framework caller (T3). First-party "
+                "traffic, Guild "
                 "tests, generic tooling, registry crawlers, anonymous or "
                 "unlinkable traffic, unproved bare members, and raw repeat "
                 "impressions never count."),

live/guild/tests/test_analytics_invariant.py

@@ -128,5 +128,5 @@ def test_operator_kill_switch_event_is_audit_only():
 
 def test_taxonomy_is_closed():
     assert set(CALLER_CLASSES) == {
-        "AG_INTERNAL", "AG_TEST", "REGISTRY_CRAWLER", "EXTERNAL_UNKNOWN",
-        "EXTERNAL_VERIFIED", "EXTERNAL_MEMBER", "OPERATOR"}
+        "AG_INTERNAL", "AG_TEST", "REGISTRY_CRAWLER", "PROPAGATION_CLIENT",
+        "EXTERNAL_UNKNOWN", "EXTERNAL_VERIFIED", "EXTERNAL_MEMBER", "OPERATOR"}

live/guild/tests/test_caller_classes.py

@@ -63,6 +63,22 @@ def test_growth_gate_is_closed_by_type():
     assert external == {"EXTERNAL_UNKNOWN", "EXTERNAL_VERIFIED", "EXTERNAL_MEMBER"}
 
 
+def test_agentguild_skill_ua_is_propagation_not_external_demand():
+    """Our published skill UA cannot prove an independent counterparty."""
+    from app.attribution import attribution_class
+
+    for ua in (
+        "agentguild-skill/1.0 (host=openclaw)",
+        "agentguild-skill/1.1 (host=codex; source=awesome-copilot)",
+        "mcp:agentguild-skill/1.1 (host=codex)",
+    ):
+        event = _e(ua=ua)
+        assert caller_class(event) == "PROPAGATION_CLIENT"
+        assert attribution_class(event) == "propagation_client"
+        assert not may_count_as_external_growth(caller_class(event))
+        assert not is_genuine_external(event)
+
+
 def test_instrumentation_exposes_caller_class_counts():
     r = client.post("/agents/register",
                     json={"name": "CCTest", "capabilities": ["x"]},

live/guild/tests/test_discovery_reach.py

@@ -92,16 +92,24 @@ def test_machine_resource_fetch_records_one_noncommercial_observation():
     } for event in new)
 
 
-def test_agent_skill_identity_is_transparent_and_qualified(tmp_path):
+def test_agent_skill_identity_is_transparent_but_not_independent(tmp_path):
     store = Store(path=str(tmp_path / "guild.json"))
     store.record_event(
         "http:skill-user", "query",
         ua="agentguild-skill/1.0 (host=openclaw)",
         endpoint="check", actor_distinct=True,
     )
+    store.record_event(
+        "mcp:skill-user", "query",
+        ua="mcp:agentguild-skill/1.1 (host=codex)",
+        endpoint="check", actor_distinct=True,
+    )
     report = store.discovery_reach()
-    assert report["qualified_distinct_autonomous_agents"] == 1
-    assert report["tiers"]["T3_framework_ua_actors"] == 1
+    assert report["qualified_distinct_autonomous_agents"] == 0
+    assert report["tiers"]["T3_framework_ua_actors"] == 0
+    assert report["evidence"]["excluded_distinct_actors_by_reason"] == {
+        "propagation_client": 2,
+    }
 
 
 def test_exact_registry_scanner_aliases_are_excluded_without_ua_overreach(