test(gate): tolerate EPIPE in the no-node test; harness asserts skill.triage

The no-node test runs the hook with an empty PATH, so cat is missing too
and the script exits before reading stdin; on Linux the test's write can
lose that race (CI run 6483). The child exiting unread is the no-node
path working. Harness: assert_skill_triage on chain and microvm — the
event must exist; agreement with what the agent read is reported.

Co-Authored-By: Claude Opus 5 <[email protected]>
Claude-Session: https://claude.ai/code/session_01WZb5A2kfVfjpdwSochkuHz

crates/cm-api/src/vm_tool_gate.rs

@@ -890,14 +890,15 @@ mod shell_tests {
             .stderr(Stdio::piped())
             .spawn()
             .expect("spawn sh");
-        child
-            .stdin
-            .as_mut()
-            .unwrap()
-            .write_all(
-                br#"{"tool_name":"Bash","tool_input":{"command":"git push --force origin main"}}"#,
-            )
-            .unwrap();
+        // With an empty PATH `cat` is missing too, so the script never reads
+        // its stdin and may exit before this write lands. EPIPE here IS the
+        // no-node path succeeding, not a failure: CI run 6483 lost that race
+        // on Linux after every earlier run had won it.
+        if let Err(e) = child.stdin.as_mut().unwrap().write_all(
+            br#"{"tool_name":"Bash","tool_input":{"command":"git push --force origin main"}}"#,
+        ) {
+            assert_eq!(e.kind(), std::io::ErrorKind::BrokenPipe, "{e}");
+        }
         let out = child.wait_with_output().expect("wait");
 
         assert_eq!(

scripts/verify-mission-delivery.sh

@@ -357,6 +357,7 @@ assert_microvm() { # <token> <mission> <report>
 
   # Project memory: the brief carries earlier verdicts on this repository.
   assert_project_memory "$mission" microvm
+  assert_skill_triage "$token" "$mission" microvm
 
   # And the verdict: judged, and by whom. `independent` is only true when the
   # judge came from a different provider family than the agent.
@@ -1106,6 +1107,7 @@ assert_chain() { # <token> <mission> <report>
     [ "$pushed" = "True" ] || fail "chain: phase $idx not pushed (commit_error=$cerr push_error=$perr)"
     case "$files" in 0|-) fail "chain: phase $idx delivered no files" ;; esac
   done <<<"$report"
+  assert_skill_triage "$token" "$mission" chain
   delivered=$(fetch_delivered "$token" "$mission" CHAIN.md) \
     || { fail "chain: could not read CHAIN.md from the pushed branch"; return 1; }
   case "$delivered" in
@@ -1222,6 +1224,40 @@ assert_credentials_revoked() { # <mission> <label>
   fi
 }
 
+# Shadow skill triage: one skill.triage event per phase, and — where the
+# delivery arm makes retrieval observable — how the oracle's "applies" set
+# compares with what the agent actually read. Informational on agreement
+# (shadow mode measures; it does not select), PASS/FAIL on the event being
+# there at all: a triage that silently stopped firing is the state to catch.
+assert_skill_triage() { # <token> <mission> <label>
+  local token="$1" mission="$2" label="$3" ev agree
+  ev=$(ssh "$HOST" "docker exec clawmates_postgres_1 psql -U postgres -d clawmates -tAc \
+    \"select count(*)::text || ' ' || coalesce(max(detail->>'model'),'-') || ' ' || \
+             coalesce(max((detail->>'latency_ms')::int),0)::text || ' ' || \
+             coalesce(max((select count(*) from jsonb_object_keys(detail->'skills'))),0)::text \
+        from mission_events where mission_id='$mission' and kind='skill.triage';\"" | head -1 | tr -d '\r')
+  set -- $ev
+  case "${1:-0}" in
+    0) fail "$label-triage: no skill.triage event — TYPESAFE_API_KEY unset on the server, or the call failed (see the server log)" ;;
+    *) pass "$label-triage: $1 event(s) by $2 over $4 skills, ${3} ms" ;;
+  esac
+  # Oracle vs agent, from the Skill-Use report: skills the oracle said apply
+  # (triage_p >= 0.5) that the agent read, and ones it read the oracle did
+  # not expect. Only meaningful when the arm is a retrieval arm.
+  agree=$(api "$token" GET "/api/missions/$mission/skill-use" | python3 -c '
+import json,sys
+d=json.load(sys.stdin); sk=d.get("skills") or []
+if not sk or all(s.get("triage_p") is None for s in sk): print("n/a"); sys.exit()
+def read(s): return s["trigger"].get("verdict") in ("pass",) if isinstance(s.get("trigger"),dict) else str(s.get("trigger")).startswith("pass")
+applies=[s for s in sk if (s.get("triage_p") or 0)>=0.5]
+hit=sum(1 for s in applies if read(s)); read_unexpected=sum(1 for s in sk if read(s) and (s.get("triage_p") or 0)<0.5)
+print(f"oracle says {len(applies)} of {len(sk)} delivered skills apply; agent read {hit} of those and {read_unexpected} it did not expect")' 2>/dev/null)
+  case "$agree" in
+    ""|n/a) pass "$label-triage: agreement n/a on this arm (no retrieval to compare against)" ;;
+    *)      pass "$label-triage: $agree" ;;
+  esac
+}
+
 # ── Scenario: multi-role with a real test suite ──────────────────
 #
 # The workload that failed with `COMMIT_EDITMSG: Permission denied` under the