fix: P4 exit code masking, SIGKILL escalation, error annotations

Author: nam20485

docs/workflow-issues.md

@@ -63,7 +63,12 @@ Now executing **Assignment 3: create-project-structure**. This requires Python a
 
 ## P4: /orchestrate-project-setup timeout and completion issues
 
-**Status: PARTIALLY FIXED** — `--thinking` + `/proc/io` watchdog fixes prevent false-positive kills during subagent work; genuine subagent stalls still cause 15m idle kill
+**Status: FIXED** (commit TBD in `ai-new-workflow-app-template`) — three infrastructure bugs fixed:
+1. **Exit code masking**: Idle-killed runs exited 0, making GitHub Actions report "succeeded" despite incomplete work. Now exits 1.
+2. **`::warning::` → `::error::`**: Idle kills and hard-ceiling kills are failures — annotated as `::error::` so they surface in the workflow summary.
+3. **SIGTERM→SIGKILL escalation**: After sending SIGTERM, waits 10s then sends SIGKILL if the process hasn't exited, preventing zombie/hung processes.
+
+Combined with P5's watchdog race-condition fix, the idle kill is now reliable: it won't false-positive during active subagent work (P5), it properly reports failure when it does fire (P4), and it cleans up stuck processes.
 
 ### Delta86 analysis (run 23332933790 — succeeded in 26m 14s)
 

run_opencode_prompt.sh

@@ -310,17 +310,29 @@ while kill -0 "$OPENCODE_PID" 2>/dev/null; do
 
     if [[ $elapsed -ge $HARD_CEILING_SECS ]]; then
         echo ""
-        echo "::warning::opencode hit ${HARD_CEILING_SECS}s hard ceiling; terminating"
+        echo "::error::opencode hit ${HARD_CEILING_SECS}s hard ceiling; terminating"
         kill "$OPENCODE_PID" 2>/dev/null
+        # Escalate to SIGKILL if SIGTERM doesn't work within 10s
+        sleep 10
+        if kill -0 "$OPENCODE_PID" 2>/dev/null; then
+            echo "::warning::opencode did not exit after SIGTERM; sending SIGKILL"
+            kill -9 "$OPENCODE_PID" 2>/dev/null
+        fi
         IDLE_KILLED=1
         break
     fi
 
     # Idle detection: only trigger when BOTH client output and server are stale
     if [[ $idle -ge $IDLE_TIMEOUT_SECS ]]; then
         echo ""
-        echo "::warning::opencode idle for $(( idle / 60 ))m (no output from client or server); terminating"
+        echo "::error::opencode idle for $(( idle / 60 ))m (no output from client or server); terminating"
         kill "$OPENCODE_PID" 2>/dev/null
+        # Escalate to SIGKILL if SIGTERM doesn't work within 10s
+        sleep 10
+        if kill -0 "$OPENCODE_PID" 2>/dev/null; then
+            echo "::warning::opencode did not exit after SIGTERM; sending SIGKILL"
+            kill -9 "$OPENCODE_PID" 2>/dev/null
+        fi
         IDLE_KILLED=1
         break
     fi
@@ -372,8 +384,11 @@ rm -f "$OUTPUT_LOG"
 
 set -e
 
+# Exit non-zero on idle kill so the workflow properly reports failure.
+# Previously this was `exit 0` which masked SIGTERM (143) as success,
+# causing incomplete runs to appear as "succeeded" in GitHub Actions.
 if [[ $IDLE_KILLED -eq 1 ]]; then
-    exit 0
+    exit 1
 fi
 
 exit ${OPENCODE_EXIT}