fix(agentic-engineering): forbid foreground remote polling

.claude-plugin/marketplace.json

@@ -27,7 +27,7 @@
     {
       "name": "agentic-engineering",
       "description": "The autonomous engineering system for repository portfolios — engineer, read-only surveyor, and meta-engineer agents; portfolio, product, spend, and improvement workflows; cross-tool instruction architecture and skill discovery; configured by the consumer AGENTS.md",
-      "version": "4.1.2",
+      "version": "4.1.3",
       "source": "./plugins/agentic-engineering"
     },
     {

.github/plugin/marketplace.json

@@ -27,7 +27,7 @@
     {
       "name": "agentic-engineering",
       "description": "The autonomous engineering system for repository portfolios — engineer, read-only surveyor, and meta-engineer agents; portfolio, product, spend, and improvement workflows; cross-tool instruction architecture and skill discovery; configured by the consumer AGENTS.md",
-      "version": "4.1.2",
+      "version": "4.1.3",
       "source": "./plugins/agentic-engineering"
     },
     {

plugins/agentic-engineering/.claude-plugin/plugin.json

@@ -1,7 +1,7 @@
 {
   "name": "agentic-engineering",
   "description": "The autonomous engineering system for repository portfolios — engineer, read-only surveyor, and meta-engineer agents; portfolio, product, spend, and improvement workflows; cross-tool instruction architecture and skill discovery; configured by the consumer AGENTS.md",
-  "version": "4.1.2",
+  "version": "4.1.3",
   "author": {
     "name": "devantler-tech",
     "url": "https://github.com/devantler-tech"

plugins/agentic-engineering/agents/agentic-engineer.agent.md

@@ -109,8 +109,11 @@ instead.
    invoking it. When the runtime exposes no per-call setting, use an equivalent bounded process
    supervisor that preserves output and exit status; when neither control exists, split the command
    into bounded targets or record the missing capability rather than launching a known-too-long
-   command. This does not authorize a long foreground remote-state poll, retry, or sleep loop. Keep
-   remote waits asynchronous where the runtime supports it and keep doing useful work.
+   command. **Never foreground-wait on remote state.** **Keep remote waits asynchronous:** for CI,
+   review, merge, or deploy state, arm
+   at most one detached watcher when the runtime supports it; never run a foreground polling or
+   sleep loop, and never poll beside an armed watcher. Continue with other actionable work. If none
+   remains, end the run and let the next scheduled tick collect the result.
 8. **Spend context deliberately.** Delegate the survey to the read-only **`portfolio-surveyor`**
    subagent (your runtime may expose this bundled agent under a plugin-scoped name — e.g.
    `agentic-engineering:portfolio-surveyor` — so select it by whatever qualified

plugins/agentic-engineering/plugin.json

@@ -1,7 +1,7 @@
 {
   "name": "agentic-engineering",
   "description": "The autonomous engineering system for repository portfolios — engineer, read-only surveyor, and meta-engineer agents; portfolio, product, spend, and improvement workflows; cross-tool instruction architecture and skill discovery; configured by the consumer AGENTS.md",
-  "version": "4.1.2",
+  "version": "4.1.3",
   "author": {
     "name": "devantler-tech",
     "url": "https://github.com/devantler-tech"

scripts/validate-manifests.sh

@@ -774,6 +774,19 @@ validate_desired_state_resources() {
       fi
     done
 
+    for remote_wait_marker in \
+      '**Never foreground-wait on remote state.**' \
+      'at most one detached watcher' \
+      'end the run and let the next scheduled tick collect'; do
+      if [ ! -f "$plugin_dir/agents/$entrypoint.agent.md" ] \
+        || ! grep -qF "$remote_wait_marker" \
+          "$plugin_dir/agents/$entrypoint.agent.md"; then
+        echo "::error::$resource: agentic-engineer must forbid foreground remote waits, missing: $remote_wait_marker"
+        failed=1
+        resource_failed=1
+      fi
+    done
+
     if [ ! -f "$plugin_dir/agents/agent-improver.agent.md" ] \
       || ! grep -qF "## Delivery ownership — finding to fix" \
         "$plugin_dir/agents/agent-improver.agent.md"; then

scripts/validate-manifests.test.sh

@@ -570,6 +570,9 @@ Use a **bounded tool timeout** from the **measured repository or CI duration** p
 When the runtime exposes no per-call setting, use an equivalent bounded process supervisor.
 Keep remote waits asynchronous.
 
+**Never foreground-wait on remote state.** Arm at most one detached watcher.
+If no other work remains, end the run and let the next scheduled tick collect the result.
+
 ## Spend stewardship
 
 - **You never move money.**
@@ -964,6 +967,18 @@ for deadline_marker in \
     "agentic-engineer must bound expected-to-run-long local commands" "$d"
 done
 
+for remote_wait_marker in \
+  'Never foreground-wait on remote state' \
+  'at most one detached watcher' \
+  'end the run and let the next scheduled tick collect'; do
+  d=$(fresh); make_desired_state "$d" alpha
+  grep -vF "$remote_wait_marker" \
+    "$d/plugins/alpha/agents/agentic-engineer.agent.md" > "$d/tmp" \
+    && mv "$d/tmp" "$d/plugins/alpha/agents/agentic-engineer.agent.md"
+  check_fail "Agentic Engineer requires remote wait marker: $remote_wait_marker" \
+    "agentic-engineer must forbid foreground remote waits" "$d"
+done
+
 d=$(fresh); make_desired_state "$d" alpha
 jq '.spec.guardrails |= map(select(startswith("Write-capable roles own selected engineering work") | not))' \
   "$d/plugins/alpha/resources/provider-neutral.desired-state.json" > "$d/tmp" \