hermes: carry goal-loop controller evidence

services/hermes/agent-deployment.yaml

@@ -25,7 +25,7 @@ spec:
         ai.bstein.dev/execution: Hermes Kanban with durable direct Codex and Claude Code CLI workers
         ai.bstein.dev/model-policy: Jetson-assisted AUTO routing, low through xhigh, cross-provider fallback
         ai.bstein.dev/placement: rpi5 preferred; Jetson deferred until state storage is available
-        ai.bstein.dev/config-rev: "20260816-goal-completion-judge-v1"
+        ai.bstein.dev/config-rev: "20260816-goal-completion-judge-v2"
         prometheus.io/scrape: "true"
         prometheus.io/path: /metrics
         prometheus.io/port: "9010"

services/hermes/scripts/cli_lane_runner.py

@@ -862,6 +862,7 @@ def execute_claim(board: str, task_id: str) -> None:
             excluded_provider = (
                 fresh_unavailable_provider() if assignee == "cli-auto" else None
             )
+            unavailable_provider = excluded_provider
             route = select_route(
                 context,
                 assignee,
@@ -914,6 +915,7 @@ def execute_claim(board: str, task_id: str) -> None:
                     remaining,
                 )
                 if result.capacity_failure:
+                    unavailable_provider = route.provider
                     retry_context = (
                         context
                         + "\n\nRouting boundary: the first provider failed from capacity/authentication. "
@@ -975,8 +977,18 @@ def execute_claim(board: str, task_id: str) -> None:
                         and goal_mode
                     ):
                         heartbeat("local goal-completion judge active")
+                        rejection_history = state.get("goal_rejections", [])
+                        if not isinstance(rejection_history, list):
+                            rejection_history = []
+                        judge_context = context
+                        if goal_turn > 1 or rejection_history:
+                            judge_context += (
+                                "\n\nAuthoritative Hermes goal-controller evidence: "
+                                f"current turn {goal_turn}/{goal_max_turns}; prior rejected "
+                                f"reports: {json.dumps(rejection_history[-5:])}."
+                            )
                         accepted, judge_reason = cli_lane_goal.judge_goal_completion(
-                            context,
+                            judge_context,
                             structured,
                         )
                         metadata["goal_judge_reason"] = judge_reason
@@ -1010,6 +1022,13 @@ def execute_claim(board: str, task_id: str) -> None:
                     and deadline - time.monotonic() > 30
                 )
                 if can_continue:
+                    rejection_history = state.get("goal_rejections", [])
+                    if not isinstance(rejection_history, list):
+                        rejection_history = []
+                    state["goal_rejections"] = [
+                        *rejection_history[-4:],
+                        completion_problem,
+                    ]
                     goal_turn += 1
                     comment(
                         f"Goal completion rejected; continuing turn {goal_turn}/{goal_max_turns}: {completion_problem}",
@@ -1020,9 +1039,9 @@ def execute_claim(board: str, task_id: str) -> None:
                         + completion_problem
                         + "\nSelect a route that can finish and verify the remaining work."
                     )
-                    excluded = (
+                    excluded = unavailable_provider
-                        fresh_unavailable_provider() if assignee == "cli-auto" else None
+                    if excluded is None and assignee == "cli-auto":
-                    )
+                        excluded = fresh_unavailable_provider()
                     next_route = select_route(
                         escalation_context,
                         assignee,
@@ -1031,6 +1050,11 @@ def execute_claim(board: str, task_id: str) -> None:
                         if excluded
                         else None,
                     )
+                    comment(
+                        f"Goal route {goal_turn}/{goal_max_turns}: "
+                        f"{next_route.provider}/{next_route.model} at {next_route.effort}; "
+                        f"classifier={next_route.classifier}; {next_route.reason}",
+                    )
                     handoff = (
                         git_handoff(workspace, result.output)
                         if next_route.provider != route.provider

testing/tests/test_hermes_cli_lanes.py

@@ -727,12 +727,29 @@ def test_goal_card_continues_after_local_judge_rejects_progress(
         "state_path",
         lambda _board, _task_id: tmp_path / "state.json",
     )
-    route = lanes.Route(
+    claude_low = lanes.Route(
-        "codex", "gpt-5.6-sol", "xhigh", "codex-xhigh", "jetson", "vote", 1, ()
+        "claude", "claude-fable-5", "low", "claude-low", "jetson", "vote", 1, ()
     )
-    monkeypatch.setattr(lanes, "select_route", lambda *_args, **_kwargs: route)
+    codex_low = lanes.Route(
+        "codex", "gpt-5.6-luna", "low", "codex-low", "manual", "fallback", 1, ()
+    )
+    codex_xhigh = lanes.Route(
+        "codex", "gpt-5.6-sol", "xhigh", "codex-xhigh", "jetson", "escalated", 1, ()
+    )
+    route_calls = []
+
+    def select_route(_prompt, assignee, **kwargs):
+        route_calls.append((assignee, kwargs))
+        if assignee == "cli-codex-low":
+            return codex_low
+        if len(route_calls) == 1:
+            return claude_low
+        return codex_xhigh
+
+    monkeypatch.setattr(lanes, "select_route", select_route)
     monkeypatch.setattr(lanes, "fresh_unavailable_provider", lambda: None)
     reports = [
+        lanes.ProcessResult(1, "authentication expired", None, True),
         lanes.ProcessResult(
             0,
             "first turn",
@@ -767,10 +784,16 @@ def test_goal_card_continues_after_local_judge_rejects_progress(
             (True, "all explicit acceptance criteria have evidence"),
         ]
     )
+    judge_contexts = []
+
+    def judge_goal_completion(objective, *_args, **_kwargs):
+        judge_contexts.append(objective)
+        return next(verdicts)
+
     monkeypatch.setattr(
         lanes.cli_lane_goal,
         "judge_goal_completion",
-        lambda *_args, **_kwargs: next(verdicts),
+        judge_goal_completion,
     )
 
     lanes.execute_claim("cassandra", "t_goal")
@@ -778,6 +801,10 @@ def test_goal_card_continues_after_local_judge_rejects_progress(
     assert calls[0][0] == "complete"
     assert calls[0][1]["metadata"]["goal_turn"] == 2
     assert any("Goal completion rejected; continuing turn 2/3" in item for item in comments)
+    assert any("Goal route 2/3: codex/gpt-5.6-sol at xhigh" in item for item in comments)
+    assert route_calls[2][1]["exclude_provider"] == "claude"
+    assert "prior rejected reports" in judge_contexts[1]
+    assert "commit, push, and remote verification are missing" in judge_contexts[1]
     assert reports == []