fix(OperationsCenter): stop Rule 8 retries on backend-capacity planner failures

src/operations_center/entrypoints/board_worker/outcomes.py

@@ -40,6 +40,11 @@
 # fault, so we never terminate on unknown). See CODE_FAILURE_RETRY_CAP.md.
 _CODE_FAILURE_CATEGORIES = frozenset({"validation_failed", "no_changes"})
 _TERMINAL_STATE_NAMES = {"cancelled", "done"}
+_BACKEND_CAPACITY_REASON_LABEL = "blocked-reason: backend-capacity"
+_BACKEND_CAPACITY_REASON_SNIPPETS = (
+    "stage planner received non-json from agent",
+    "session limit or error",
+)
 
 
 # ── Atomic failure transition ─────────────────────────────────────────────────
@@ -97,6 +102,13 @@ def read_improve_output(workspace: Path) -> list[dict]:
     return [item for item in raw[:5] if isinstance(item, dict) and item.get("title")]
 
 
+def _is_backend_capacity_failure(category: str, reason: str) -> bool:
+    if category != "backend_error":
+        return False
+    lowered = reason.lower()
+    return all(snippet in lowered for snippet in _BACKEND_CAPACITY_REASON_SNIPPETS)
+
+
 # ── Success handler ───────────────────────────────────────────────────────────
 
 
@@ -510,6 +522,11 @@ def handle_failure(
             # 26 policy-blocked re-dispatches in one hour). This label marks the block
             # as policy-driven so Rule 8 can exclude it.
             add_label(client, issue, "blocked-reason: policy")
+        elif _is_backend_capacity_failure(category, reason):
+            # Fallback dispatch hit a backend-capacity/error signature and returned
+            # no structured planner output; immediate clean-retry recycling just
+            # burns the next backend slot with the same non-JSON failure.
+            add_label(client, issue, _BACKEND_CAPACITY_REASON_LABEL)
         if executor_signal:
             add_label(client, issue, f"executor-signal: {executor_signal}")
             if "sigkill" in executor_signal.lower():

src/operations_center/entrypoints/maintenance/board_unblock.py

@@ -79,6 +79,9 @@
       - Not "blocked-reason: policy" (deterministic policy gate — e.g. review.required —
         that will re-block identically on every retry; recycling it Blocked->Backlog->
         Ready for AI is a closed loop, not a transient infra failure)
+      - Not "blocked-reason: backend-capacity" (planner/backend returned no structured
+        output under a backend-capacity condition; immediate recycle just repeats the
+        same non-JSON/session-limit failure on the next backend slot)
       - Blocked for at least --clean-blocked-min-minutes (default 5) minutes
     → move to Backlog for retry.
     These represent pre-execution failures (workspace preparation errors, missing
@@ -268,6 +271,7 @@ def _dispatch_cooldown_reason(
 _HANDOFF_IMPROVEMENT_LABEL = "handoff-reason: improvement_applied"
 _PR_URL_PREFIX = "pr-url:"
 _BLOCKED_REASON_POLICY_LABEL = "blocked-reason: policy"
+_BLOCKED_REASON_BACKEND_CAPACITY_LABEL = "blocked-reason: backend-capacity"
 _OPEN_PR_GATE_LABEL = "OPEN_PR_GATE"
 
 
@@ -803,6 +807,7 @@ def _apply_rules(
             and not _has_label_prefix(labels, "executor-exit-code:")
             and not _has_label_prefix(labels, _BLOCKED_BY_PREFIX)
             and not _has_label(labels, _BLOCKED_REASON_POLICY_LABEL)
+            and not _has_label(labels, _BLOCKED_REASON_BACKEND_CAPACITY_LABEL)
         )
         if is_clean_blocked:
             updated_at = _parse_updated_at(issue)

tests/unit/entrypoints/board_worker/test_outcomes_cov.py

@@ -711,6 +711,23 @@ def test_handle_failure_non_policy_category_no_policy_label():
     assert "blocked-reason: policy" not in _all_update_labels(client)
 
 
+def test_handle_failure_backend_capacity_adds_structured_label():
+    client = _make_client()
+    issue = {"id": "b1", "labels": []}
+    outcomes.handle_failure(
+        client,
+        issue,
+        "goal",
+        "goal",
+        {
+            "failure_category": "backend_error",
+            "failure_reason": "Stage planner received non-JSON from agent (session limit or error): ",
+        },
+        _make_settings(),
+    )
+    assert "blocked-reason: backend-capacity" in _all_update_labels(client)
+
+
 def test_handle_failure_transient_category_does_not_count():
     # Env/transient categories are NOT code failures — must not bump the counter.
     for cat in ("backend_error", "timeout", "unknown", "scope_too_wide"):

tests/unit/entrypoints/maintenance/test_board_unblock.py

@@ -80,6 +80,19 @@ def test_rule8_skip_spec_author_too_young():
     assert len(retry_actions) == 0
 
 
+def test_rule8_skip_backend_capacity_blocked():
+    """Rule 8 must NOT re-queue backend-capacity failures as clean retries."""
+    issue = _issue(
+        "t3b",
+        state="Blocked",
+        labels=["task-kind: spec-author", "blocked-reason: backend-capacity"],
+        updated_at="2026-05-28T10:00:00+00:00",
+    )
+    actions = _apply_rules([issue], **_RULES_KWARGS)
+    retry_actions = [a for a in actions if a["rule"] == "CLEAN_BLOCKED_RETRY"]
+    assert len(retry_actions) == 0
+
+
 # --- Rule 9: SPEC_AUTHOR_BACKLOG_PROMOTE ---
 
 

tests/unit/entrypoints/maintenance/test_board_unblock_cov.py

@@ -740,6 +740,20 @@ def test_rule8_policy_blocked_excluded():
     assert not _by_rule(actions, "CLEAN_BLOCKED_RETRY")
 
 
+def test_rule8_backend_capacity_blocked_excluded():
+    actions = _run(
+        [
+            _issue(
+                "1",
+                state="Blocked",
+                labels=["task-kind: goal", "blocked-reason: backend-capacity"],
+                updated_at=_STALE,
+            )
+        ]
+    )
+    assert not _by_rule(actions, "CLEAN_BLOCKED_RETRY")
+
+
 # ---------------------------------------------------------------------------
 # Rule 9 — SPEC_AUTHOR_BACKLOG_PROMOTE
 # ---------------------------------------------------------------------------