Skip to content

The agent that grows with you

fix(guardrails): no exact-replay warning on a throttled retry after earlier failures

Teknium7d0fb393b715
diff --git a/agent/tool_guardrails.py b/agent/tool_guardrails.py
index 489b6b05d7b..aa35790f82d 100644
--- a/agent/tool_guardrails.py
+++ b/agent/tool_guardrails.py
@@ -393,10 +393,12 @@ class ToolCallGuardrailController:
if self._progress_since_failure.pop(signature, False):
self._exact_failure_counts.pop(signature, None)
# A throttled call retried after its cooldown is the correct move, not a replay:
- # keep it out of the exact-args streak (whose warning says "stop retrying unchanged").
+ # keep it out of the exact-args streak and never give it that streak's "stop retrying
+ # unchanged" warning, even after earlier ordinary failures with the same args.
# It still counts per tool below, so same_tool_failure_halt bounds a dead endpoint.
+ throttled = is_rate_limited(result)
exact_count = self._exact_failure_counts.get(signature, 0)
- if not is_rate_limited(result):
+ if not throttled:
exact_count = self._exact_failure_counts[signature] = exact_count + 1
same_count = self._same_tool_failure_counts[tool_name] = self._same_tool_failure_counts.get(tool_name, 0) + 1
self._no_progress.pop(signature, None)
@@ -414,7 +416,7 @@ class ToolCallGuardrailController:
and same_count >= self.config.same_tool_failure_halt_after
):
return self._decide("halt", "same_tool_failure_halt", tool_name, same_count, signature)
- if warnings and exact_count >= self.config.exact_failure_warn_after:
+ if warnings and not throttled and exact_count >= self.config.exact_failure_warn_after:
return self._decide("warn", "repeated_exact_failure_warning", tool_name, exact_count, signature)
if warnings and same_count >= self.config.same_tool_failure_warn_after:
return self._decide(
diff --git a/tests/agent/test_tool_guardrails.py b/tests/agent/test_tool_guardrails.py
index b9368d7aeab..daf989cf869 100644
--- a/tests/agent/test_tool_guardrails.py
+++ b/tests/agent/test_tool_guardrails.py
@@ -482,21 +482,27 @@ def test_execute_code_replay_streak_notice_fires_on_warn_only_desktop_config():
assert controller.halt_decision is None, "warn-only surfaces must not halt"
[email protected]("result,second_attempt", [
- (json.dumps({"error": "Slack API error: ratelimited. Retry after 1 seconds."}), "allow"),
- (json.dumps({"ok": False, "error": "busy", "retry_after": 30}), "allow"),
- (json.dumps({"error": "channel_not_found"}), "warn"), # control: a real replay still warns
[email protected]("prior_ordinary_failures", [0, 2])
[email protected]("result,throttled", [
+ (json.dumps({"error": "Slack API error: ratelimited. Retry after 1 seconds."}), True),
+ (json.dumps({"ok": False, "error": "busy", "retry_after": 30}), True),
+ (json.dumps({"error": "channel_not_found"}), False), # control: a real replay still warns
])
-def test_rate_limited_retry_is_not_a_replay_but_stays_bounded(result, second_attempt):
+def test_rate_limited_retry_is_not_a_replay_but_stays_bounded(result, throttled, prior_ordinary_failures):
"""Retrying a throttled read after its cooldown must not draw the "stop retrying unchanged"
- warning, yet an endpoint that never recovers still halts unattended runs on the per-tool cap."""
+ warning (even after earlier ordinary failures with the same args), yet an endpoint that never
+ recovers still halts unattended runs on the per-tool cap."""
controller = ToolCallGuardrailController(ToolCallGuardrailConfig.from_mapping({}, platform="webhook"))
- args = {"channel": "C123"}
- actions = []
+ tool, args = "mcp_slack_history", {"channel": "C123"}
+ for _ in range(prior_ordinary_failures):
+ controller.after_call(tool, args, json.dumps({"error": "upstream timeout"}))
+ controller.after_call(tool, args, result)
+ retry = controller.after_call(tool, args, result)
+ assert (retry.code == "repeated_exact_failure_warning") is not throttled, retry
for _ in range(10):
- if not controller.before_call("mcp_slack_history", args).allows_execution:
- actions.append("block")
+ if not controller.before_call(tool, args).allows_execution:
break
- actions.append(controller.after_call("mcp_slack_history", args, result).action)
- assert actions[1] == second_attempt
- assert actions[-1] in {"halt", "block"}, actions
+ if controller.after_call(tool, args, result).action == "halt":
+ break
+ else:
+ raise AssertionError("a permanently failing endpoint must still halt or block")