From 0ebe70d574246cb0fdcff9d73af60b1fb03fda18 Mon Sep 17 00:00:00 2001
From: Teknium <127238744+teknium1@users.noreply.github.com>
Date: Tue, 1 Sep 2026 15:46:06 -0700
Subject: [PATCH] fix(agent): long-context tier recovery also rechecks the
 rebuilt request

The Anthropic long-context 429 handler restarts on row count alone,
the same shape #100614 fixed in the generic overflow handler. Arm the
same provider-overflow recovery flag there so the rebuilt request is
measured against the reduced window before the provider is retried.

The 413 (byte-scored) and output-cap (max_tokens) handlers are a
different yardstick and are left as-is.
---
 agent/conversation_loop.py              |  6 ++
 tests/run_agent/test_413_compression.py | 78 +++++++++++++++++++++++++
 2 files changed, 84 insertions(+)

diff --git a/agent/conversation_loop.py b/agent/conversation_loop.py
index 26df1c362a..4469ab395e 100644
--- a/agent/conversation_loop.py
+++ b/agent/conversation_loop.py
@@ -5825,6 +5825,12 @@ def run_conversation(
                                 )
                             )
                             time.sleep(2)
+                            # Same class as the generic overflow handler below:
+                            # the provider proved the request does not fit the
+                            # (now-reduced) window, and row count alone is not
+                            # proof the rebuilt request does. Recheck the
+                            # complete request before the next provider call.
+                            _provider_overflow_recovery_pending = True
                             _retry.restart_with_compressed_messages = True
                             break
                     # Fall through to normal error handling if compression
diff --git a/tests/run_agent/test_413_compression.py b/tests/run_agent/test_413_compression.py
index 20a386d6e7..2c99ff03a0 100644
--- a/tests/run_agent/test_413_compression.py
+++ b/tests/run_agent/test_413_compression.py
@@ -1051,6 +1051,84 @@ class TestPreflightCompression:
         assert mock_compress.call_count == 2
         assert agent.client.chat.completions.create.call_count == 1
 
+    def test_long_context_tier_recovery_rechecks_complete_request_before_retry(self, agent):
+        """The Anthropic long-context 429 handler is the same recovery class.
+
+        It compacts and restarts on row count alone, exactly like the generic
+        overflow handler. The rebuilt request must be measured against the
+        (now-reduced) window before the provider is retried, so a compaction
+        that drops rows but stays oversized fails closed instead of being
+        sent again.
+        """
+        agent.compression_enabled = True
+        agent.max_compression_attempts = 2
+        agent.context_compressor.context_length = 1_000_000
+        agent.context_compressor.threshold_tokens = 500_000
+
+        tier_error = Exception(
+            "Extra usage is required for long context requests."
+        )
+        tier_error.status_code = 429
+        agent.client.chat.completions.create.side_effect = [tier_error]
+
+        history = [
+            {"role": "user", "content": "earlier question"},
+            {"role": "assistant", "content": "earlier answer"},
+        ]
+        compress_calls = 0
+
+        def _request_pressure(*_args, **_kwargs):
+            if agent.client.chat.completions.create.call_count == 0:
+                return 30_000
+            return 250_000
+
+        def _compress(_messages, *_args, **_kwargs):
+            nonlocal compress_calls
+            compress_calls += 1
+            return (
+                [
+                    {"role": "user", "content": f"summary {compress_calls}"},
+                    {"role": "assistant", "content": "summary acknowledged"},
+                ],
+                "rebuilt prompt remains oversized",
+            )
+
+        def _update_model(*, context_length, **_kwargs):
+            agent.context_compressor.context_length = context_length
+            agent.context_compressor.threshold_tokens = context_length // 2
+
+        with (
+            patch(
+                "agent.turn_context.estimate_request_tokens_rough",
+                return_value=30_000,
+            ),
+            patch(
+                "agent.conversation_loop._midturn_request_pressure_tokens",
+                side_effect=_request_pressure,
+            ),
+            patch.object(
+                agent.context_compressor,
+                "should_defer_preflight_to_real_usage",
+                return_value=True,
+            ),
+            patch.object(
+                agent.context_compressor, "update_model", side_effect=_update_model
+            ),
+            patch.object(agent, "_compress_context", side_effect=_compress) as mock_compress,
+            patch.object(agent, "_persist_session"),
+            patch.object(agent, "_save_trajectory"),
+            patch.object(agent, "_cleanup_task_resources"),
+        ):
+            result = agent.run_conversation(
+                "continue",
+                conversation_history=history,
+            )
+
+        assert result["completed"] is False
+        assert result["compression_exhausted"] is True
+        assert mock_compress.call_count == 2
+        assert agent.client.chat.completions.create.call_count == 1
+
 
     def test_interrupt_before_first_provider_call_restores_preflight_display_seed(self, agent):
         """Interrupted turns must not keep a speculative preflight display seed.