From 6b6435a8748859cc256c7b1d1b4bf1a195d99197 Mon Sep 17 00:00:00 2001
From: wz-heng <68931789+wz-heng@users.noreply.github.com>
Date: Thu, 30 Jul 2026 20:39:08 +0800
Subject: [PATCH] feat(cache): enable DeepSeek caching on OpenCode

---
 agent/agent_runtime_helpers.py                | 30 ++++++++------
 .../test_anthropic_prompt_cache_policy.py     | 41 ++++++++++++++++++-
 2 files changed, 58 insertions(+), 13 deletions(-)

diff --git a/agent/agent_runtime_helpers.py b/agent/agent_runtime_helpers.py
index 11f98c82e945d..5f72bedfb5628 100644
--- a/agent/agent_runtime_helpers.py
+++ b/agent/agent_runtime_helpers.py
@@ -1860,12 +1860,12 @@ def anthropic_prompt_cache_policy(
     gateway implements the Anthropic cache_control contract
     (MiniMax, Zhipu GLM, LiteLLM's Anthropic proxy mode all do).
 
-    Qwen / Alibaba-family models on OpenCode, OpenCode Go, and direct
-    Alibaba (DashScope) also honour Anthropic-style ``cache_control``
-    markers on OpenAI-wire chat completions. Upstream pi-mono #3392 /
-    pi #3393 documented this for opencode-go Qwen. Without markers
-    these providers serve zero cache hits, re-billing the full prompt
-    on every turn.
+    Qwen models on OpenCode and direct Alibaba (DashScope), plus DeepSeek
+    models on OpenCode, also honour Anthropic-style ``cache_control`` markers
+    on OpenAI-wire chat completions. Upstream pi-mono #3392 / pi #3393
+    documented this for opencode-go Qwen; #24617 reports the same gateway
+    contract for DeepSeek. Without markers these providers serve zero cache
+    hits, re-billing the full prompt on every turn.
     """
     eff_provider = (provider if provider is not None else agent.provider) or ""
     eff_base_url = base_url if base_url is not None else (agent.base_url or "")
@@ -1980,16 +1980,22 @@ def anthropic_prompt_cache_policy(
         if is_minimax_provider or is_minimax_host:
             return True, True
 
-    # Qwen/Alibaba on OpenCode (Zen/Go) and native DashScope: OpenAI-wire
-    # transport that accepts Anthropic-style cache_control markers and
-    # rewards them with real cache hits.  Without this branch
-    # qwen3.6-plus on opencode-go reports 0% cached tokens and burns
-    # through the subscription on every turn.
+    # Qwen on OpenCode (Zen/Go) and native DashScope, plus DeepSeek on
+    # OpenCode only: OpenAI-wire transports that accept Anthropic-style
+    # cache_control markers and reward them with real cache hits. Keep direct
+    # Alibaba specific to Qwen; its catalog does not establish the same
+    # contract for DeepSeek.
     model_is_qwen = "qwen" in model_lower
+    model_is_deepseek = "deepseek" in model_lower
+    provider_is_opencode = provider_lower in {
+        "opencode", "opencode-zen", "opencode-go",
+    }
     provider_is_alibaba_family = provider_lower in {
         "opencode", "opencode-zen", "opencode-go", "alibaba",
     }
-    if provider_is_alibaba_family and model_is_qwen:
+    if (provider_is_alibaba_family and model_is_qwen) or (
+        provider_is_opencode and model_is_deepseek
+    ):
         # Envelope layout (native_anthropic=False): markers on inner
         # content parts, not top-level tool messages.  Matches
         # pi-mono's "alibaba" cacheControlFormat.
diff --git a/tests/run_agent/test_anthropic_prompt_cache_policy.py b/tests/run_agent/test_anthropic_prompt_cache_policy.py
index 34cc748a2e4bc..f456c107eefd6 100644
--- a/tests/run_agent/test_anthropic_prompt_cache_policy.py
+++ b/tests/run_agent/test_anthropic_prompt_cache_policy.py
@@ -10,6 +10,8 @@ from __future__ import annotations
 
 from unittest.mock import MagicMock
 
+import pytest
+
 from run_agent import AIAgent
 
 
@@ -256,6 +258,44 @@ class TestQwenAlibabaFamily:
         assert agent._anthropic_prompt_cache_policy() == (False, False)
 
 
+class TestDeepSeekOpenCode:
+    """DeepSeek uses OpenCode's envelope-layout cache markers (#24617)."""
+
+    @pytest.mark.parametrize(
+        "provider",
+        ["opencode", "opencode-zen", "opencode-go"],
+    )
+    def test_deepseek_on_opencode_caches_with_envelope_layout(self, provider):
+        agent = _make_agent(
+            provider=provider,
+            base_url="https://opencode.ai/v1",
+            api_mode="chat_completions",
+            model="deepseek-v4-pro",
+        )
+
+        assert agent._anthropic_prompt_cache_policy() == (True, False)
+
+    def test_deepseek_on_direct_alibaba_does_not_cache(self):
+        agent = _make_agent(
+            provider="alibaba",
+            base_url="https://dashscope.aliyuncs.com/compatible-mode/v1",
+            api_mode="chat_completions",
+            model="deepseek-v4-pro",
+        )
+
+        assert agent._anthropic_prompt_cache_policy() == (False, False)
+
+    def test_deepseek_on_openrouter_does_not_cache(self):
+        agent = _make_agent(
+            provider="openrouter",
+            base_url="https://openrouter.ai/api/v1",
+            api_mode="chat_completions",
+            model="deepseek/deepseek-chat",
+        )
+
+        assert agent._anthropic_prompt_cache_policy() == (False, False)
+
+
 class TestNousPortalAnthropicWire:
     def test_portal_claude_on_the_messages_wire_uses_the_native_layout(self):
         agent = _make_agent(
@@ -315,4 +355,3 @@ class TestExplicitOverrides:
 # ─────────────────────────────────────────────────────────────────────
 # Long-lived prefix cache policy (cross-session 1h tier)
 # ─────────────────────────────────────────────────────────────────────
-