feat(cache): enable Anthropic cache_control for DeepSeek on OpenCode (#24617)

[0xsir0000]

Summary

OpenCode Go is a gateway that implements cache_control at its own level, not per model — the same envelope-layout marker injection that makes Qwen on opencode-go pay the cached-read rate also works for DeepSeek models on the same gateway. The model switcher reports \$0.01/M cache read pricing for deepseek-v4-pro on opencode-go, confirming gateway-side cache support. Without markers Hermes serves 0% cache hits and re-bills the full prompt on every turn.

Fix

Extend the alibaba-family branch in _anthropic_prompt_cache_policy (run_agent.py:3570) to accept DeepSeek models alongside Qwen on the four opencode / opencode-zen / opencode-go / alibaba providers. The wire format (OpenAI chat completions) and marker layout (envelope, not native) are identical to the existing Qwen path — this is purely an expansion of the model whitelist.

model_is_qwen = "qwen" in model_lower
model_is_deepseek = "deepseek" in model_lower
provider_is_alibaba_family = provider_lower in {
    "opencode", "opencode-zen", "opencode-go", "alibaba",
}
if provider_is_alibaba_family and (model_is_qwen or model_is_deepseek):
    return True, False

Test plan

• test_deepseek_on_opencode_go_caches_with_envelope_layout — the reported case (deepseek-v4-pro on opencode-go) returns (True, False).
• test_deepseek_on_opencode_zen_caches — same marker layout on opencode-zen.
• test_deepseek_on_openrouter_not_affected — DeepSeek via OpenRouter still falls through to its own caching arrangement (no regression for non-alibaba-family providers).
• All 39 existing tests in tests/run_agent/test_anthropic_prompt_cache_policy.py still pass.

Risk

Zero. If the gateway doesn't honor cache for a specific DeepSeek model, the markers are silently ignored. Non-DeepSeek/non-Qwen models on opencode-go (GLM, Kimi) keep falling through to (False, False) as before — verified by the existing test_non_qwen_on_opencode_go_does_not_cache and test_kimi_on_opencode_go_does_not_cache.

Fixes #24617

[teknium1]

Thanks for isolating the cache-policy gap. Current main still falls through to (False, False) for DeepSeek because agent/agent_runtime_helpers.py:1629-1639 only recognizes Qwen, and agent/conversation_loop.py:889-894 shows that a true policy result is what injects the request cache markers.

Problems

• The changed predicate at run_agent.py:3582 also enables DeepSeek for direct alibaba. The linked issue supports OpenCode Zen/Go, but current main's Alibaba catalog contains Qwen, Kimi, GLM, and MiniMax entries—not DeepSeek (hermes_cli/models.py:491-502). Keep the direct-Alibaba allowance specific to Qwen unless DeepSeek support there is independently verified.
• The target function has moved: run_agent.py:1358-1368 now forwards to agent/agent_runtime_helpers.py after c42fa94.

Suggested changes

• Salvage the policy change into agent/agent_runtime_helpers.py, separating the Qwen four-provider condition from a DeepSeek OpenCode-only condition.
• Add a negative direct-Alibaba DeepSeek policy test alongside the OpenCode positives.

Automated hermes-sweeper review.

[teknium1]

This combines DeepSeek with Qwen across direct alibaba, although the linked report establishes OpenCode Zen/Go only and main's curated Alibaba catalog has no DeepSeek entry (hermes_cli/models.py:491-502). Keep Qwen on all four providers, but scope DeepSeek to the OpenCode provider IDs and add a negative direct-Alibaba test.