Skip to content

Compress tool outputs, logs, files, and RAG chunks before they reach the LLM. 20% fewer tokens for coding agents, 60-95% fewer tokens for JSON, same answers. Library, proxy, MCP server.

fix(opencode): keep Claude models off OpenAI provider

Suliman Abdulrazzaq5e53b8aa0a72
diff --git a/headroom/providers/opencode/config.py b/headroom/providers/opencode/config.py
index 6c64544fe..08a6f5748 100644
--- a/headroom/providers/opencode/config.py
+++ b/headroom/providers/opencode/config.py
@@ -31,24 +31,18 @@ _MCP_BLOCK_RE = re.compile(
)
HEADROOM_OPENCODE_PLUGIN = "headroom-opencode"
-# Models exposed by the injected `headroom` provider. OpenCode only resolves
-# `headroom/<id>` for ids listed in the provider's `models` map, so an empty
-# map means every documented `headroom/*` model fails with "Model not found".
-# Keep in sync with DEFAULT_MODELS in plugins/opencode/src/provider.ts and the
-# table in plugins/opencode/README.md.
+# Models exposed by the injected `headroom` provider. This provider uses
+# ``@ai-sdk/openai-compatible`` and the proxy's ``/v1/chat/completions`` path,
+# which is routed to the configured OpenAI upstream. Do not advertise Claude
+# models here: OpenCode would send them through the OpenAI endpoint and report
+# an ``invalid_api_key`` error instead of reaching Anthropic. Claude models are
+# available through OpenCode's native ``anthropic`` provider, whose base URL is
+# also redirected to Headroom by ``build_opencode_config_content``.
+#
+# OpenCode only resolves ``headroom/<id>`` for ids listed in this map, so an
+# empty map means every documented ``headroom/*`` model fails with "Model not
+# found".
HEADROOM_OPENCODE_MODELS: dict[str, Any] = {
- "claude-sonnet-4-6": {
- "name": "Claude Sonnet 4.6",
- "limit": {"context": 200000, "output": 16384},
- },
- "claude-opus-4-6": {
- "name": "Claude Opus 4.6",
- "limit": {"context": 200000, "output": 16384},
- },
- "claude-haiku-4-5-20251001": {
- "name": "Claude Haiku 4.5",
- "limit": {"context": 200000, "output": 8192},
- },
"gpt-4o": {
"name": "GPT-4o",
"limit": {"context": 128000, "output": 16384},
diff --git a/tests/test_providers_opencode_config.py b/tests/test_providers_opencode_config.py
index 658ba1ea4..aec387013 100644
--- a/tests/test_providers_opencode_config.py
+++ b/tests/test_providers_opencode_config.py
@@ -199,7 +199,10 @@ def test_inject_provider_config_creates_file(
assert config["provider"]["headroom"]["npm"] == "@ai-sdk/openai-compatible"
# Bare model ids: OpenCode resolves them as "headroom/<id>" (#1657).
models = config["provider"]["headroom"]["models"]
- assert "claude-sonnet-4-6" in models
+ assert set(models) == {"gpt-4o", "gpt-4.1"}
+ # The injected provider is OpenAI-compatible. Claude models must remain on
+ # OpenCode's native Anthropic provider so they are not sent to OpenAI.
+ assert not any(model_id.startswith("claude-") for model_id in models)
assert all(not model_id.startswith("headroom/") for model_id in models)
assert "mcp" not in config
assert "model" not in config # headroom provider is a transparent pass-through
@@ -456,10 +459,12 @@ def test_build_opencode_config_content_without_mcp(
providers = config["provider"]
assert providers["anthropic"]["options"]["baseURL"] == "http://127.0.0.1:8787/v1"
assert providers["openai"]["options"]["baseURL"] == "http://127.0.0.1:8787/v1"
- # The headroom provider exposes explicit models so "headroom/<id>" resolves (#1657).
+ # The headroom provider exposes only models supported by its
+ # OpenAI-compatible endpoint so "headroom/<id>" resolves safely (#1657).
assert providers["headroom"]["options"]["baseURL"] == "http://127.0.0.1:8787/v1"
models = providers["headroom"]["models"]
- assert "claude-sonnet-4-6" in models
+ assert set(models) == {"gpt-4o", "gpt-4.1"}
+ assert not any(model_id.startswith("claude-") for model_id in models)
assert all(not model_id.startswith("headroom/") for model_id in models)
# The transport plugin is injected by absolute path (opencode loads it directly).
assert config["plugin"] == [str(plugin)]