feat: add support for OpenCode Go provider

dev-notes/opencode-go-support.md

@@ -0,0 +1,64 @@
+# OpenCode Go & Zen Free Models Capabilities
+
+| Model Name | Display Name | Base URL | API | Thinking | Default Level | Thinking Modes | API Parameter / Payload Mapped | Context Window | Max Output / Reasoning Tokens | Additional Details |
+|---|---|---|---|---|---|---|---|---|---|---|
+| **Go Models** | | | | | | | | | | |
+| `deepseek-v4-pro` | DeepSeek V4 Pro | `https://opencode.ai/zen/go/v1` | OpenAI | Always on | `high` | `high`, `xhigh` (max) | `reasoning_effort` | 1,000,000 | 384,000 | Effort values: `"high"`, `"max"`. Inherits from `deepseek/deepseek-v4-pro`. |
+| `deepseek-v4-flash` | DeepSeek V4 Flash | `https://opencode.ai/zen/go/v1` | OpenAI | Always on | `high` | `high`, `xhigh` (max) | `reasoning_effort` | 1,000,000 | 384,000 | Effort values: `"high"`, `"max"`. Inherits from `deepseek/deepseek-v4-flash`. |
+| `glm-5.2` | GLM 5.2 | `https://opencode.ai/zen/go/v1` | OpenAI | Always on | `high` | `high`, `xhigh` (max) | `reasoning_effort` | 1,000,000 | 131,072 | Effort values: `"high"`, `"max"`. Inherits from `zhipuai/glm-5.2`. |
+| `glm-5.1` | GLM 5.1 | `https://opencode.ai/zen/go/v1` | OpenAI | Always on | N/A | None | None | 202,752 | 32,768 | Always-on reasoning. No reasoning options defined in TOML. |
+| `grok-4.5` | Grok 4.5 | `https://opencode.ai/zen/go/v1` | OpenAI | Toggleable | `medium` | `off` (none), `low`, `medium`, `high` | `reasoning_effort` | 1,000,000 | — | Tiered pricing: 200K+ tokens cost 2x. Inherits from `xai/grok-4.5`. |
+| `kimi-k2.7-code` | Kimi K2.7 Code | `https://opencode.ai/zen/go/v1` | OpenAI | Always on | N/A | None | None | 262,144 | 262,144 | Always-on reasoning. Resolved from compiled models.dev catalog. |
+| `kimi-k2.6` | Kimi K2.6 | `https://opencode.ai/zen/go/v1` | OpenAI | Always on | N/A | None | None | 262,144 | 65,536 | Always-on reasoning. No reasoning options defined in TOML. |
+| `kimi-k3` | Kimi K3 | `https://opencode.ai/zen/go/v1` | OpenAI | Toggleable | `xhigh` | `xhigh` (max) | `reasoning_effort` | 262,144 | — | Effort value: `"max"` only. Inherits from `moonshotai/kimi-k3`. 2x usage pricing. |
+| `mimo-v2.5-pro` | MiMo V2.5 Pro | `https://opencode.ai/zen/go/v1` | OpenAI | Always on | N/A | None | None | 1,048,576 | 128,000 | Always-on reasoning. No reasoning options defined in TOML. |
+| `mimo-v2.5` | MiMo V2.5 | `https://opencode.ai/zen/go/v1` | OpenAI | Always on | N/A | None | None | 1,000,000 | 128,000 | Always-on reasoning. No reasoning options defined in TOML. |
+| `minimax-m3` | MiniMax M3 | `https://opencode.ai/zen/go/v1` | Anthropic | Toggleable | `high` | `off`, `high` (on) | `thinking` type (`adaptive` / `disabled`) | 1,000,000 | 131,072 | Mapped toggle: `high` -> `"adaptive"`. Inherits from `minimax/MiniMax-M3`. |
+| `minimax-m2.7` | MiniMax M2.7 | `https://opencode.ai/zen/go/v1` | Anthropic | Always on | N/A | None | None | 204,800 | 131,072 | Always-on reasoning. No reasoning options defined in TOML. |
+| `qwen3.7-max` | Qwen 3.7 Max | `https://opencode.ai/zen/go/v1` | Anthropic | Toggleable | `high` | `off` (disabled), `low` (2048), `medium` (4096), `high` (8192), `xhigh` (16384) | `output_config.task_budget` | 1,000,000 | 65,536 | Supports `toggle` and `budget_tokens` (max 262,144 reasoning tokens). |
+| `qwen3.7-plus` | Qwen 3.7 Plus | `https://opencode.ai/zen/go/v1` | Anthropic | Toggleable | `high` | `off` (disabled), `low` (2048), `medium` (4096), `high` (8192), `xhigh` (16384) | `output_config.task_budget` | 1,000,000 | 65,536 | Supports `toggle` and `budget_tokens` (max 262,144 reasoning tokens). |
+| `qwen3.6-plus` | Qwen 3.6 Plus | `https://opencode.ai/zen/go/v1` | Anthropic | Toggleable | `high` | `off` (disabled), `low` (2048), `medium` (4096), `high` (8192), `xhigh` (8192/16384) | `output_config.task_budget` | 1,000,000 | 65,536 | Supports `toggle` and `budget_tokens` (max 81,920 reasoning tokens). |
+| **Zen Free Models** | | | | | | | | | | |
+| `big-pickle` | Big Pickle | `https://opencode.ai/zen/v1` | OpenAI | Always on | N/A | None | None | 200,000 | 32,000 | Always-on reasoning. Routed to Zen v1. |
+| `deepseek-v4-flash-free` | DeepSeek V4 Flash Free | `https://opencode.ai/zen/v1` | OpenAI | Toggleable | `high` | `off`, `high`, `xhigh` (max) | `reasoning_effort` | 200,000 | 128,000 | Mapped to Zen v1. Effort values: `"high"`, `"max"`. |
+| `hy3-free` | Hy3 Free | `https://opencode.ai/zen/v1` | OpenAI | Always on | N/A | None | None | 256,000 | 64,000 | Always-on reasoning. Routed to Zen v1. |
+| `mimo-v2.5-free` | MiMo V2.5 Free | `https://opencode.ai/zen/v1` | OpenAI | Always on | N/A | None | None | 200,000 | 32,000 | Always-on reasoning. Routed to Zen v1. |
+| `nemotron-3-ultra-free` | Nemotron 3 Ultra Free | `https://opencode.ai/zen/v1` | OpenAI | Always on | N/A | None | None | 1,000,000 | 128,000 | Always-on reasoning. Routed to Zen v1. |
+| `north-mini-code-free` | North Mini Code Free | `https://opencode.ai/zen/v1` | OpenAI | Toggleable | `high` | `off` (none), `high` | `reasoning_effort` | 256,000 | 64,000 | Inherits from `cohere/north-mini-code-1-0`. Effort values: `"none"`, `"high"`. |
+
+## Sources Used
+
+- `anomalyco/models.dev` repository on GitHub (`dev` branch):
+  - [opencode-go/models/](https://github.com/anomalyco/models.dev/tree/dev/providers/opencode-go/models)
+  - [opencode/models/](https://github.com/anomalyco/models.dev/tree/dev/providers/opencode/models)
+  - Compiled [models.json](https://models.dev/models.json) (compiled models.dev catalog registry)
+- `anomalyco/opencode` repository on GitHub (`dev` branch):
+  - [opencode.ai/zen/go/v1](https://github.com/anomalyco/opencode/tree/dev)
+
+## Gaps & Upstream Capabilities Limitations
+
+The following capabilities of OpenCode Go / Zen Free models are currently unsupported in upstream Tau and require local implementations:
+
+### 1. Hybrid Protocols per Provider (Mixed OpenAI & Anthropic API models)
+- **Problem**: Upstream Tau strictly assumes a provider serves models via a single client/API protocol class (either `openai-compatible` or `anthropic`/Messages API). Under a unified aggregator like `opencode-go`, some models use the OpenAI `/chat/completions` schema, while others (like Qwen and Minimax) are served via the Anthropic `/messages` API.
+- **Solution**: Check model-specific metadata configuration `api = "anthropic-messages"` inside `create_model_provider` and dynamically morph the configuration class into `AnthropicProviderConfig` at runtime.
+
+### 2. Toggle-Only Anthropic Thinking (MiniMax M3)
+- **Problem**: Upstream's Anthropic provider handles thinking via either `budget` mode (which expects an integer token count) or `adaptive` mode (which always appends `output_config.effort` to the request payload). Toggle-only models like `minimax-m3` reject the `effort` config payload and require a pure `{ "thinking": { "type": "adaptive" } }` payload.
+- **Solution**: Introduce the `thinking_type: Literal["adaptive", "disabled"]` runtime parameter to bypass the `output_config` payload and send a pure toggle to the endpoint.
+
+### 3. Model-Specific Thinking Level Display Labels
+- **Problem**: Provider-level thinking levels are mapped to standard canonical levels (like `high`, `xhigh`). Some models require mapping these canonical levels to custom UI display labels (e.g., mapping `high` to `"on"` or `"none"` in the CLI/TUI). Upstream does not support model-specific display label overrides.
+- **Solution**: Implement `thinking_level_labels` mapping on the model catalog metadata to translate canonical levels to display strings for CLI/TUI widgets.
+
+### 4. Always-On Reasoning Models
+- **Problem**: Upstream has no native way to declare always-on reasoning models (models where reasoning cannot be toggled off or customized, but thinking parameters must NOT be sent in the API payload to prevent errors). In upstream, excluding a model from the provider's `thinking_models` list disables its reasoning display entirely (rendering it as `"unavailable"` in the UI).
+- **Solution**: Set `reasoning = true` on the model metadata and exclude the model from the provider's `thinking_models` list. Our implementation detects these models by verifying that `reasoning` is `True` while supported `levels` is empty `()`. The TUI and CLI then format these models with the static `"always on"` label and disable custom level adjustments.
+
+### 5. Anthropic-Protocol Model Routing (minimax-m3, minimax-m2.7, qwen3.7-max, qwen3.7-plus, qwen3.6-plus)
+- **Problem**: These models use `@ai-sdk/anthropic` but were routed through the OpenAI-compatible endpoint because the catalog did not specify `api = "anthropic-messages"` in model_metadata.
+- **Solution**: Added `api = "anthropic-messages"` and `thinking_parameter = "anthropic.thinking"` to model_metadata for each Anthropic-protocol model. The runtime now correctly morphs to `AnthropicProvider` with model-specific thinking configuration (budget tokens for Qwen, adaptive toggle for MiniMax M3, no controls for always-on models).
+
+### 6. Model-Specific Thinking Parameter Override
+- **Problem**: Provider-level `thinking_parameter` cannot describe mixed-transport providers where some models use `reasoning_effort` (OpenAI) and others use `anthropic.thinking` (Anthropic).
+- **Solution**: Added `thinking_parameter` field to `ProviderModelMetadata` and `ModelCatalogMetadata`. Model-specific values override the provider-level default when constructing runtime configs.

src/tau_ai/anthropic.py

@@ -362,9 +362,14 @@ def _build_messages_payload(
         ),
         "messages": [_anthropic_message(message) for message in messages],
     }
-    if thinking_mode == "adaptive" and thinking_effort is not None:
+    if thinking_mode == "disabled":
+        payload["thinking"] = {"type": "disabled"}
+    elif thinking_mode == "adaptive" and thinking_effort is not None:
         payload["thinking"] = {"type": "adaptive", "display": "summarized"}
         payload["output_config"] = {"effort": thinking_effort}
+    elif thinking_mode == "adaptive":
+        # Toggle-only model with thinking enabled (no effort field)
+        payload["thinking"] = {"type": "adaptive", "display": "summarized"}
     elif thinking_budget_tokens is not None:
         payload["thinking"] = {
             "type": "enabled",

src/tau_coding/catalog_loader.py

@@ -80,7 +80,9 @@ class _CatalogModelMetadata(BaseModel):
     headers: dict[_NonEmptyString, _NonEmptyString] = {}
     compat: dict[_NonEmptyString, Any] = {}
     thinking_level_map: dict[ThinkingLevel, _NonEmptyString] = {}
+    thinking_level_labels: dict[ThinkingLevel, _NonEmptyString] = {}
     unsupported_thinking_levels: tuple[ThinkingLevel, ...] = ()
+    thinking_parameter: ThinkingParameter | None = None
 
 
 class _CatalogProvider(BaseModel):
@@ -396,6 +398,8 @@ def _model_metadata_from_provider(metadata: _CatalogModelMetadata) -> ModelCatal
         headers=dict(metadata.headers),
         compat=_json_object(metadata.compat, "model_metadata.compat"),
         thinking_level_map=thinking_level_map,
+        thinking_level_labels=dict(metadata.thinking_level_labels),
+        thinking_parameter=metadata.thinking_parameter,
     )
 
 
@@ -523,6 +527,10 @@ def _raw_model_metadata_from_entry(metadata: ModelCatalogMetadata) -> dict[str,
         raw["thinking_level_map"] = thinking_level_map
     if unsupported:
         raw["unsupported_thinking_levels"] = unsupported
+    if metadata.thinking_level_labels:
+        raw["thinking_level_labels"] = dict(metadata.thinking_level_labels)
+    if metadata.thinking_parameter is not None:
+        raw["thinking_parameter"] = metadata.thinking_parameter
     return raw
 
 

src/tau_coding/commands.py

@@ -66,6 +66,9 @@ def context_window_tokens(self) -> int: ...
     @property
     def thinking_level(self) -> str: ...
 
+    @property
+    def thinking_level_label(self) -> str: ...
+
     @property
     def available_thinking_levels(self) -> Sequence[str]: ...
 
@@ -629,6 +632,12 @@ def _thinking_command(context: CommandContext) -> CommandResult:
         return CommandResult(handled=True, message="\n".join(lines))
 
     if not available:
+        always_on = getattr(session, "thinking_is_always_on", False)
+        if always_on:
+            return CommandResult(
+                handled=True,
+                message=f"Thinking mode is always on for {session.provider_name}:{session.model}",
+            )
         message = f"Thinking controls are unavailable for {session.provider_name}:{session.model}"
         reason = _thinking_unavailable_reason(session)
         if reason:
@@ -656,7 +665,9 @@ def _thinking_command(context: CommandContext) -> CommandResult:
 
 def _thinking_status_lines(session: CommandSession) -> list[str]:
     if tuple(session.available_thinking_levels):
-        return [f"Thinking mode: {session.thinking_level}"]
+        return [f"Thinking mode: {session.thinking_level_label}"]
+    if getattr(session, "thinking_is_always_on", False):
+        return ["Thinking mode: always on"]
     lines = ["Thinking mode: unavailable"]
     reason = _thinking_unavailable_reason(session)
     if reason:

src/tau_coding/data/catalog.toml

@@ -6268,28 +6268,169 @@ kind = "openai-compatible"
 base_url = "https://opencode.ai/zen/go/v1"
 api_key_env = "OPENCODE_API_KEY"
 credential_name = "opencode-go"
-models = ["deepseek-v4-flash", "deepseek-v4-pro", "glm-5.1", "glm-5.2", "kimi-k2.6", "kimi-k2.7-code", "mimo-v2.5", "mimo-v2.5-pro", "minimax-m2.7", "minimax-m3", "qwen3.6-plus", "qwen3.7-max", "qwen3.7-plus"]
-default_model = "kimi-k2.7-code"
+models = [
+  "deepseek-v4-pro",
+  "deepseek-v4-flash",
+  "glm-5.2",
+  "glm-5.1",
+  "grok-4.5",
+  "kimi-k2.7-code",
+  "kimi-k2.6",
+  "kimi-k3",
+  "minimax-m3",
+  "minimax-m2.7",
+  "mimo-v2.5-pro",
+  "mimo-v2.5",
+  "qwen3.7-max",
+  "qwen3.7-plus",
+  "qwen3.6-plus",
+  "big-pickle",
+  "deepseek-v4-flash-free",
+  "hy3-free",
+  "mimo-v2.5-free",
+  "nemotron-3-ultra-free",
+  "north-mini-code-free",
+]
+default_model = "deepseek-v4-pro"
 docs_url = "https://opencode.ai/docs/go"
-api = "openai-completions"
-thinking_levels = ["off", "minimal", "low", "medium", "high", "xhigh"]
+thinking_levels = ["off", "low", "medium", "high", "xhigh"]
+thinking_models = [
+  "deepseek-v4-pro",
+  "deepseek-v4-flash",
+  "glm-5.2",
+  "grok-4.5",
+  "kimi-k3",
+  "minimax-m3",
+  "qwen3.7-max",
+  "qwen3.7-plus",
+  "qwen3.6-plus",
+  "deepseek-v4-flash-free",
+  "north-mini-code-free",
+]
 thinking_default = "medium"
 thinking_parameter = "reasoning_effort"
 
 [providers.context_windows]
-"deepseek-v4-flash" = 1000000
-"deepseek-v4-pro" = 1000000
+"glm-5.2" = 1000000
 "glm-5.1" = 202752
-"glm-5.2" = 202752
-"kimi-k2.6" = 262144
 "kimi-k2.7-code" = 262144
-"mimo-v2.5" = 262144
-"mimo-v2.5-pro" = 262144
+"kimi-k2.6" = 262144
+"deepseek-v4-pro" = 1000000
+"deepseek-v4-flash" = 1000000
+"mimo-v2.5" = 1000000
+"mimo-v2.5-pro" = 1048576
+"minimax-m3" = 1000000
 "minimax-m2.7" = 204800
-"minimax-m3" = 204800
-"qwen3.6-plus" = 1000000
 "qwen3.7-max" = 1000000
 "qwen3.7-plus" = 1000000
+"qwen3.6-plus" = 1000000
+"deepseek-v4-flash-free" = 200000
+"mimo-v2.5-free" = 200000
+"big-pickle" = 200000
+"north-mini-code-free" = 256000
+"nemotron-3-ultra-free" = 1000000
+"hy3-free" = 256000
+"kimi-k3" = 262144
+"grok-4.5" = 1000000
+
+[providers.model_metadata."deepseek-v4-flash-free"]
+base_url = "https://opencode.ai/zen/v1"
+thinking_level_map = {high = "high", xhigh = "max"}
+thinking_level_labels = {xhigh = "max"}
+unsupported_thinking_levels = ["minimal", "low", "medium"]
+
+[providers.model_metadata."mimo-v2.5-free"]
+base_url = "https://opencode.ai/zen/v1"
+reasoning = true
+
+[providers.model_metadata."big-pickle"]
+base_url = "https://opencode.ai/zen/v1"
+reasoning = true
+
+[providers.model_metadata."north-mini-code-free"]
+base_url = "https://opencode.ai/zen/v1"
+thinking_level_map = {off = "none", high = "high"}
+unsupported_thinking_levels = ["minimal", "low", "medium", "xhigh"]
+
+[providers.model_metadata."nemotron-3-ultra-free"]
+base_url = "https://opencode.ai/zen/v1"
+reasoning = true
+
+[providers.model_metadata."hy3-free"]
+base_url = "https://opencode.ai/zen/v1"
+reasoning = true
+
+[providers.model_metadata."glm-5.2"]
+thinking_level_map = {high = "high", xhigh = "max"}
+thinking_level_labels = {xhigh = "max"}
+unsupported_thinking_levels = ["off", "minimal", "low", "medium"]
+
+[providers.model_metadata."glm-5.1"]
+reasoning = true
+
+[providers.model_metadata."kimi-k2.7-code"]
+reasoning = true
+
+[providers.model_metadata."kimi-k2.6"]
+reasoning = true
+
+[providers.model_metadata."deepseek-v4-pro"]
+thinking_level_map = {high = "high", xhigh = "max"}
+thinking_level_labels = {xhigh = "max"}
+unsupported_thinking_levels = ["off", "minimal", "low", "medium"]
+
+[providers.model_metadata."deepseek-v4-flash"]
+thinking_level_map = {high = "high", xhigh = "max"}
+thinking_level_labels = {xhigh = "max"}
+unsupported_thinking_levels = ["off", "minimal", "low", "medium"]
+
+[providers.model_metadata."kimi-k3"]
+thinking_level_map = {xhigh = "max"}
+thinking_level_labels = {xhigh = "max"}
+unsupported_thinking_levels = ["off", "minimal", "low", "medium", "high"]
+
+[providers.model_metadata."grok-4.5"]
+thinking_level_map = {off = "none", low = "low", medium = "medium", high = "high"}
+thinking_level_labels = {}
+unsupported_thinking_levels = ["minimal", "xhigh"]
+
+[providers.model_metadata."minimax-m3"]
+api = "anthropic-messages"
+reasoning = true
+thinking_level_map = {off = "disabled", high = "adaptive"}
+thinking_level_labels = {off = "off", high = "on"}
+unsupported_thinking_levels = ["minimal", "low", "medium", "xhigh"]
+compat = {forceAdaptiveThinking = true}
+thinking_parameter = "anthropic.thinking"
+
+[providers.model_metadata."minimax-m2.7"]
+api = "anthropic-messages"
+reasoning = true
+unsupported_thinking_levels = ["off", "minimal", "low", "medium", "high", "xhigh"]
+
+[providers.model_metadata."qwen3.7-max"]
+api = "anthropic-messages"
+reasoning = true
+thinking_level_map = {off = "disabled", low = "low", medium = "medium", high = "high", xhigh = "xhigh"}
+thinking_level_labels = {}
+unsupported_thinking_levels = ["minimal"]
+thinking_parameter = "anthropic.thinking"
+
+[providers.model_metadata."qwen3.7-plus"]
+api = "anthropic-messages"
+reasoning = true
+thinking_level_map = {off = "disabled", low = "low", medium = "medium", high = "high", xhigh = "xhigh"}
+thinking_level_labels = {}
+unsupported_thinking_levels = ["minimal"]
+thinking_parameter = "anthropic.thinking"
+
+[providers.model_metadata."qwen3.6-plus"]
+api = "anthropic-messages"
+reasoning = true
+thinking_level_map = {off = "disabled", low = "low", medium = "medium", high = "high", xhigh = "xhigh"}
+thinking_level_labels = {}
+unsupported_thinking_levels = ["minimal"]
+thinking_parameter = "anthropic.thinking"
 
 [[providers]]
 name = "opencode"

src/tau_coding/provider_catalog.py

@@ -52,6 +52,8 @@ class ModelCatalogMetadata:
     headers: dict[str, str] = field(default_factory=dict)
     compat: dict[str, JSONValue] = field(default_factory=dict)
     thinking_level_map: ThinkingLevelMap = field(default_factory=dict)
+    thinking_level_labels: dict[ThinkingLevel, str] = field(default_factory=dict)
+    thinking_parameter: ThinkingParameter | None = None
 
 
 def model_cost_for_input_tokens(