dostuff 0.2.1__tar.gz → 0.2.3__tar.gz

{dostuff-0.2.1 → dostuff-0.2.3}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: dostuff
-Version: 0.2.1
+Version: 0.2.3
 Summary: DoStuff - a terminal coding agent for any LLM: TUI and headless mode, permissions, MCP, skills, web tools, sessions and git worktrees
 Author-email: Varun Kumawat <varunkumawatleap2@gmail.com>
 License-Expression: MIT
@@ -32,6 +32,7 @@ Requires-Dist: opentelemetry-api>=1.20
 Requires-Dist: opentelemetry-sdk>=1.20
 Requires-Dist: textual>=0.50.0
 Requires-Dist: opentelemetry-exporter-otlp-proto-grpc>=1.20
+Requires-Dist: keyring>=24
 Provides-Extra: memory
 Requires-Dist: chromadb>=1.5.0; extra == "memory"
 Provides-Extra: search
@@ -40,8 +41,6 @@ Provides-Extra: full
 Requires-Dist: dostuff[memory,search]; extra == "full"
 Requires-Dist: onnxruntime>=1.16; extra == "full"
 Requires-Dist: grpcio>=1.52; extra == "full"
-Provides-Extra: keyring
-Requires-Dist: keyring>=24; extra == "keyring"
 Provides-Extra: dev
 Requires-Dist: pytest>=8; extra == "dev"
 Requires-Dist: mypy>=1.10; extra == "dev"
@@ -110,15 +109,19 @@ pip install dostuff
 
 This installs the `dostuff` command on your `PATH`. Done.
 
-Long-term memory (the agent remembers facts and past decisions across sessions) needs ChromaDB,
-a large download, so it's an extra:
+Optional features come as extras:
 
-```bash
-pip install "dostuff[memory]"
-```
+| Install | Adds |
+|---|---|
+| `pip install dostuff` | Everything except the extras below, including MCP (with OAuth tokens kept in the OS keyring) |
+| `pip install "dostuff[memory]"` | Long-term memory (facts and past decisions across sessions) and semantic MCP tool search, via ChromaDB. Large download |
+| `pip install "dostuff[search]"` | Steadier `web_search` through the `ddgs` package (several engines) instead of DuckDuckGo's HTML page |
+| `pip install "dostuff[full]"` | `memory` + `search` |
+| `pip install "dostuff[dev]"` | `pytest` and `mypy`, for working on dostuff itself |
 
-Without it everything else works: the memory tools are hidden from the model and MCP tool search
-matches keywords instead of meaning. `dostuff doctor` shows whether memory is on.
+Extras combine, e.g. `pip install "dostuff[memory,search]"`. Without `memory` everything else
+works: the memory tools are hidden from the model and MCP tool search matches keywords instead of
+meaning. `dostuff doctor` shows whether memory is on.
 
 ### Other install methods
 
@@ -126,6 +129,7 @@ For an isolated install with [pipx](https://pypa.github.io/pipx/):
 
 ```bash
 pipx install dostuff
+pipx install "dostuff[full]"   # with extras
 ```
 
 For development (editable mode, from a clone):
@@ -135,8 +139,8 @@ git clone https://github.com/kVarunkk/DoStuff.git
 cd DoStuff
 pip install -e .           # core
 pip install -e ".[memory]" # + long-term memory (chromadb)
-pip install -e ".[full]"   # + memory + onnxruntime + grpcio
-pip install -e ".[keyring]" # persist MCP OAuth tokens in the OS keyring
+pip install -e ".[search]" # + ddgs web search
+pip install -e ".[full]"   # + memory + search
 pip install -e ".[dev]"    # + pytest + mypy
 
 python -m pytest -q tests
@@ -348,7 +352,7 @@ models:
 loop:
   max_iterations: 100
   subagent_max_iterations: 25
-  # context_window: 128000       # override when litellm doesn't know the model
+  # context_window: 128000       # override when dostuff/litellm don't know the model
   max_tool_output_chars: 30000   # longer tool results keep head + tail
   max_retries: 6                 # retries on rate limits, timeouts, 5xx (0 = none)
   retry_max_delay: 60            # cap on each wait, seconds; Retry-After is honored
@@ -412,6 +416,8 @@ or `/model` inside the TUI. Model lists come from litellm's catalog (chat models
 
 OpenCode models (`opencode/<id>` for Zen, `opencode-go/<id>` for a Go subscription) use the ids from the [OpenCode docs](https://opencode.ai/docs/zen/). litellm has no OpenCode provider, so DoStuff sends each model to the gateway endpoint for its family: Claude models (and a few Qwen / MiniMax models) to the Anthropic Messages endpoint, GPT / Grok / Muse models to the Responses endpoint, Gemini models to the Gemini endpoint, and all other models to chat completions. One console key in `OPENCODE_API_KEY` works for Zen and Go. OpenCode only allows its free models from its own app, so use paid models or Go.
 
+The gateways don't publish a model's context window in `/v1/models`, so `helpers/models/opencode.py` carries one per model, taken from [models.dev](https://models.dev)'s `opencode` and `opencode-go` entries (the same id can differ per gateway: `opencode/minimax-m3` is 512k, `opencode-go/minimax-m3` is 1M). `/context`, the status bar and auto-compaction size themselves from it; set `loop.context_window` to override.
+
 ---
 
 ## MCP servers
@@ -462,7 +468,7 @@ MCP servers connect concurrently in the background, so the TUI never blanks duri
 
 MCP tools are named `<server>__<tool>` (e.g. `github__create_issue`), so two servers can expose a tool with the same name. `search_mcp_tools` only returns tools of servers that are connected right now.
 
-**OAuth (HTTP servers):** a browser login opens when the server answers 401. The flow uses PKCE, a random `state` and an RFC 8707 `resource` indicator; the local callback on port 8085 times out after 120 s. Tokens are kept in memory, or in the OS keyring if you install `dostuff[keyring]`, never in a plain file. A tool call that fails with 401 re-authenticates once and retries.
+**OAuth (HTTP servers):** a browser login opens when the server answers 401. The flow uses PKCE, a random `state` and an RFC 8707 `resource` indicator; the local callback on port 8085 times out after 120 s. Tokens are stored in the OS keyring (memory only if no keyring backend is available), never in a plain file. Expired access tokens are refreshed with the refresh token, before a request or after a 401, without opening the browser; you sign in again only when the refresh token itself is rejected.
 
 To force a re-registration, delete:
 

{dostuff-0.2.1 → dostuff-0.2.3}/README.md

@@ -61,15 +61,19 @@ pip install dostuff
 
 This installs the `dostuff` command on your `PATH`. Done.
 
-Long-term memory (the agent remembers facts and past decisions across sessions) needs ChromaDB,
-a large download, so it's an extra:
+Optional features come as extras:
 
-```bash
-pip install "dostuff[memory]"
-```
+| Install | Adds |
+|---|---|
+| `pip install dostuff` | Everything except the extras below, including MCP (with OAuth tokens kept in the OS keyring) |
+| `pip install "dostuff[memory]"` | Long-term memory (facts and past decisions across sessions) and semantic MCP tool search, via ChromaDB. Large download |
+| `pip install "dostuff[search]"` | Steadier `web_search` through the `ddgs` package (several engines) instead of DuckDuckGo's HTML page |
+| `pip install "dostuff[full]"` | `memory` + `search` |
+| `pip install "dostuff[dev]"` | `pytest` and `mypy`, for working on dostuff itself |
 
-Without it everything else works: the memory tools are hidden from the model and MCP tool search
-matches keywords instead of meaning. `dostuff doctor` shows whether memory is on.
+Extras combine, e.g. `pip install "dostuff[memory,search]"`. Without `memory` everything else
+works: the memory tools are hidden from the model and MCP tool search matches keywords instead of
+meaning. `dostuff doctor` shows whether memory is on.
 
 ### Other install methods
 
@@ -77,6 +81,7 @@ For an isolated install with [pipx](https://pypa.github.io/pipx/):
 
 ```bash
 pipx install dostuff
+pipx install "dostuff[full]"   # with extras
 ```
 
 For development (editable mode, from a clone):
@@ -86,8 +91,8 @@ git clone https://github.com/kVarunkk/DoStuff.git
 cd DoStuff
 pip install -e .           # core
 pip install -e ".[memory]" # + long-term memory (chromadb)
-pip install -e ".[full]"   # + memory + onnxruntime + grpcio
-pip install -e ".[keyring]" # persist MCP OAuth tokens in the OS keyring
+pip install -e ".[search]" # + ddgs web search
+pip install -e ".[full]"   # + memory + search
 pip install -e ".[dev]"    # + pytest + mypy
 
 python -m pytest -q tests
@@ -299,7 +304,7 @@ models:
 loop:
   max_iterations: 100
   subagent_max_iterations: 25
-  # context_window: 128000       # override when litellm doesn't know the model
+  # context_window: 128000       # override when dostuff/litellm don't know the model
   max_tool_output_chars: 30000   # longer tool results keep head + tail
   max_retries: 6                 # retries on rate limits, timeouts, 5xx (0 = none)
   retry_max_delay: 60            # cap on each wait, seconds; Retry-After is honored
@@ -363,6 +368,8 @@ or `/model` inside the TUI. Model lists come from litellm's catalog (chat models
 
 OpenCode models (`opencode/<id>` for Zen, `opencode-go/<id>` for a Go subscription) use the ids from the [OpenCode docs](https://opencode.ai/docs/zen/). litellm has no OpenCode provider, so DoStuff sends each model to the gateway endpoint for its family: Claude models (and a few Qwen / MiniMax models) to the Anthropic Messages endpoint, GPT / Grok / Muse models to the Responses endpoint, Gemini models to the Gemini endpoint, and all other models to chat completions. One console key in `OPENCODE_API_KEY` works for Zen and Go. OpenCode only allows its free models from its own app, so use paid models or Go.
 
+The gateways don't publish a model's context window in `/v1/models`, so `helpers/models/opencode.py` carries one per model, taken from [models.dev](https://models.dev)'s `opencode` and `opencode-go` entries (the same id can differ per gateway: `opencode/minimax-m3` is 512k, `opencode-go/minimax-m3` is 1M). `/context`, the status bar and auto-compaction size themselves from it; set `loop.context_window` to override.
+
 ---
 
 ## MCP servers
@@ -413,7 +420,7 @@ MCP servers connect concurrently in the background, so the TUI never blanks duri
 
 MCP tools are named `<server>__<tool>` (e.g. `github__create_issue`), so two servers can expose a tool with the same name. `search_mcp_tools` only returns tools of servers that are connected right now.
 
-**OAuth (HTTP servers):** a browser login opens when the server answers 401. The flow uses PKCE, a random `state` and an RFC 8707 `resource` indicator; the local callback on port 8085 times out after 120 s. Tokens are kept in memory, or in the OS keyring if you install `dostuff[keyring]`, never in a plain file. A tool call that fails with 401 re-authenticates once and retries.
+**OAuth (HTTP servers):** a browser login opens when the server answers 401. The flow uses PKCE, a random `state` and an RFC 8707 `resource` indicator; the local callback on port 8085 times out after 120 s. Tokens are stored in the OS keyring (memory only if no keyring backend is available), never in a plain file. Expired access tokens are refreshed with the refresh token, before a request or after a 401, without opening the browser; you sign in again only when the refresh token itself is rejected.
 
 To force a re-registration, delete:
 

dostuff-0.2.3/dostuff/helpers/agent/get_model_token_limit.py

@@ -0,0 +1,54 @@
+import litellm
+from dostuff.config import get_config
+from dostuff.helpers.models.opencode import context_window, is_opencode
+from dostuff.lib.model import get_litellm_model, get_model
+
+FALLBACK_CONTEXT_WINDOW = 128_000
+_warned_models: set[str] = set()
+
+
+def _litellm_limit(model: str) -> int | None:
+    """litellm's `max_input_tokens` for `model`, or None when it doesn't map it."""
+    try:
+        limit = litellm.get_model_info(model=model).get("max_input_tokens")
+    except Exception:
+        return None
+    return int(limit) if limit else None
+
+
+def get_model_token_limit() -> int:
+    """Context window for the active model: `loop.context_window`, else the OpenCode table, else
+    litellm's model info, else 128k.
+
+    OpenCode ids are checked against `helpers/models/opencode._CONTEXT_WINDOWS` first: litellm only
+    knows their upstream `openai/` / `anthropic/` route for a few of them, and the same id can have a
+    different window on each gateway.
+    """
+    configured = get_config().context_window
+    if configured > 0:
+        return configured
+    opencode = is_opencode(get_config().provider)
+    if opencode:
+        model_id = get_model().split("/", 1)[-1]
+        known = context_window(get_config().provider, model_id)
+        if known:
+            return known
+    model = get_litellm_model()
+    limit = _litellm_limit(model)
+    if limit:
+        return limit
+    if opencode:
+        # litellm keys some ids only without the route prefix, e.g. `deepseek-v4-pro` (1M) exists
+        # while `openai/deepseek-v4-pro` doesn't.
+        limit = _litellm_limit(get_model().split("/", 1)[-1])
+        if limit:
+            return limit
+    if model not in _warned_models:
+        _warned_models.add(model)
+        from dostuff.helpers.ui.emit import emit
+        emit(
+            f"Unknown context window for '{model}'; assuming {FALLBACK_CONTEXT_WINDOW:,} tokens. "
+            "Set loop.context_window in config.yaml to override.",
+            msg_type="system",
+        )
+    return FALLBACK_CONTEXT_WINDOW

dostuff-0.2.3/dostuff/helpers/mcp/token_store.py

@@ -0,0 +1,81 @@
+"""OAuth token persistence in the OS keyring.
+
+If `keyring` can't be imported or has no usable backend, tokens live in memory only; they are
+never written to a plaintext file.
+
+Windows Credential Manager caps a secret at 2560 bytes (1280 UTF-16 chars) and JWT access plus
+refresh tokens easily exceed that, so long values are split across `<url>#<i>` entries and the
+main entry holds `chunks:<n>`.
+"""
+import json
+from typing import Any
+
+SERVICE = "dostuff-mcp"
+CHUNK_CHARS = 1000
+_CHUNK_PREFIX = "chunks:"
+
+try:
+    import keyring as _keyring
+except ImportError:  # a required dependency, but don't crash a broken install
+    _keyring = None  # type: ignore[assignment]
+
+
+def keyring_available() -> bool:
+    return _keyring is not None
+
+
+def _chunk_count(head: str | None) -> int:
+    if head and head.startswith(_CHUNK_PREFIX):
+        try:
+            return int(head[len(_CHUNK_PREFIX):])
+        except ValueError:
+            return 0
+    return 0
+
+
+def load_token(server_url: str) -> dict[str, Any] | None:
+    if _keyring is None:
+        return None
+    try:
+        raw = _keyring.get_password(SERVICE, server_url)
+        if n := _chunk_count(raw):
+            parts = [_keyring.get_password(SERVICE, f"{server_url}#{i}") for i in range(n)]
+            chunks = [p for p in parts if p is not None]
+            raw = "".join(chunks) if len(chunks) == n else None
+        data = json.loads(raw) if raw else None
+    except Exception:
+        return None
+    return data if isinstance(data, dict) and data.get("access_token") else None
+
+
+def save_token(server_url: str, token: dict[str, Any]) -> bool:
+    """True when the token reached the keyring."""
+    if _keyring is None:
+        return False
+    raw = json.dumps(token)
+    try:
+        delete_token(server_url)
+        if len(raw) <= CHUNK_CHARS:
+            _keyring.set_password(SERVICE, server_url, raw)
+            return True
+        parts = [raw[i:i + CHUNK_CHARS] for i in range(0, len(raw), CHUNK_CHARS)]
+        for i, part in enumerate(parts):
+            _keyring.set_password(SERVICE, f"{server_url}#{i}", part)
+        _keyring.set_password(SERVICE, server_url, f"{_CHUNK_PREFIX}{len(parts)}")  # last: marks it complete
+        return True
+    except Exception:
+        return False  # no usable keyring backend; keep the in-memory token
+
+
+def delete_token(server_url: str) -> None:
+    if _keyring is None:
+        return
+    try:
+        n = _chunk_count(_keyring.get_password(SERVICE, server_url))
+    except Exception:
+        n = 0
+    for name in [server_url, *(f"{server_url}#{i}" for i in range(n))]:
+        try:
+            _keyring.delete_password(SERVICE, name)
+        except Exception:
+            pass

dostuff-0.2.3/dostuff/helpers/models/opencode.py

@@ -0,0 +1,200 @@
+"""OpenCode Zen (`opencode/<id>`) and OpenCode Go (`opencode-go/<id>`) models.
+
+litellm has no OpenCode provider. Each gateway serves a model family on the endpoint of its
+upstream API (https://opencode.ai/docs/zen/), so a model is sent through litellm's openai,
+anthropic or gemini route with the gateway as api_base.
+"""
+import os
+from typing import Any, Literal
+
+OPENCODE_PROVIDERS = {
+    "opencode": "https://opencode.ai/inference",
+    "opencode-go": "https://opencode.ai/inference/go",
+}
+OPENCODE_API_KEY_ENV = "OPENCODE_API_KEY"  # one console key works for Zen and Go
+# Served on the Anthropic Messages endpoint besides Claude; they differ between the gateways
+_ANTHROPIC_IDS = {
+    "opencode": {"qwen3.8-flash", "qwen3.6-plus", "qwen3.5-plus"},
+    "opencode-go": {"qwen3.8-max", "qwen3.8-flash", "qwen3.7-plus", "minimax-m3"},
+}
+# Input-token limit per model, from models.dev's `opencode` / `opencode-go` entries
+# (https://models.dev — the catalog OpenCode's own clients size their context with). The gateways
+# don't publish limits in /v1/models, litellm doesn't map their ids, and the same model can have a
+# different window on each gateway, so this table is the only reliable source. Refresh it when
+# `provider_models()` lists an id that is missing here (those fall back to litellm, then 128k).
+_CONTEXT_WINDOWS = {
+    "opencode": {
+        "beam-free": 1_048_576,
+        "big-pickle": 160_000,
+        "claude-fable-5": 1_000_000,
+        "claude-fable-5-1": 1_000_000,
+        "claude-haiku-4-5": 200_000,
+        "claude-haiku-5-5": 1_000_000,
+        "claude-opus-4-5": 200_000,
+        "claude-opus-4-6": 1_000_000,
+        "claude-opus-4-7": 1_000_000,
+        "claude-opus-4-8": 1_000_000,
+        "claude-opus-5": 1_000_000,
+        "claude-opus-5-5": 1_000_000,
+        "claude-sonnet-4-5": 1_000_000,
+        "claude-sonnet-4-6": 1_000_000,
+        "claude-sonnet-5": 1_000_000,
+        "claude-sonnet-5-5": 1_000_000,
+        "deepseek-v4-pro": 1_000_000,
+        "deepseek-v4.1-flash": 1_000_000,
+        "deepseek-v4.1-flash-free": 200_000,
+        "flint-40b-free": 262_144,
+        "gemini-3-flash": 1_048_576,
+        "gemini-3.1-pro": 1_048_576,
+        "gemini-3.5-flash": 1_048_576,
+        "gemini-3.5-flash-lite": 1_048_576,
+        "gemini-3.6-flash": 1_048_576,
+        "gemini-3.7-flash": 1_048_576,
+        "gemini-3.8-flash": 1_048_576,
+        "gemini-4-argon": 1_048_576,
+        "glm-5": 204_800,
+        "glm-5.1": 204_800,
+        "glm-5.2": 1_000_000,
+        "glm-5.3": 1_000_000,
+        "glm-5.3-flash": 1_000_000,
+        "glm-5.3-flash-free": 200_000,
+        "gpt-5": 272_000,
+        "gpt-5-codex": 272_000,
+        "gpt-5-nano": 272_000,
+        "gpt-5.1": 272_000,
+        "gpt-5.1-codex": 272_000,
+        "gpt-5.1-codex-max": 272_000,
+        "gpt-5.1-codex-mini": 272_000,
+        "gpt-5.2": 272_000,
+        "gpt-5.2-codex": 272_000,
+        "gpt-5.3-codex": 272_000,
+        "gpt-5.3-codex-spark": 128_000,
+        "gpt-5.4": 922_000,
+        "gpt-5.4-mini": 272_000,
+        "gpt-5.4-nano": 272_000,
+        "gpt-5.4-pro": 922_000,
+        "gpt-5.5": 922_000,
+        "gpt-5.5-pro": 922_000,
+        "gpt-5.6-luna": 922_000,
+        "gpt-5.6-sol": 922_000,
+        "gpt-5.6-terra": 922_000,
+        "gpt-6-astra": 922_000,
+        "gpt-6-luna": 922_000,
+        "gpt-6-sol": 922_000,
+        "gpt-6.1-sol": 922_000,
+        "grok-4.5": 500_000,
+        "grok-4.6": 500_000,
+        "grok-4.7": 500_000,
+        "grok-build-0.1": 256_000,
+        "kimi-k2.5": 262_144,
+        "kimi-k2.6": 262_144,
+        "kimi-k2.7-code": 262_144,
+        "kimi-k3": 1_048_576,
+        "mimo-v2.6-flash-free": 200_000,
+        "minimax-m2.5": 204_800,
+        "minimax-m2.7": 204_800,
+        "minimax-m3": 512_000,
+        "mistral-large-4": 524_288,
+        "muse-spark-1.2": 1_048_576,
+        "muse-spark-1.3": 1_048_576,
+        "muse-wonder-1.0": 1_048_576,
+        "nemotron-3.5-lightning-free": 262_144,
+        "nemotron-3.5-super-vl-free": 1_048_576,
+        "qwen3.5-plus": 262_144,
+        "qwen3.6-plus": 262_144,
+        "qwen3.8-flash": 1_000_000,
+        "qwen3.8-max": 262_144,
+    },
+    "opencode-go": {
+        "beam-free": 1_048_576,
+        "claude-haiku-5-5": 1_000_000,
+        "deepseek-v4-pro": 1_000_000,
+        "deepseek-v4.1-flash": 1_000_000,
+        "gemini-4-argon": 1_048_576,
+        "glm-5.2": 1_000_000,
+        "glm-5.3": 1_000_000,
+        "glm-5.3-flash": 1_000_000,
+        "gpt-5.6-luna": 922_000,
+        "gpt-6-luna": 922_000,
+        "grok-4.7": 500_000,
+        "hy3": 192_000,
+        "hy4-preview": 1_024_000,
+        "iceman-contributor-free": 262_144,
+        "kimi-k2.7-code": 262_144,
+        "kimi-k3": 1_048_576,
+        "kimi-k3.1-pro": 1_048_576,
+        "longcat-2.0": 1_000_000,
+        "mimo-v2.6-flash": 1_048_576,
+        "mimo-v2.6-pro": 1_048_576,
+        "minimax-m3": 1_000_000,
+        "minimax-m3.1-flash": 1_000_000,
+        "muse-spark-1.2-contributor": 1_048_576,
+        "muse-spark-1.3-contributor": 1_048_576,
+        "qwen3.7-plus": 1_000_000,
+        "qwen3.8-flash": 1_000_000,
+        "qwen3.8-max": 1_000_000,
+        "step-5-preview": 1_000_000,
+    },
+}
+# In /v1/models but not chat models
+_NOT_CHAT = ("jev-", "test")
+
+Endpoint = Literal["chat", "responses", "anthropic", "google"]
+
+
+def is_opencode(provider: str) -> bool:
+    return provider in OPENCODE_PROVIDERS
+
+
+def context_window(provider: str, model_id: str) -> int | None:
+    """Input-token limit for an OpenCode model id, or None if the table doesn't list it."""
+    return _CONTEXT_WINDOWS.get(provider, {}).get(model_id)
+
+
+def endpoint_for(provider: str, model_id: str) -> Endpoint:
+    """Which upstream API the gateway serves `model_id` on."""
+    if model_id.startswith("claude-") or model_id in _ANTHROPIC_IDS[provider]:
+        return "anthropic"
+    if model_id.startswith(("gpt-", "grok-", "muse-")):
+        return "responses"
+    if model_id.startswith("gemini-"):
+        return "google"
+    return "chat"
+
+
+def litellm_route(provider: str, model_id: str) -> tuple[str, str]:
+    """(litellm model, api_base) for an OpenCode model id."""
+    base = OPENCODE_PROVIDERS[provider]
+    endpoint = endpoint_for(provider, model_id)
+    if endpoint == "anthropic":
+        return f"anthropic/{model_id}", f"{base}/anthropic"  # litellm appends /v1/messages
+    if endpoint == "google":
+        return f"gemini/{model_id}", f"{base}/google/v1beta"
+    if endpoint == "responses":
+        return f"openai/responses/{model_id}", f"{base}/openai/v1"
+    return f"openai/{model_id}", f"{base}/openai/v1"
+
+
+def opencode_kwargs(provider: str, model_id: str, api_key: str | None, session_id: str, version: str) -> dict[str, Any]:
+    """litellm kwargs for an OpenCode model. Go asks clients to name themselves and send a
+    stable per-conversation id; Zen gets the same headers."""
+    model, api_base = litellm_route(provider, model_id)
+    return {
+        "model": model,
+        "api_base": api_base,
+        # Never fall back to OPENAI_API_KEY / ANTHROPIC_API_KEY: those would go to opencode.ai
+        "api_key": api_key or os.environ.get(OPENCODE_API_KEY_ENV) or "missing-opencode-key",
+        "extra_headers": {"User-Agent": f"dostuff/{version}", "x-opencode-session": session_id},
+    }
+
+
+def models_url(provider: str) -> str:
+    return f"{OPENCODE_PROVIDERS[provider]}/v1/models"
+
+
+def chat_model_ids(data: dict) -> list[str]:
+    """Model ids from a /v1/models response, without the non-chat ones."""
+    return [
+        m["id"] for m in data.get("data") or []
+        if isinstance(m, dict) and m.get("id") and not str(m["id"]).startswith(_NOT_CHAT)
+    ]