Merge branch 'NousResearch:main' into main

Author: devorun

.env.example

@@ -59,6 +59,15 @@ OPENCODE_ZEN_API_KEY=
 # OpenCode Go provides access to open models (GLM-5, Kimi K2.5, MiniMax M2.5)
 # $10/month subscription. Get your key at: https://opencode.ai/auth
 OPENCODE_GO_API_KEY=
+
+# =============================================================================
+# LLM PROVIDER (Hugging Face Inference Providers)
+# =============================================================================
+# Hugging Face routes to 20+ open models via unified OpenAI-compatible endpoint.
+# Free tier included ($0.10/month), no markup on provider rates.
+# Get your token at: https://huggingface.co/settings/tokens
+# Required permission: "Make calls to Inference Providers"
+HF_TOKEN=
 # OPENCODE_GO_BASE_URL=https://opencode.ai/zen/go/v1  # Override default base URL
 
 # =============================================================================

agent/anthropic_adapter.py

@@ -35,6 +35,54 @@
     "minimal": "low",
 }
 
+# ── Max output token limits per Anthropic model ───────────────────────
+# Source: Anthropic docs + Cline model catalog.  Anthropic's API requires
+# max_tokens as a mandatory field.  Previously we hardcoded 16384, which
+# starves thinking-enabled models (thinking tokens count toward the limit).
+_ANTHROPIC_OUTPUT_LIMITS = {
+    # Claude 4.6
+    "claude-opus-4-6":   128_000,
+    "claude-sonnet-4-6":  64_000,
+    # Claude 4.5
+    "claude-opus-4-5":    64_000,
+    "claude-sonnet-4-5":  64_000,
+    "claude-haiku-4-5":   64_000,
+    # Claude 4
+    "claude-opus-4":      32_000,
+    "claude-sonnet-4":    64_000,
+    # Claude 3.7
+    "claude-3-7-sonnet": 128_000,
+    # Claude 3.5
+    "claude-3-5-sonnet":   8_192,
+    "claude-3-5-haiku":    8_192,
+    # Claude 3
+    "claude-3-opus":       4_096,
+    "claude-3-sonnet":     4_096,
+    "claude-3-haiku":      4_096,
+}
+
+# For any model not in the table, assume the highest current limit.
+# Future Anthropic models are unlikely to have *less* output capacity.
+_ANTHROPIC_DEFAULT_OUTPUT_LIMIT = 128_000
+
+
+def _get_anthropic_max_output(model: str) -> int:
+    """Look up the max output token limit for an Anthropic model.
+
+    Uses substring matching against _ANTHROPIC_OUTPUT_LIMITS so date-stamped
+    model IDs (claude-sonnet-4-5-20250929) and variant suffixes (:1m, :fast)
+    resolve correctly.  Longest-prefix match wins to avoid e.g. "claude-3-5"
+    matching before "claude-3-5-sonnet".
+    """
+    m = model.lower()
+    best_key = ""
+    best_val = _ANTHROPIC_DEFAULT_OUTPUT_LIMIT
+    for key, val in _ANTHROPIC_OUTPUT_LIMITS.items():
+        if key in m and len(key) > len(best_key):
+            best_key = key
+            best_val = val
+    return best_val
+
 
 def _supports_adaptive_thinking(model: str) -> bool:
     """Return True for Claude 4.6 models that support adaptive thinking."""
@@ -59,6 +107,7 @@ def _supports_adaptive_thinking(model: str) -> bool:
 # The version must stay reasonably current — Anthropic rejects OAuth requests
 # when the spoofed user-agent version is too far behind the actual release.
 _CLAUDE_CODE_VERSION_FALLBACK = "2.1.74"
+_claude_code_version_cache: Optional[str] = None
 
 
 def _detect_claude_code_version() -> str:
@@ -86,11 +135,18 @@ def _detect_claude_code_version() -> str:
     return _CLAUDE_CODE_VERSION_FALLBACK
 
 
-_CLAUDE_CODE_VERSION = _detect_claude_code_version()
 _CLAUDE_CODE_SYSTEM_PREFIX = "You are Claude Code, Anthropic's official CLI for Claude."
 _MCP_TOOL_PREFIX = "mcp_"
 
 
+def _get_claude_code_version() -> str:
+    """Lazily detect the installed Claude Code version when OAuth headers need it."""
+    global _claude_code_version_cache
+    if _claude_code_version_cache is None:
+        _claude_code_version_cache = _detect_claude_code_version()
+    return _claude_code_version_cache
+
+
 def _is_oauth_token(key: str) -> bool:
     """Check if the key is an OAuth/setup token (not a regular Console API key).
 
@@ -132,7 +188,7 @@ def build_anthropic_client(api_key: str, base_url: str = None):
         kwargs["auth_token"] = api_key
         kwargs["default_headers"] = {
             "anthropic-beta": ",".join(all_betas),
-            "user-agent": f"claude-cli/{_CLAUDE_CODE_VERSION} (external, cli)",
+            "user-agent": f"claude-cli/{_get_claude_code_version()} (external, cli)",
             "x-app": "cli",
         }
     else:
@@ -241,7 +297,7 @@ def _refresh_oauth_token(creds: Dict[str, Any]) -> Optional[str]:
 
     headers = {
         "Content-Type": "application/json",
-        "User-Agent": f"claude-cli/{_CLAUDE_CODE_VERSION} (external, cli)",
+        "User-Agent": f"claude-cli/{_get_claude_code_version()} (external, cli)",
     }
 
     for endpoint in token_endpoints:
@@ -810,9 +866,15 @@ def build_anthropic_kwargs(
     tool_choice: Optional[str] = None,
     is_oauth: bool = False,
     preserve_dots: bool = False,
+    context_length: Optional[int] = None,
 ) -> Dict[str, Any]:
     """Build kwargs for anthropic.messages.create().
 
+    When *max_tokens* is None, the model's native output limit is used
+    (e.g. 128K for Opus 4.6, 64K for Sonnet 4.6).  If *context_length*
+    is provided, the effective limit is clamped so it doesn't exceed
+    the context window.
+
     When *is_oauth* is True, applies Claude Code compatibility transforms:
     system prompt prefix, tool name prefixing, and prompt sanitization.
 
@@ -823,7 +885,12 @@ def build_anthropic_kwargs(
     anthropic_tools = convert_tools_to_anthropic(tools) if tools else []
 
     model = normalize_model_name(model, preserve_dots=preserve_dots)
-    effective_max_tokens = max_tokens or 16384
+    effective_max_tokens = max_tokens or _get_anthropic_max_output(model)
+
+    # Clamp to context window if the user set a lower context_length
+    # (e.g. custom endpoint with limited capacity).
+    if context_length and effective_max_tokens > context_length:
+        effective_max_tokens = max(context_length - 1, 1)
 
     # ── OAuth: Claude Code identity ──────────────────────────────────
     if is_oauth:

agent/auxiliary_client.py

@@ -1137,7 +1137,13 @@ def _finalize(resolved_provider: str, sync_client: Any, default_model: Optional[
         return "custom", client, final_model
 
     if requested == "auto":
-        for candidate in get_available_vision_backends():
+        ordered = list(_VISION_AUTO_PROVIDER_ORDER)
+        preferred = _preferred_main_vision_provider()
+        if preferred in ordered:
+            ordered.remove(preferred)
+            ordered.insert(0, preferred)
+
+        for candidate in ordered:
             sync_client, default_model = _resolve_strict_vision_backend(candidate)
             if sync_client is not None:
                 return _finalize(candidate, sync_client, default_model)
@@ -1210,6 +1216,39 @@ def auxiliary_max_tokens_param(value: int) -> dict:
 _client_cache_lock = threading.Lock()
 
 
+def neuter_async_httpx_del() -> None:
+    """Monkey-patch ``AsyncHttpxClientWrapper.__del__`` to be a no-op.
+
+    The OpenAI SDK's ``AsyncHttpxClientWrapper.__del__`` schedules
+    ``self.aclose()`` via ``asyncio.get_running_loop().create_task()``.
+    When an ``AsyncOpenAI`` client is garbage-collected while
+    prompt_toolkit's event loop is running (the common CLI idle state),
+    the ``aclose()`` task runs on prompt_toolkit's loop but the
+    underlying TCP transport is bound to a *different* loop (the worker
+    thread's loop that the client was originally created on).  If that
+    loop is closed or its thread is dead, the transport's
+    ``self._loop.call_soon()`` raises ``RuntimeError("Event loop is
+    closed")``, which prompt_toolkit surfaces as "Unhandled exception
+    in event loop ... Press ENTER to continue...".
+
+    Neutering ``__del__`` is safe because:
+    - Cached clients are explicitly cleaned via ``_force_close_async_httpx``
+      on stale-loop detection and ``shutdown_cached_clients`` on exit.
+    - Uncached clients' TCP connections are cleaned up by the OS when the
+      process exits.
+    - The OpenAI SDK itself marks this as a TODO (``# TODO(someday):
+      support non asyncio runtimes here``).
+
+    Call this once at CLI startup, before any ``AsyncOpenAI`` clients are
+    created.
+    """
+    try:
+        from openai._base_client import AsyncHttpxClientWrapper
+        AsyncHttpxClientWrapper.__del__ = lambda self: None  # type: ignore[assignment]
+    except (ImportError, AttributeError):
+        pass  # Graceful degradation if the SDK changes its internals
+
+
 def _force_close_async_httpx(client: Any) -> None:
     """Mark the httpx AsyncClient inside an AsyncOpenAI client as closed.
 
@@ -1257,6 +1296,25 @@ def shutdown_cached_clients() -> None:
         _client_cache.clear()
 
 
+def cleanup_stale_async_clients() -> None:
+    """Force-close cached async clients whose event loop is closed.
+
+    Call this after each agent turn to proactively clean up stale clients
+    before GC can trigger ``AsyncHttpxClientWrapper.__del__`` on them.
+    This is defense-in-depth — the primary fix is ``neuter_async_httpx_del``
+    which disables ``__del__`` entirely.
+    """
+    with _client_cache_lock:
+        stale_keys = []
+        for key, entry in _client_cache.items():
+            client, _default, cached_loop = entry
+            if cached_loop is not None and cached_loop.is_closed():
+                _force_close_async_httpx(client)
+                stale_keys.append(key)
+        for key in stale_keys:
+            del _client_cache[key]
+
+
 def _get_cached_client(
     provider: str,
     model: str = None,
@@ -1558,6 +1616,62 @@ def call_llm(
         raise
 
 
+def extract_content_or_reasoning(response) -> str:
+    """Extract content from an LLM response, falling back to reasoning fields.
+
+    Mirrors the main agent loop's behavior when a reasoning model (DeepSeek-R1,
+    Qwen-QwQ, etc.) returns ``content=None`` with reasoning in structured fields.
+
+    Resolution order:
+      1. ``message.content`` — strip inline think/reasoning blocks, check for
+         remaining non-whitespace text.
+      2. ``message.reasoning`` / ``message.reasoning_content`` — direct
+         structured reasoning fields (DeepSeek, Moonshot, Novita, etc.).
+      3. ``message.reasoning_details`` — OpenRouter unified array format.
+
+    Returns the best available text, or ``""`` if nothing found.
+    """
+    import re
+
+    msg = response.choices[0].message
+    content = (msg.content or "").strip()
+
+    if content:
+        # Strip inline think/reasoning blocks (mirrors _strip_think_blocks)
+        cleaned = re.sub(
+            r"<(?:think|thinking|reasoning|REASONING_SCRATCHPAD)>"
+            r".*?"
+            r"</(?:think|thinking|reasoning|REASONING_SCRATCHPAD)>",
+            "", content, flags=re.DOTALL | re.IGNORECASE,
+        ).strip()
+        if cleaned:
+            return cleaned
+
+    # Content is empty or reasoning-only — try structured reasoning fields
+    reasoning_parts: list[str] = []
+    for field in ("reasoning", "reasoning_content"):
+        val = getattr(msg, field, None)
+        if val and isinstance(val, str) and val.strip() and val not in reasoning_parts:
+            reasoning_parts.append(val.strip())
+
+    details = getattr(msg, "reasoning_details", None)
+    if details and isinstance(details, list):
+        for detail in details:
+            if isinstance(detail, dict):
+                summary = (
+                    detail.get("summary")
+                    or detail.get("content")
+                    or detail.get("text")
+                )
+                if summary and summary not in reasoning_parts:
+                    reasoning_parts.append(summary.strip() if isinstance(summary, str) else str(summary))
+
+    if reasoning_parts:
+        return "\n\n".join(reasoning_parts)
+
+    return ""
+
+
 async def async_call_llm(
     task: str = None,
     *,

agent/context_references.py

@@ -286,12 +286,16 @@ def _expand_git_reference(
     args: list[str],
     label: str,
 ) -> tuple[str | None, str | None]:
-    result = subprocess.run(
-        ["git", *args],
-        cwd=cwd,
-        capture_output=True,
-        text=True,
-    )
+    try:
+        result = subprocess.run(
+            ["git", *args],
+            cwd=cwd,
+            capture_output=True,
+            text=True,
+            timeout=30,
+        )
+    except subprocess.TimeoutExpired:
+        return f"{ref.raw}: git command timed out (30s)", None
     if result.returncode != 0:
         stderr = (result.stderr or "").strip() or "git command failed"
         return f"{ref.raw}: {stderr}", None
@@ -449,9 +453,12 @@ def _rg_files(path: Path, cwd: Path, limit: int) -> list[Path] | None:
             cwd=cwd,
             capture_output=True,
             text=True,
+            timeout=10,
         )
     except FileNotFoundError:
         return None
+    except subprocess.TimeoutExpired:
+        return None
     if result.returncode != 0:
         return None
     files = [Path(line.strip()) for line in result.stdout.splitlines() if line.strip()]

agent/display.py

@@ -699,7 +699,7 @@ def format_context_pressure(
         threshold_percent: Compaction threshold as a fraction of context window.
         compression_enabled: Whether auto-compression is active.
     """
-    pct_int = int(compaction_progress * 100)
+    pct_int = min(int(compaction_progress * 100), 100)
     filled = min(int(compaction_progress * _BAR_WIDTH), _BAR_WIDTH)
     bar = _BAR_FILLED * filled + _BAR_EMPTY * (_BAR_WIDTH - filled)
 
@@ -729,7 +729,7 @@ def format_context_pressure_gateway(
     No ANSI — just Unicode and plain text suitable for Telegram/Discord/etc.
     The percentage shows progress toward the compaction threshold.
     """
-    pct_int = int(compaction_progress * 100)
+    pct_int = min(int(compaction_progress * 100), 100)
     filled = min(int(compaction_progress * _BAR_WIDTH), _BAR_WIDTH)
     bar = _BAR_FILLED * filled + _BAR_EMPTY * (_BAR_WIDTH - filled)
 

agent/model_metadata.py

@@ -113,6 +113,15 @@ def _strip_provider_prefix(model: str) -> str:
     "glm": 202752,
     # Kimi
     "kimi": 262144,
+    # Hugging Face Inference Providers — model IDs use org/name format
+    "Qwen/Qwen3.5-397B-A17B": 131072,
+    "Qwen/Qwen3.5-35B-A3B": 131072,
+    "deepseek-ai/DeepSeek-V3.2": 65536,
+    "moonshotai/Kimi-K2.5": 262144,
+    "moonshotai/Kimi-K2-Thinking": 262144,
+    "MiniMaxAI/MiniMax-M2.5": 204800,
+    "XiaomiMiMo/MiMo-V2-Flash": 32768,
+    "zai-org/GLM-5": 202752,
 }
 
 _CONTEXT_LENGTH_KEYS = (