tau-coding-agent 0.6.9__tar.gz → 0.7.1__tar.gz

{tau_coding_agent-0.6.9/tau_coding_agent.egg-info → tau_coding_agent-0.7.1}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: tau-coding-agent
-Version: 0.6.9
+Version: 0.7.1
 Summary: A self-extensible agent CLI with terminal UI, multi-provider LLM support, session management with branching, and a plugin system for tools and commands
 Requires-Python: >=3.12
 Description-Content-Type: text/markdown

{tau_coding_agent-0.6.9 → tau_coding_agent-0.7.1}/pyproject.toml

@@ -4,7 +4,7 @@ build-backend = "setuptools.build_meta"
 
 [project]
 name = "tau-coding-agent"
-version = "0.6.9"
+version = "0.7.1"
 description = "A self-extensible agent CLI with terminal UI, multi-provider LLM support, session management with branching, and a plugin system for tools and commands"
 readme = "README.md"
 requires-python = ">=3.12"

{tau_coding_agent-0.6.9 → tau_coding_agent-0.7.1}/tau/builtins/extensions/sandbox/__init__.py

@@ -74,7 +74,14 @@ def register(tau: ExtensionAPI) -> None:
             asyncio.ensure_future(_prewarm())
 
     @tau.on("session_shutdown")
-    async def _on_shutdown(_event: Any, _ctx: ExtensionContext) -> None:
+    async def _on_session_shutdown(_event: Any, _ctx: ExtensionContext) -> None:
+        await manager.stop()
+
+    @tau.on("runtime_stop")
+    async def _on_runtime_stop(_event: Any, _ctx: ExtensionContext) -> None:
+        # session_shutdown only fires on session transitions (new/resume/clone),
+        # not on actual process exit, so the microVM would otherwise keep running
+        # until its own idle_timeout — reap it here too, same as the lsp extension.
         await manager.stop()
 
     async def cmd_sandbox(ctx: ExtensionContext, args: list[str]) -> None:

{tau_coding_agent-0.6.9 → tau_coding_agent-0.7.1}/tau/builtins/models/text.py

@@ -75,13 +75,33 @@ models = [
         output=_TEXT,
     ),
     # Anthropic Claude Code (OAuth)
+    Model(
+        id="claude-fable-5",
+        name="Claude Fable 5",
+        provider="anthropic-claude-code",
+        cost=Cost(),
+        thinking=True,
+        context_window=1_048_576,
+        input=_TEXT_IMAGE,
+        output=_TEXT,
+    ),
+    Model(
+        id="claude-opus-4-8",
+        name="Claude Opus 4.8",
+        provider="anthropic-claude-code",
+        cost=Cost(),
+        thinking=True,
+        context_window=1_048_576,
+        input=_TEXT_IMAGE,
+        output=_TEXT,
+    ),
     Model(
         id="claude-opus-4-7",
         name="Claude Opus 4.7",
         provider="anthropic-claude-code",
         cost=Cost(),
         thinking=True,
-        context_window=200_000,
+        context_window=1_048_576,
         input=_TEXT_IMAGE,
         output=_TEXT,
     ),
@@ -91,7 +111,7 @@ models = [
         provider="anthropic-claude-code",
         cost=Cost(),
         thinking=True,
-        context_window=1_000_000,
+        context_window=1_048_576,
         input=_TEXT_IMAGE,
         output=_TEXT,
     ),
@@ -101,6 +121,16 @@ models = [
         provider="anthropic-claude-code",
         cost=Cost(),
         thinking=True,
+        context_window=1_048_576,
+        input=_TEXT_IMAGE,
+        output=_TEXT,
+    ),
+    Model(
+        id="claude-sonnet-4-5-20250929",
+        name="Claude Sonnet 4.5",
+        provider="anthropic-claude-code",
+        cost=Cost(),
+        thinking=True,
         context_window=200_000,
         input=_TEXT_IMAGE,
         output=_TEXT,
@@ -122,26 +152,29 @@ models = [
         cost=Cost(),
         thinking=True,
         context_window=1_048_576,
+        max_output_tokens=65_536,
         input=_TEXT_IMAGE_VIDEO,
         output=_TEXT,
     ),
     Model(
-        id="gemini-3.1-pro-preview",
-        name="Gemini 3.1 Pro",
+        id="gemini-3.1-pro-high",
+        name="Gemini 3.1 Pro (High)",
         provider="google-antigravity",
         cost=Cost(),
         thinking=True,
         context_window=1_048_576,
+        max_output_tokens=65_535,
         input=_TEXT_IMAGE_VIDEO,
         output=_TEXT,
     ),
     Model(
-        id="gemini-3-flash-preview",
-        name="Gemini 3 Flash Preview",
+        id="gemini-3-flash-agent",
+        name="Gemini 3.5 Flash (High)",
         provider="google-antigravity",
         cost=Cost(),
         thinking=True,
         context_window=1_048_576,
+        max_output_tokens=65_536,
         input=_TEXT_IMAGE_VIDEO,
         output=_TEXT,
     ),
@@ -151,6 +184,7 @@ models = [
         provider="google-antigravity",
         cost=Cost(),
         context_window=1_048_576,
+        max_output_tokens=65_535,
         input=_TEXT_IMAGE_VIDEO,
         output=_TEXT,
     ),
@@ -161,6 +195,7 @@ models = [
         cost=Cost(),
         thinking=True,
         context_window=1_048_576,
+        max_output_tokens=65_535,
         input=_TEXT_IMAGE_VIDEO,
         output=_TEXT,
     ),
@@ -171,6 +206,7 @@ models = [
         cost=Cost(),
         thinking=True,
         context_window=1_048_576,
+        max_output_tokens=65_535,
         input=_TEXT_IMAGE_VIDEO,
         output=_TEXT,
     ),
@@ -180,36 +216,41 @@ models = [
         provider="google-antigravity",
         cost=Cost(),
         context_window=1_048_576,
+        max_output_tokens=65_535,
         input=_TEXT_IMAGE_VIDEO,
         output=_TEXT,
     ),
     Model(
-        id="claude-sonnet-4-20250514",
+        id="claude-sonnet-4-6",
         name="Claude Sonnet 4.6 (Thinking)",
         provider="google-antigravity",
         cost=Cost(),
         thinking=True,
-        context_window=200_000,
+        context_window=250_000,
+        max_output_tokens=64_000,
         input=_TEXT_IMAGE,
         output=_TEXT,
     ),
     Model(
-        id="claude-opus-4-20250514",
+        id="claude-opus-4-6-thinking",
         name="Claude Opus 4.6 (Thinking)",
         provider="google-antigravity",
         cost=Cost(),
         thinking=True,
-        context_window=200_000,
+        context_window=250_000,
+        max_output_tokens=64_000,
         input=_TEXT_IMAGE,
         output=_TEXT,
     ),
     Model(
-        id="gpt-oss-120",
-        name="GPT-OSS 120",
+        id="gpt-oss-120b-medium",
+        name="GPT-OSS 120B (Medium)",
         provider="google-antigravity",
         cost=Cost(),
-        context_window=128_000,
-        input=_TEXT_IMAGE,
+        thinking=True,
+        context_window=131_072,
+        max_output_tokens=32_768,
+        input=_TEXT,
         output=_TEXT,
     ),
     # Google AI Studio (Gemini Developer API — API key)
@@ -370,7 +411,7 @@ models = [
         id="claude-fable-5",
         name="Claude Fable 5",
         provider="anthropic-vertex",
-        cost=Cost(input=5.0, output=25.0, cache_read=0.5),
+        cost=Cost(input=10.0, output=50.0, cache_read=1.0),
         thinking=True,
         context_window=1_048_576,
         input=_TEXT_IMAGE,
@@ -380,7 +421,7 @@ models = [
         id="claude-opus-4-8",
         name="Claude Opus 4.8",
         provider="anthropic-vertex",
-        cost=Cost(input=15.0, output=75.0, cache_read=1.5),
+        cost=Cost(input=5.0, output=25.0, cache_read=0.5),
         thinking=True,
         context_window=1_048_576,
         input=_TEXT_IMAGE,
@@ -390,7 +431,7 @@ models = [
         id="claude-opus-4-7",
         name="Claude Opus 4.7",
         provider="anthropic-vertex",
-        cost=Cost(input=15.0, output=75.0, cache_read=1.5),
+        cost=Cost(input=5.0, output=25.0, cache_read=0.5),
         thinking=True,
         context_window=1_048_576,
         input=_TEXT_IMAGE,
@@ -400,7 +441,7 @@ models = [
         id="claude-opus-4-6",
         name="Claude Opus 4.6",
         provider="anthropic-vertex",
-        cost=Cost(input=15.0, output=75.0, cache_read=1.5),
+        cost=Cost(input=5.0, output=25.0, cache_read=0.5),
         thinking=True,
         context_window=1_048_576,
         input=_TEXT_IMAGE,
@@ -430,7 +471,7 @@ models = [
         id="claude-opus-4-5@20251101",
         name="Claude Opus 4.5",
         provider="anthropic-vertex",
-        cost=Cost(input=15.0, output=75.0, cache_read=1.5),
+        cost=Cost(input=5.0, output=25.0, cache_read=0.5),
         thinking=True,
         context_window=200_000,
         input=_TEXT_IMAGE,
@@ -449,7 +490,7 @@ models = [
         id="claude-haiku-4-5@20251001",
         name="Claude Haiku 4.5",
         provider="anthropic-vertex",
-        cost=Cost(input=0.8, output=4.0, cache_read=0.08),
+        cost=Cost(input=1.0, output=5.0, cache_read=0.10),
         context_window=200_000,
         input=_TEXT_IMAGE,
         output=_TEXT,
@@ -620,13 +661,33 @@ models = [
         output=_TEXT,
     ),
     # Anthropic (API key)
+    Model(
+        id="claude-fable-5",
+        name="Claude Fable 5",
+        provider="anthropic",
+        cost=Cost(input=10.0, output=50.0, cache_read=1.0, cache_write=12.5),
+        thinking=True,
+        context_window=1_048_576,
+        input=_TEXT_IMAGE,
+        output=_TEXT,
+    ),
+    Model(
+        id="claude-opus-4-8",
+        name="Claude Opus 4.8",
+        provider="anthropic",
+        cost=Cost(input=5.0, output=25.0, cache_read=0.5, cache_write=6.25),
+        thinking=True,
+        context_window=1_048_576,
+        input=_TEXT_IMAGE,
+        output=_TEXT,
+    ),
     Model(
         id="claude-opus-4-7",
         name="Claude Opus 4.7",
         provider="anthropic",
-        cost=Cost(input=15.0, output=75.0, cache_read=1.5, cache_write=3.75),
+        cost=Cost(input=5.0, output=25.0, cache_read=0.5, cache_write=6.25),
         thinking=True,
-        context_window=200_000,
+        context_window=1_048_576,
         input=_TEXT_IMAGE,
         output=_TEXT,
     ),
@@ -634,9 +695,10 @@ models = [
         id="claude-sonnet-5",
         name="Claude Sonnet 5",
         provider="anthropic",
-        cost=Cost(input=3.0, output=15.0, cache_read=0.30, cache_write=3.75),
+        # Introductory pricing through 2026-08-31; standard rate is $3/$15 after.
+        cost=Cost(input=2.0, output=10.0, cache_read=0.20, cache_write=2.5),
         thinking=True,
-        context_window=1_000_000,
+        context_window=1_048_576,
         input=_TEXT_IMAGE,
         output=_TEXT,
     ),
@@ -646,7 +708,7 @@ models = [
         provider="anthropic",
         cost=Cost(input=3.0, output=15.0, cache_read=0.30, cache_write=3.75),
         thinking=True,
-        context_window=200_000,
+        context_window=1_048_576,
         input=_TEXT_IMAGE,
         output=_TEXT,
     ),
@@ -654,7 +716,7 @@ models = [
         id="claude-haiku-4-5-20251001",
         name="Claude Haiku 4.5",
         provider="anthropic",
-        cost=Cost(input=0.80, output=4.0, cache_read=0.08, cache_write=1.0),
+        cost=Cost(input=1.0, output=5.0, cache_read=0.10, cache_write=1.25),
         context_window=200_000,
         input=_TEXT_IMAGE,
         output=_TEXT,

{tau_coding_agent-0.6.9 → tau_coding_agent-0.7.1}/tau/console/commands/update.py

@@ -2,11 +2,17 @@ from __future__ import annotations
 
 import asyncio
 import sys
+import threading
+from collections.abc import Callable, Iterator
+from contextlib import contextmanager
 from pathlib import Path
 from typing import TYPE_CHECKING
 
 import click
 
+from tau.tui.style import RESET, Style
+from tau.tui.widgets.symbols import FILL_HORIZONTAL
+
 if TYPE_CHECKING:
     from tau.packages.manager import PackageManager
     from tau.settings.manager import SettingsManager
@@ -88,7 +94,6 @@ def _update_tau() -> None:
     from tau.settings.paths import get_app_name, get_package_name
 
     app = get_package_name()
-    click.echo(f"Updating {get_app_name()}…")
 
     # Pick the upgrade tool that matches how this copy was installed, inferred
     # from the venv it runs in, so we upgrade the right managed environment.
@@ -103,8 +108,58 @@ def _update_tau() -> None:
     else:
         cmd = [sys.executable, "-m", "pip", "install", "--upgrade", app]
 
-    result = subprocess.run(cmd, capture_output=True, text=True)
+    with _progress_bar(f"Updating {get_app_name()}…"):
+        result = subprocess.run(cmd, capture_output=True, text=True)
+
     if result.returncode == 0:
         click.echo(click.style(f"✓ {get_app_name()} updated successfully", fg="green"))
     else:
         raise click.ClickException(result.stderr.strip() or "Update failed.")
+
+
+# Width of the sweep track and the highlighted window within it, in cells.
+_BAR_WIDTH = 24
+_SWEEP_WIDTH = 8
+_FRAME_INTERVAL = 0.08
+
+
+def _sweep_frame(frame: int) -> str:
+    """Render one frame of an indeterminate progress bar: a filled window that
+    bounces back and forth across the track, built from the same fill glyph
+    and Style the tui's own ``LineGauge`` widget uses (tau/tui/widgets/gauge.py)."""
+    track_len = max(1, _BAR_WIDTH - _SWEEP_WIDTH)
+    period = track_len * 2
+    t = frame % period
+    pos = t if t <= track_len else period - t
+    fill = FILL_HORIZONTAL[-1]
+    cells = [fill if pos <= i < pos + _SWEEP_WIDTH else " " for i in range(_BAR_WIDTH)]
+    style = Style().with_fg("bright_cyan")
+    return f"[{style.sgr()}{''.join(cells)}{RESET}]"
+
+
+@contextmanager
+def _progress_bar(message: str) -> Iterator[Callable[[], None]]:
+    """Show an indeterminate sweep bar on stderr while a blocking call runs.
+
+    ``subprocess.run(capture_output=True)`` gives no feedback until it
+    returns, which can be tens of seconds for an installer download — animate
+    on a background thread so the terminal doesn't look hung, then erase it
+    in favor of the caller's own result line.
+    """
+    stop = threading.Event()
+
+    def _animate() -> None:
+        frame = 0
+        while not stop.is_set():
+            click.echo(f"\r{_sweep_frame(frame)} {message}", nl=False, err=True)
+            frame += 1
+            stop.wait(_FRAME_INTERVAL)
+        click.echo("\r" + " " * (_BAR_WIDTH + len(message) + 3) + "\r", nl=False, err=True)
+
+    thread = threading.Thread(target=_animate, daemon=True)
+    thread.start()
+    try:
+        yield stop.set
+    finally:
+        stop.set()
+        thread.join()

{tau_coding_agent-0.6.9 → tau_coding_agent-0.7.1}/tau/inference/api/text/anthropic_claude_code.py

@@ -1,5 +1,7 @@
 from __future__ import annotations
 
+import hashlib
+import os
 from collections.abc import AsyncGenerator
 from typing import TYPE_CHECKING, Any
 
@@ -55,11 +57,127 @@ _DEFAULT_MAX_TOKENS = 8096
 
 
 _OAUTH_HEADERS = {
-    "anthropic-beta": "oauth-2025-04-20",
     "x-app": "cli",
     "User-Agent": "claude-cli/2.1.122 (external, sdk-cli)",
 }
 
+# System identity that Anthropic's API requires as its own system[] entry for
+# OAuth-authenticated (Claude Pro/Max) requests to be accepted and billed
+# against the subscription rather than rejected with a 400.
+_SYSTEM_IDENTITY = "You are Claude Code, Anthropic's official CLI for Claude."
+
+_CC_VERSION = "2.1.185"
+_BILLING_SALT = "59cf53e54c78"
+
+# Beta flags Claude Code itself sends on every OAuth request.
+_BASE_BETAS = [
+    "claude-code-20250219",
+    "oauth-2025-04-20",
+    "interleaved-thinking-2025-05-14",
+    "prompt-caching-scope-2026-01-05",
+    "context-management-2025-06-27",
+    "advisor-tool-2026-03-01",
+    "thinking-token-count-2026-05-13",
+    "extended-cache-ttl-2025-04-11",
+    "effort-2025-11-24",
+]
+
+# Per-model beta overrides, matched by substring against the lowercased model id.
+_MODEL_BETA_OVERRIDES: dict[str, dict[str, list[str]]] = {
+    "haiku": {"exclude": ["interleaved-thinking-2025-05-14"]},
+    "4-6": {"add": ["effort-2025-11-24"]},
+    "4-7": {"add": ["effort-2025-11-24"]},
+}
+
+
+def _model_betas(model_id: str) -> list[str]:
+    """Return the anthropic-beta flags for `model_id`, applying per-model overrides."""
+    betas = list(_BASE_BETAS)
+    lower = model_id.lower()
+    for pattern, override in _MODEL_BETA_OVERRIDES.items():
+        if pattern in lower:
+            exclude = override.get("exclude", [])
+            betas = [b for b in betas if b not in exclude]
+            for b in override.get("add", []):
+                if b not in betas:
+                    betas.append(b)
+            break
+    return betas
+
+
+def _first_user_message_text(messages: list[dict[str, Any]]) -> str:
+    """Extract the text of the first user message's first text block."""
+    for message in messages:
+        if message.get("role") != "user":
+            continue
+        content = message.get("content")
+        if isinstance(content, str):
+            return content
+        if isinstance(content, list):
+            for block in content:
+                if isinstance(block, dict) and block.get("type") == "text":
+                    return block.get("text", "")
+        return ""
+    return ""
+
+
+def _billing_header_value(messages: list[dict[str, Any]], entrypoint: str) -> str:
+    """Build the `x-anthropic-billing-header` value Claude Code embeds as system[0].
+
+    Mirrors Claude Code's internal cch/version-suffix computation so OAuth
+    requests are billed against the subscription instead of being rejected.
+    """
+    text = _first_user_message_text(messages)
+    version = os.environ.get("ANTHROPIC_CLI_VERSION", _CC_VERSION)
+    sampled = "".join(text[i] if i < len(text) else "0" for i in (4, 7, 20))
+    suffix = hashlib.sha256(f"{_BILLING_SALT}{sampled}{version}".encode()).hexdigest()[:3]
+    cch = hashlib.sha256(text.encode()).hexdigest()[:5]
+    return (
+        f"x-anthropic-billing-header: cc_version={version}.{suffix}; "
+        f"cc_entrypoint={entrypoint}; cch={cch};"
+    )
+
+
+def _build_system_blocks(
+    system_text: str | None, messages: list[dict[str, Any]]
+) -> tuple[list[dict[str, Any]], list[dict[str, Any]]]:
+    """Split the system prompt for OAuth requests.
+
+    Anthropic's API validates the system[] array for OAuth-authenticated
+    requests: only the billing header and the Claude Code identity string may
+    live there. Any other system content triggers a 400 ("out of extra
+    usage"). Everything else is relocated to the front of the first user
+    message, which is functionally equivalent for the model.
+
+    Returns (system_blocks, patched_messages).
+    """
+    entrypoint = os.environ.get("CLAUDE_CODE_ENTRYPOINT", "sdk-cli")
+    system_blocks: list[dict[str, Any]] = [
+        {"type": "text", "text": _billing_header_value(messages, entrypoint)},
+        {"type": "text", "text": _SYSTEM_IDENTITY, "cache_control": {"type": "ephemeral"}},
+    ]
+
+    if not system_text:
+        return system_blocks, messages
+
+    patched = list(messages)
+    prefix = system_text
+    for i, message in enumerate(patched):
+        if message.get("role") != "user":
+            continue
+        content = message.get("content")
+        if isinstance(content, str):
+            patched[i] = {**message, "content": f"{prefix}\n\n{content}"}
+        elif isinstance(content, list):
+            patched[i] = {
+                **message,
+                "content": [{"type": "text", "text": prefix}, *content],
+            }
+        else:
+            patched[i] = {**message, "content": prefix}
+        break
+    return system_blocks, patched
+
 
 class AnthropicClaudeCodeAPI(BaseAPI):
     """Anthropic Messages API using OAuth token auth (Claude Pro/Max).
@@ -100,10 +218,8 @@ class AnthropicClaudeCodeAPI(BaseAPI):
         }
         if not _suppress_temp:
             params["temperature"] = self.options.temperature
-        if system:
-            params["system"] = [
-                {"type": "text", "text": system, "cache_control": {"type": "ephemeral"}}
-            ]
+        system_blocks, params["messages"] = _build_system_blocks(system, params["messages"])
+        params["system"] = system_blocks
         if (
             self.options.thinking_level is not None
             and self.options.thinking_level != ThinkingLevel.Off
@@ -186,7 +302,8 @@ class AnthropicClaudeCodeAPI(BaseAPI):
 
         yield StartEvent()
 
-        async with self._client.messages.stream(**params) as stream:
+        extra_headers = {"anthropic-beta": ",".join(_model_betas(model.id))}
+        async with self._client.messages.stream(**params, extra_headers=extra_headers) as stream:
             async for event in stream:
                 if self._cancelled():
                     yield ErrorEvent(reason=StopReason.Abort, error="Cancelled")

{tau_coding_agent-0.6.9 → tau_coding_agent-0.7.1}/tau/inference/api/text/google_antigravity.py

@@ -181,7 +181,48 @@ def _requires_tool_call_id(model_id: str) -> bool:
     omitting it fails multi-turn tool use with ``tool_use.id: Field required``.
     Gemini models tolerate name-based correlation without it.
     """
-    return model_id.startswith("claude-")
+    return "claude" in model_id
+
+
+def _claude_strict_schema(node: Any) -> Any:
+    """Recursively adapt a tool schema for Claude's strict JSON Schema draft
+    2020-12 validation, which Gemini tolerates violating but Claude 400s on:
+
+    - every object-shaped node needs an explicit ``type: "object"``, at any depth
+    - genuine multi-branch unions (``anyOf``/``oneOf``/``allOf`` with more than
+      one surviving branch, e.g. a Pydantic ``str | SomeModel`` field) are
+      rejected outright by Claude's tool-use engine — collapsed to an untyped
+      (permissive) schema, since Claude has no equivalent of a JSON Schema
+      union for structured tool input.
+    """
+    if isinstance(node, list):
+        return [_claude_strict_schema(item) for item in node]
+    if not isinstance(node, dict):
+        return node
+
+    fixed = dict(node)
+    for key in ("anyOf", "oneOf", "allOf"):
+        if key in fixed and isinstance(fixed[key], list):
+            branches = [_claude_strict_schema(item) for item in fixed[key]]
+            if len(branches) == 1:
+                fixed.pop(key)
+                fixed.update(branches[0])
+            else:
+                # Claude rejects real unions; drop the constraint, keep the
+                # description so the model still knows what shape to send.
+                fixed.pop(key)
+                fixed.pop("type", None)
+    if "properties" in fixed and isinstance(fixed["properties"], dict):
+        fixed["properties"] = {
+            key: _claude_strict_schema(value) for key, value in fixed["properties"].items()
+        }
+        fixed.setdefault("type", "object")
+    if "items" in fixed:
+        fixed["items"] = _claude_strict_schema(fixed["items"])
+    if fixed.get("type") == "object" and "properties" not in fixed:
+        fixed["properties"] = {}
+    return fixed
+
 
 def _messages_to_contents(
     messages: list[LLMMessage],
@@ -337,9 +378,19 @@ class GoogleAntigravityAPI(BaseAPI):
         return self._project_id
 
     @staticmethod
-    def _tools_to_declarations(tools: list[Tool]) -> list[dict[str, Any]]:
-        """Convert Tool objects to Gemini functionDeclarations format."""
+    def _tools_to_declarations(tools: list[Tool], model_id: str = "") -> list[dict[str, Any]]:
+        """Convert Tool objects to Gemini functionDeclarations format.
+
+        The wire envelope always uses ``functionDeclarations[].parameters`` —
+        cloudcode-pa rejects an ``input_schema`` key outright, even when the
+        model routes to Claude server-side. But Claude's backend validates that
+        schema strictly against JSON Schema draft 2020-12 and 400s if any node
+        with `properties` is missing `type: object` — including nested nodes
+        inside `items`/`anyOf`/`oneOf`/`allOf`, which Gemini tolerates but
+        Claude does not — so the fix has to walk the whole tree, not just the root.
+        """
         check_strict_tools_supported(tools)
+        is_claude = "claude" in model_id
         seen: set[str] = set()
         decls = []
         for t in tools:
@@ -348,6 +399,8 @@ class GoogleAntigravityAPI(BaseAPI):
             seen.add(t.name)
             raw = t.schema.model_json_schema() if t.schema else {}
             schema = gemini_tool_schema(raw)
+            if is_claude:
+                schema = _claude_strict_schema(schema)
             decls.append({"name": t.name, "description": t.description or "", "parameters": schema})
         return decls
 
@@ -388,7 +441,7 @@ class GoogleAntigravityAPI(BaseAPI):
         if generation_config:
             inner["generationConfig"] = generation_config
         if tools:
-            decls = self._tools_to_declarations(tools)
+            decls = self._tools_to_declarations(tools, model.id)
             if decls:
                 inner["tools"] = [{"functionDeclarations": decls}]
 
@@ -411,6 +464,10 @@ class GoogleAntigravityAPI(BaseAPI):
             response_format=context.response_format,
         )
         headers = _antigravity_headers(self.options.api_key or "")
+        if "claude" in model.id and model.thinking:
+            # Antigravity's Claude backend disables interleaved thinking (thinking
+            # between tool calls) unless this beta flag is explicitly present.
+            headers["anthropic-beta"] = "interleaved-thinking-2025-05-14"
 
         if self.options.on_payload:
             modified = self.options.on_payload(body)