Toolsets API

create_console_toolset

pydantic_ai_backends.toolsets.console.create_console_toolset(id=None, include_execute=True, require_write_approval=False, require_execute_approval=True, default_ignore_hidden=True, permissions=None, ask_callback=None, ask_fallback='error', max_retries=1, image_support=False, max_image_bytes=DEFAULT_MAX_IMAGE_BYTES, document_support=False, max_document_bytes=DEFAULT_MAX_DOCUMENT_BYTES, edit_format='str_replace', descriptions=None, profile=DEFAULT_PROFILE)

Create the console tools, operating on the run's Pydantic AI workspace.

Every tool reaches ctx.workspace: whatever a workspace capability on the agent supplied — DockerWorkspace, SandboxdWorkspace, StateWorkspace, Pydantic AI's LocalWorkspace, or the harness's E2B and Modal sandboxes. A run without a workspace gets an error from every tool rather than a crash.

Parameters:

Name	Type	Description	Default
id	str | None	Optional unique ID for the toolset.	None
include_execute	bool	Include the execute tool. On a workspace that cannot run commands — read-only, or StateWorkspace — it answers with the workspace's refusal.	True
require_write_approval	bool	Whether write_file and the edit tool require approval. Ignored when permissions is given.	False
require_execute_approval	bool	Whether execute requires approval. Ignored when permissions is given.	True
default_ignore_hidden	bool	Default for grep's hidden-file handling.	True
permissions	PermissionRuleset | None	Ruleset deciding which tools exist and which need approval: an operation defaulting to "deny" drops its tools entirely, one defaulting to "ask" marks them as requiring approval.	None
max_retries	int	Times a tool may retry within one run, with the message fed back to the model — pydantic-ai's own argument validation, and the mistakes toolsets/_failures.py steers on: a missing file, an old_string that is absent or matches twice, a stale read. Past the budget the message is returned rather than raised, so a run never ends on one.	1
image_support	bool	Return recognized image files (.png, .jpg, .jpeg, .gif, .webp) as BinaryContent a multimodal model can see, instead of garbled text.	False
max_image_bytes	int	Largest image returned; bigger ones yield an error.	DEFAULT_MAX_IMAGE_BYTES
document_support	bool	Return recognized documents (.pdf) as BinaryContent for models that understand documents natively. Kept separate from image_support so the two can evolve apart.	False
max_document_bytes	int	Largest document returned; bigger ones yield an error.	DEFAULT_MAX_DOCUMENT_BYTES
edit_format	EditFormat	"str_replace" matches exact strings; "hashline" tags each line with a content hash so the model references lines by number:hash instead of reproducing text.	'str_replace'
descriptions	Mapping[str, str | ToolText] | None	Per-tool text overrides, keyed by tool name: ls, read_file, write_file, edit_file, hashline_edit, glob, grep, execute. A string replaces the tool's description and leaves its argument text alone; a :class:ToolText replaces both. An unknown key raises UserError rather than being ignored, since a silent override is one nobody discovers.	None
profile	Profile	How much guidance the descriptions carry. "coding" includes the guidance written for an agent working in a repository — git, dependencies, debugging a failed command — and "agent" leaves it out, which is about 250 tokens a request an agent with a scratch workspace was paying for advice it could not use.	DEFAULT_PROFILE

Example

Python

from pydantic_ai import Agent
from pydantic_ai.capabilities import LocalWorkspace

from pydantic_ai_backends import create_console_toolset
from pydantic_ai_backends.permissions import DEFAULT_RULESET

toolset = create_console_toolset()
agent = Agent(
    "anthropic:claude-opus-5-5", capabilities=[LocalWorkspace(".")], toolsets=[toolset]
)

hashline = create_console_toolset(edit_format="hashline")
multimodal = create_console_toolset(image_support=True, document_support=True)
guarded = create_console_toolset(permissions=DEFAULT_RULESET)

Source code in src/pydantic_ai_backends/toolsets/console.py

def create_console_toolset(  # noqa: C901
    id: str | None = None,
    include_execute: bool = True,
    require_write_approval: bool = False,
    require_execute_approval: bool = True,
    default_ignore_hidden: bool = True,
    permissions: PermissionRuleset | None = None,
    ask_callback: AskCallback | None = None,
    ask_fallback: AskFallback = "error",
    max_retries: int = 1,
    image_support: bool = False,
    max_image_bytes: int = DEFAULT_MAX_IMAGE_BYTES,
    document_support: bool = False,
    max_document_bytes: int = DEFAULT_MAX_DOCUMENT_BYTES,
    edit_format: EditFormat = "str_replace",
    descriptions: Mapping[str, str | ToolText] | None = None,
    profile: Profile = DEFAULT_PROFILE,
) -> FunctionToolset[Any]:
    """Create the console tools, operating on the run's Pydantic AI workspace.

    Every tool reaches `ctx.workspace`: whatever a workspace capability on the
    agent supplied — `DockerWorkspace`, `SandboxdWorkspace`, `StateWorkspace`,
    Pydantic AI's `LocalWorkspace`, or the harness's E2B and Modal sandboxes. A
    run without a workspace gets an error from every tool rather than a crash.

    Args:
        id: Optional unique ID for the toolset.
        include_execute: Include the `execute` tool. On a workspace that cannot
            run commands — read-only, or `StateWorkspace` — it answers with the
            workspace's refusal.
        require_write_approval: Whether `write_file` and the edit tool require
            approval. Ignored when `permissions` is given.
        require_execute_approval: Whether `execute` requires approval. Ignored
            when `permissions` is given.
        default_ignore_hidden: Default for `grep`'s hidden-file handling.
        permissions: Ruleset deciding which tools exist and which need approval:
            an operation defaulting to "deny" drops its tools entirely, one
            defaulting to "ask" marks them as requiring approval.
        max_retries: Times a tool may retry within one run, with the message
            fed back to the model — pydantic-ai's own argument validation, and
            the mistakes `toolsets/_failures.py` steers on: a missing file, an
            `old_string` that is absent or matches twice, a stale read. Past the
            budget the message is returned rather than raised, so a run never
            ends on one.
        image_support: Return recognized image files (`.png`, `.jpg`, `.jpeg`,
            `.gif`, `.webp`) as `BinaryContent` a multimodal model can see,
            instead of garbled text.
        max_image_bytes: Largest image returned; bigger ones yield an error.
        document_support: Return recognized documents (`.pdf`) as
            `BinaryContent` for models that understand documents natively. Kept
            separate from `image_support` so the two can evolve apart.
        max_document_bytes: Largest document returned; bigger ones yield an error.
        edit_format: `"str_replace"` matches exact strings; `"hashline"` tags
            each line with a content hash so the model references lines by
            `number:hash` instead of reproducing text.
        descriptions: Per-tool text overrides, keyed by tool name: `ls`,
            `read_file`, `write_file`, `edit_file`, `hashline_edit`, `glob`,
            `grep`, `execute`. A string replaces the tool's description and leaves
            its argument text alone; a :class:`ToolText` replaces both. An
            unknown key raises `UserError` rather than being ignored, since a
            silent override is one nobody discovers.
        profile: How much guidance the descriptions carry. `"coding"` includes
            the guidance written for an agent working in a repository — git,
            dependencies, debugging a failed command — and `"agent"` leaves it
            out, which is about 250 tokens a request an agent with a scratch
            workspace was paying for advice it could not use.

    Example:
        ```python
        from pydantic_ai import Agent
        from pydantic_ai.capabilities import LocalWorkspace

        from pydantic_ai_backends import create_console_toolset
        from pydantic_ai_backends.permissions import DEFAULT_RULESET

        toolset = create_console_toolset()
        agent = Agent(
            "anthropic:claude-opus-5-5", capabilities=[LocalWorkspace(".")], toolsets=[toolset]
        )

        hashline = create_console_toolset(edit_format="hashline")
        multimodal = create_console_toolset(image_support=True, document_support=True)
        guarded = create_console_toolset(permissions=DEFAULT_RULESET)
        ```
    """
    overrides: Mapping[str, str | ToolText] = descriptions or {}
    unknown = sorted(set(overrides) - OVERRIDE_KEYS)
    if unknown:
        raise UserError(
            f"Unknown tool name(s) in `descriptions`: {', '.join(unknown)}. "
            f"Valid names: {', '.join(sorted(OVERRIDE_KEYS))}."
        )

    views: weakref.WeakKeyDictionary[Workspace, _ReadView] = weakref.WeakKeyDictionary()

    def backend_for(ctx: RunContext[Any]) -> FileOps:
        """The run's workspace operations, with the ruleset applied to them.

        The one place every tool resolves what it operates on, which is why the
        guard goes here: applying per-path rules needs a path, and a path only
        exists per call.
        """
        return _guard.guarding(
            WorkspaceOps(ctx.workspace),
            permissions,
            workspace=ctx.workspace,
            ask_callback=ask_callback,
            ask_fallback=ask_fallback,
        )

    def view_for(ctx: RunContext[Any]) -> _ReadView:
        """What this toolset has read in the run's workspace, for as long as it lives.

        Stable across tool calls, so an edit to a file changed since it was read
        is refused, and holding nothing of the workspace, so a finished one can
        be collected while a long-lived agent keeps this toolset.
        """
        view = views.get(ctx.workspace)
        if view is None:
            view = views[ctx.workspace] = _ReadView()
        return view

    write_approval = _ruleset.requires_approval(permissions, "write", require_write_approval)
    execute_approval = _ruleset.requires_approval(permissions, "execute", require_execute_approval)

    toolset: FunctionToolset[Any] = FunctionToolset(id=id, max_retries=max_retries)

    def described(
        text_id: str,
        tool_name: str | None = None,
        *,
        requires_approval: bool = False,
    ) -> Callable[[_ToolFn], _ToolFn]:
        """Register a tool with the text this configuration gives it.

        One `ToolText` supplies both halves of what the model reads, but they
        travel separately: the description is passed to the decorator, while the
        per-argument text reaches the JSON schema through the function's
        docstring and through nothing else — hence the assignment. It is also
        why the tools below carry a one-line docstring rather than a second copy
        of the argument text, which is a copy that drifts.

        Args:
            text_id: Key in `TOOL_TEXT`. Differs from the tool name only for
                `read_file`, which has one text per edit format.
            tool_name: Name a caller overrides this tool by, when it is not the
                text id.
            requires_approval: Whether the tool call is suspended for a human.
        """
        name = tool_name or text_id
        override = overrides.get(name)
        text = override if isinstance(override, ToolText) else TOOL_TEXT[text_id]
        description = override if isinstance(override, str) else text.render(profile)

        def register(fn: _ToolFn) -> _ToolFn:
            cast("Any", fn).__doc__ = text.docstring()
            registered = toolset.tool(description=description, requires_approval=requires_approval)(
                fn
            )
            return cast("_ToolFn", registered)

        return register

    async def binary_content(
        target: FileOps, path: str
    ) -> Any | None:  # pragma: no cover - exercised through read_file
        """Image or document content for `path`, when either is enabled."""
        if image_support:
            image = await image_content(target, path, max_image_bytes)
            if image is not None:
                return image
        if document_support:
            return await document_content(target, path, max_document_bytes)
        return None

    @described("ls")
    @_degrade_on_error
    async def ls(
        ctx: RunContext[Any],
        path: str = ".",
    ) -> str:
        """List files and directories at the given path."""
        entries = await backend_for(ctx).ls_info(path)
        if not entries:
            return f"Directory '{path}' is empty or does not exist"

        lines = [f"Contents of {path}:"]
        for entry in entries:
            if entry["is_dir"]:
                lines.append(f"  {entry['name']}/")
            else:
                size = entry.get("size")
                lines.append(f"  {entry['name']}{f' ({size} bytes)' if size is not None else ''}")
        return "\n".join(lines)

    if edit_format == "hashline":

        @described("hashline_read_file", "read_file")
        @_degrade_on_error
        async def read_file(
            ctx: RunContext[Any],
            path: str,
            offset: int = 0,
            limit: int = 2000,
        ) -> Any:
            """Read file content with hashline tags."""
            binary = await binary_content(backend_for(ctx), path)
            if binary is not None:
                return binary

            from pydantic_ai_backends.hashline import format_hashline_output

            backend = backend_for(ctx)
            if not await backend.exists(path):
                return _failures.steer(
                    ctx,
                    f"Error: File '{path}' not found. Check the path with `ls` or "
                    "`glob`, then read it again.",
                )

            raw = await backend.read_bytes(path)
            _tracking.record_read(view_for(ctx), path, raw)
            return format_hashline_output(raw.decode("utf-8", errors="replace"), offset, limit)

    else:

        @described("read_file")
        @_degrade_on_error
        async def read_file(
            ctx: RunContext[Any],
            path: str,
            offset: int = 0,
            limit: int = 2000,
        ) -> Any:
            """Read file content with line numbers."""
            binary = await binary_content(backend_for(ctx), path)
            if binary is not None:
                return binary

            backend = backend_for(ctx)
            result = await backend.read(path, offset, limit)
            if result.startswith("Error"):
                return _failures.steer(ctx, result)
            await _tracking.record_path_read(backend, view_for(ctx), path)
            return result

    @described("write_file", requires_approval=write_approval)
    @_degrade_on_error
    async def write_file(
        ctx: RunContext[Any],
        path: str,
        content: str,
    ) -> str:
        """Write content to a file."""
        result = await backend_for(ctx).write(path, content)
        if result.error:
            return _failures.steer(ctx, f"Error: {result.error}")

        # The agent knows this file's content now, so an immediate edit must not
        # be refused as stale.
        _tracking.record_read(view_for(ctx), path, content.encode("utf-8"))
        return f"Wrote {len(content.splitlines())} lines to {result.path}"

    if edit_format == "hashline":

        @described("hashline_edit", requires_approval=write_approval)
        @_degrade_on_error
        async def hashline_edit(
            ctx: RunContext[Any],
            path: str,
            start_line: int,
            start_hash: str,
            new_content: str,
            end_line: int | None = None,
            end_hash: str | None = None,
            insert_after: bool = False,
        ) -> str:
            """Edit a file by referencing lines with their content hashes."""
            from pydantic_ai_backends.hashline import apply_hashline_edit_with_summary

            backend = backend_for(ctx)

            async with _tracking.edit_lock(view_for(ctx), path):
                if not await backend.exists(path):
                    return _failures.steer(ctx, f"Error: File '{path}' not found")

                current = (await backend.read_bytes(path)).decode("utf-8", errors="replace")
                new_text, error, summary = apply_hashline_edit_with_summary(
                    current,
                    start_line,
                    start_hash,
                    new_content,
                    end_line,
                    end_hash,
                    insert_after,
                )
                if error:
                    return _failures.steer(ctx, f"Error: {error}")

                written = await backend.write(path, new_text)
                if written.error:
                    return _failures.steer(ctx, f"Error: {written.error}")
                return f"Edited {written.path}: {summary}"

    else:

        @described("edit_file", requires_approval=write_approval)
        @_degrade_on_error
        async def edit_file(
            ctx: RunContext[Any],
            path: str,
            old_string: str,
            new_string: str,
            replace_all: bool = False,
        ) -> str:
            """Edit a file by performing exact string replacement."""
            backend = backend_for(ctx)
            view = view_for(ctx)

            # Locked for the same reason `hashline_edit` is: every backend's
            # `edit` is a read, a replace and a write, so two edits to one path
            # in flight together lose one of them. The staleness check belongs
            # inside the lock too — checked outside, it is answered before the
            # other edit's write and passes on content that no longer exists.
            async with _tracking.edit_lock(view, path):
                stale = await _tracking.staleness_error(backend, view, path)
                if stale is not None:
                    return _failures.steer(ctx, stale)

                result = await backend.edit(path, old_string, new_string, replace_all)
                if result.error:
                    return _failures.steer(ctx, f"Error: {result.error}")

                # The agent's view is the post-edit content now, so a follow-up
                # edit must not be flagged as stale.
                await _tracking.record_path_read(backend, view, path)
                return f"Edited {result.path}: replaced {result.occurrences} occurrence(s)"

    @described("glob")
    @_degrade_on_error
    async def glob(
        ctx: RunContext[Any],
        pattern: str,
        path: str = ".",
    ) -> str:
        """Find files matching a glob pattern."""
        entries = await backend_for(ctx).glob_info(pattern, path)
        if not entries:
            return f"No files matching '{pattern}' in {path}"

        lines = [f"Found {len(entries)} file(s) matching '{pattern}':"]
        lines.extend(f"  {entry['path']}" for entry in entries[:GLOB_RESULT_LIMIT])
        if len(entries) > GLOB_RESULT_LIMIT:
            lines.append(f"  ... and {len(entries) - GLOB_RESULT_LIMIT} more")
        return "\n".join(lines)

    @described("grep")
    @_degrade_on_error
    async def grep(
        ctx: RunContext[Any],
        pattern: str,
        path: str | None = None,
        glob_pattern: str | None = None,
        output_mode: Literal["content", "files_with_matches", "count"] = "files_with_matches",
        ignore_hidden: bool = default_ignore_hidden,
    ) -> str:
        """Search for a regex pattern across files."""
        result = await backend_for(ctx).grep_raw(pattern, path, glob_pattern, ignore_hidden)
        if isinstance(result, str):
            return result
        if not result:
            return f"No matches for '{pattern}'"

        matches: list[GrepMatch] = result
        if output_mode == "count":
            return f"Found {len(matches)} match(es) for '{pattern}'"

        if output_mode == "files_with_matches":
            files = sorted({match["path"] for match in matches})
            return _truncated_list(f"Files containing '{pattern}':", files, "more files")

        rendered = [
            f"{m['path']}:{m['line_number']}: {m['line'][:GREP_LINE_WIDTH]}" for m in matches
        ]
        return _truncated_list(f"Matches for '{pattern}':", rendered, "more matches")

    # Exposed for the test suite.
    cast(_ConsoleToolsetTestAttrs, toolset)._console_default_ignore_hidden = default_ignore_hidden
    cast(_ConsoleToolsetTestAttrs, toolset)._console_grep_impl = grep

    if include_execute:

        @described("execute", requires_approval=execute_approval)
        @_degrade_on_error
        async def execute(
            ctx: RunContext[Any],
            command: str,
            timeout: int | None = DEFAULT_EXECUTE_TIMEOUT,
        ) -> str:
            """Execute a shell command in the working directory."""
            result = await backend_for(ctx).execute(command, timeout)

            output = result.output
            if result.truncated:
                output += "\n\n... (output truncated)"
            if result.exit_code is not None and result.exit_code != 0:
                return f"Command failed (exit code {result.exit_code}):\n{output}"
            return str(output)

        # Exposed for the test suite.
        cast(_ConsoleToolsetTestAttrs, toolset)._console_execute_impl = execute

    for tool_name in _denied_tools(permissions):
        toolset.tools.pop(tool_name, None)

    return toolset

get_console_system_prompt

pydantic_ai_backends.toolsets.console.get_console_system_prompt(edit_format='str_replace')

The system prompt describing the console tools.

Parameters:

Name	Type	Description	Default
edit_format	EditFormat	Which edit format to describe.	'str_replace'

Source code in src/pydantic_ai_backends/toolsets/console.py

def get_console_system_prompt(edit_format: EditFormat = "str_replace") -> str:
    """The system prompt describing the console tools.

    Args:
        edit_format: Which edit format to describe.
    """
    if edit_format == "hashline":
        return HASHLINE_CONSOLE_PROMPT
    return CONSOLE_SYSTEM_PROMPT

ToolText

pydantic_ai_backends.toolsets.descriptions.ToolText dataclass

Everything the model reads about one tool.

Held as fields rather than one string because the parts have different destinations: summary, usage, coding and returns are composed into the tool's description, while args becomes the per-argument text in its JSON schema. Splitting them is also what lets a host show summary in its own catalogue and know it is the first sentence the model reads, rather than a paraphrase written in another repository.

Source code in src/pydantic_ai_backends/toolsets/descriptions.py

@dataclass(frozen=True)
class ToolText:
    """Everything the model reads about one tool.

    Held as fields rather than one string because the parts have different
    destinations: `summary`, `usage`, `coding` and `returns` are composed into
    the tool's description, while `args` becomes the per-argument text in its
    JSON schema. Splitting them is also what lets a host show `summary` in its
    own catalogue and know it is the first sentence the model reads, rather than
    a paraphrase written in another repository.
    """

    summary: str
    """One sentence: what the tool does. Also what a catalogue should show."""

    usage: str = ""
    """When to use it, when to use another tool, and what it will not do."""

    coding: str = ""
    """Guidance only an agent working in a repository needs.

    Rendered under the `"coding"` profile and omitted under `"agent"`. Anything
    true of any workspace belongs in `usage` instead.
    """

    args: Mapping[str, str] = field(default_factory=dict)
    """One entry per argument, keyed exactly as the parameter is named."""

    returns: str = ""
    """The shape of the result, including its failures and its truncation."""

    def render(self, profile: Profile = DEFAULT_PROFILE) -> str:
        """The description handed to the model.

        Shaped the way pydantic-ai shapes a docstring that has a `Returns:`
        section - the prose inside `<summary>`, the return description inside
        `<returns>` - because that is what every tool built from a docstring
        already sends, and a host registering these beside its own would
        otherwise put two conventions in one tool list. A prose `Returns:`
        paragraph was the first attempt and is what that inconsistency looked
        like. `tests/test_tool_text.py` pins the shape against a tool the
        framework renders itself, so a change there fails here rather than
        drifting quietly.

        Args:
            profile: Which audience to write for.
        """
        parts = [self.summary]
        if self.usage:
            parts.append(self.usage)
        if self.coding and profile == "coding":
            parts.append(self.coding)
        body = "\n\n".join(parts)
        if not self.returns:
            return body
        return (
            f"<summary>{body}</summary>\n"
            f"<returns>\n<description>{self.returns}</description>\n</returns>"
        )

    def docstring(self) -> str:
        """A Google-style docstring carrying the argument text.

        Set on the tool function before it is registered, because per-argument
        descriptions reach the JSON schema through the docstring and through
        nothing else — which is why they cannot simply live in `render`. The
        summary is repeated here for a reader of the generated docstring; the
        model reads `render` instead, since an explicit description wins over
        the docstring's own summary.
        """
        lines = [self.summary]
        if self.args:
            lines.append("")
            lines.append("Args:")
            lines.extend(f"    {name}: {text}" for name, text in self.args.items())
        return "\n".join(lines)

summary instance-attribute

One sentence: what the tool does. Also what a catalogue should show.

usage = '' class-attribute instance-attribute

When to use it, when to use another tool, and what it will not do.

coding = '' class-attribute instance-attribute

Guidance only an agent working in a repository needs.

Rendered under the "coding" profile and omitted under "agent". Anything true of any workspace belongs in usage instead.

args = field(default_factory=dict) class-attribute instance-attribute

One entry per argument, keyed exactly as the parameter is named.

returns = '' class-attribute instance-attribute

The shape of the result, including its failures and its truncation.

render(profile=DEFAULT_PROFILE)

The description handed to the model.

Shaped the way pydantic-ai shapes a docstring that has a Returns: section - the prose inside <summary>, the return description inside <returns> - because that is what every tool built from a docstring already sends, and a host registering these beside its own would otherwise put two conventions in one tool list. A prose Returns: paragraph was the first attempt and is what that inconsistency looked like. tests/test_tool_text.py pins the shape against a tool the framework renders itself, so a change there fails here rather than drifting quietly.

Parameters:

Name	Type	Description	Default
profile	Profile	Which audience to write for.	DEFAULT_PROFILE

Source code in src/pydantic_ai_backends/toolsets/descriptions.py

def render(self, profile: Profile = DEFAULT_PROFILE) -> str:
    """The description handed to the model.

    Shaped the way pydantic-ai shapes a docstring that has a `Returns:`
    section - the prose inside `<summary>`, the return description inside
    `<returns>` - because that is what every tool built from a docstring
    already sends, and a host registering these beside its own would
    otherwise put two conventions in one tool list. A prose `Returns:`
    paragraph was the first attempt and is what that inconsistency looked
    like. `tests/test_tool_text.py` pins the shape against a tool the
    framework renders itself, so a change there fails here rather than
    drifting quietly.

    Args:
        profile: Which audience to write for.
    """
    parts = [self.summary]
    if self.usage:
        parts.append(self.usage)
    if self.coding and profile == "coding":
        parts.append(self.coding)
    body = "\n\n".join(parts)
    if not self.returns:
        return body
    return (
        f"<summary>{body}</summary>\n"
        f"<returns>\n<description>{self.returns}</description>\n</returns>"
    )

docstring()

A Google-style docstring carrying the argument text.

Set on the tool function before it is registered, because per-argument descriptions reach the JSON schema through the docstring and through nothing else — which is why they cannot simply live in render. The summary is repeated here for a reader of the generated docstring; the model reads render instead, since an explicit description wins over the docstring's own summary.

Source code in src/pydantic_ai_backends/toolsets/descriptions.py

def docstring(self) -> str:
    """A Google-style docstring carrying the argument text.

    Set on the tool function before it is registered, because per-argument
    descriptions reach the JSON schema through the docstring and through
    nothing else — which is why they cannot simply live in `render`. The
    summary is repeated here for a reader of the generated docstring; the
    model reads `render` instead, since an explicit description wins over
    the docstring's own summary.
    """
    lines = [self.summary]
    if self.args:
        lines.append("")
        lines.append("Args:")
        lines.extend(f"    {name}: {text}" for name, text in self.args.items())
    return "\n".join(lines)

Console Tools

The toolset registers these tools. What each one says — its description and the text describing every argument — is not written beside the function: it lives in TOOL_TEXT, keyed by tool name, and is assigned when the tool is registered. Read TOOL_TEXT["grep"].render() to see exactly what the model is handed.

Python

async def ls(ctx, path: str = ".") -> str: ...
async def read_file(ctx, path: str, offset: int = 0, limit: int = 2000) -> str: ...
async def write_file(ctx, path: str, content: str) -> str: ...
async def edit_file(
    ctx, path: str, old_string: str, new_string: str, replace_all: bool = False
) -> str: ...
async def hashline_edit(
    ctx,
    path: str,
    start_line: int,
    start_hash: str,
    new_content: str,
    end_line: int | None = None,
    end_hash: str | None = None,
    insert_after: bool = False,
) -> str: ...
async def glob(ctx, pattern: str, path: str = ".") -> str: ...
async def grep(
    ctx,
    pattern: str,
    path: str | None = None,
    glob_pattern: str | None = None,
    output_mode: Literal["content", "files_with_matches", "count"] = "files_with_matches",
    ignore_hidden: bool = True,
) -> str: ...
async def execute(ctx, command: str, timeout: int | None = 120) -> str: ...