deepy-cli 0.2.25__tar.gz → 0.2.26__tar.gz

{deepy_cli-0.2.25 → deepy_cli-0.2.26}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.3
 Name: deepy-cli
-Version: 0.2.25
+Version: 0.2.26
 Summary: Deepy - Vibe coding for DeepSeek models in your terminal
 Keywords: deepseek,coding-agent,terminal,cli,agents
 Author: kirineko

{deepy_cli-0.2.25 → deepy_cli-0.2.26}/pyproject.toml

@@ -1,6 +1,6 @@
 [project]
 name = "deepy-cli"
-version = "0.2.25"
+version = "0.2.26"
 description = "Deepy - Vibe coding for DeepSeek models in your terminal"
 readme = "README.md"
 authors = [

{deepy_cli-0.2.25 → deepy_cli-0.2.26}/src/deepy/__init__.py

@@ -1,6 +1,6 @@
 from __future__ import annotations
 
-__version__ = "0.2.25"
+__version__ = "0.2.26"
 
 
 def main() -> None:

{deepy_cli-0.2.25 → deepy_cli-0.2.26}/src/deepy/cli.py

@@ -30,6 +30,7 @@ from .config import (
 from .config.settings import DEFAULT_UI_THEME, UI_THEMES
 from .errors import format_error_display
 from .llm.cache_context import format_cache_usage
+from .llm.multimodal import redact_image_data_urls
 from .llm.provider import build_provider_bundle
 from .llm.runner import DEFAULT_MAX_TURNS, run_prompt_once
 from .sessions import DeepySession, list_session_entries
@@ -614,7 +615,7 @@ def _cmd_sessions(args: argparse.Namespace) -> int:
                     else 0,
                     "cache_break_reason": entry.cache_break_reason if entry is not None else None,
                     "cache_usage": entry.cache_usage if entry is not None else None,
-                    "items": items,
+                    "items": redact_image_data_urls(items),
                 }
             )
         )

{deepy_cli-0.2.25 → deepy_cli-0.2.26}/src/deepy/config/settings.py

@@ -59,6 +59,7 @@ class ModelInfo:
     label: str
     description: str
     supports_thinking: bool = True
+    supports_image_input: bool = False
     default_reasoning_mode: str = "max"
 
 
@@ -102,6 +103,7 @@ OPENROUTER_MODEL_CATALOG = (
         name="xiaomi/mimo-v2.5",
         label="MiMo V2.5",
         description="Xiaomi MiMo V2.5 via OpenRouter.",
+        supports_image_input=True,
         default_reasoning_mode="enabled",
     ),
 )
@@ -116,6 +118,7 @@ XIAOMI_MODEL_CATALOG = (
         name="mimo-v2.5",
         label="MiMo V2.5",
         description="Xiaomi official MiMo V2.5.",
+        supports_image_input=True,
         default_reasoning_mode="enabled",
     ),
 )

{deepy_cli-0.2.25 → deepy_cli-0.2.26}/src/deepy/data/tools/test_shell.md

@@ -11,6 +11,8 @@ redirection, command substitution, heredocs, and background operators.
 
 Low-risk verification commands run immediately and return command, cwd,
 exit-code, elapsed time, stdout, stderr, and truncation metadata. Medium-risk
-commands return `approval_required` with an `approvalToken`; the main Deepy
-agent must ask the user before retrying the same command with that token.
-Destructive, publishing, mutating, or unsupported commands are denied.
+commands are routed through Deepy's outer audit approval flow when an audit
+policy is active; after approval they still execute through this constrained
+tool. Without an active audit policy, medium-risk commands return
+`approval_required` with an `approvalToken` for same-command retry. Destructive,
+publishing, mutating, or unsupported commands are denied.

{deepy_cli-0.2.25 → deepy_cli-0.2.26}/src/deepy/llm/compaction.py

@@ -132,7 +132,7 @@ async def ensure_context_ready(
     prefix_snapshot: CachePrefixSnapshot | None = None,
     prefix_tools: list[Any] | None = None,
     prefix_mcp_servers: list[Any] | None = None,
-    additional_input: str | None = None,
+    additional_input: Any | None = None,
 ) -> ContextReadiness:
     additional_tokens = estimate_tokens_for_item(additional_input or "")
     state = session.context_token_state()

{deepy_cli-0.2.25 → deepy_cli-0.2.26}/src/deepy/llm/context.py

@@ -4,6 +4,8 @@ from math import ceil
 from typing import Any
 
 from deepy.config import Settings
+from deepy.llm.multimodal import item_contains_image_content, strip_image_content_from_items
+from deepy.llm.multimodal import supports_image_input
 from deepy.types.sdk import SessionInputCallback
 from deepy.utils import json as json_utils
 
@@ -28,6 +30,8 @@ def estimate_tokens_for_text(text: str) -> int:
 
 
 def estimate_tokens_for_item(item: Any) -> int:
+    if item_contains_image_content(item):
+        return _estimate_multimodal_item_tokens(item)
     if isinstance(item, str):
         return estimate_tokens_for_text(item)
     if isinstance(item, dict):
@@ -37,13 +41,34 @@ def estimate_tokens_for_item(item: Any) -> int:
     return estimate_tokens_for_text(str(item))
 
 
+def _estimate_multimodal_item_tokens(item: Any) -> int:
+    if not isinstance(item, dict):
+        return estimate_tokens_for_text(str(item))
+    content = item.get("content")
+    if not isinstance(content, list):
+        return estimate_tokens_for_text(json_utils.dumps(item))
+    tokens = 0
+    for part in content:
+        if not isinstance(part, dict):
+            tokens += estimate_tokens_for_item(part)
+            continue
+        if part.get("type") in {"input_image", "image", "image_url"} or "image_url" in part:
+            tokens += 1024
+            continue
+        tokens += estimate_tokens_for_item(part)
+    return max(tokens, 1)
+
+
 def estimate_tokens_for_items(items: list[dict[str, Any]]) -> int:
     return sum(estimate_tokens_for_item(item) for item in items)
 
 
 def build_session_input_callback(settings: Settings) -> SessionInputCallback:
     def callback(history: list[Any], new_input: list[Any]) -> list[Any]:
-        return [*history, *new_input]
+        items = [*history, *new_input]
+        if not supports_image_input(settings):
+            return strip_image_content_from_items(items)
+        return items
 
     return callback
 

deepy_cli-0.2.26/src/deepy/llm/multimodal.py

@@ -0,0 +1,279 @@
+from __future__ import annotations
+
+import base64
+import re
+from dataclasses import dataclass
+from typing import Any
+
+from deepy.config import Settings
+
+
+SUPPORTED_IMAGE_MIME_TYPES = frozenset(
+    {
+        "image/png",
+        "image/jpeg",
+        "image/webp",
+        "image/gif",
+    }
+)
+DEFAULT_MAX_IMAGE_BYTES = 50 * 1024 * 1024
+UNSUPPORTED_IMAGE_INPUT_MESSAGE = "当前模型不支持图片输入，已忽略粘贴的图片。"
+IMAGE_ONLY_DEFAULT_TEXT = "请描述这张图片的内容，不要执行工具或修改文件。"
+IMAGE_DATA_URL_RE = re.compile(r"^data:image/[a-zA-Z0-9.+-]+;base64,", re.IGNORECASE)
+
+
+class ImageAttachmentError(ValueError):
+    pass
+
+
+class UnsupportedImageInputError(RuntimeError):
+    pass
+
+
+@dataclass(frozen=True)
+class PromptImageAttachment:
+    label: str
+    mime_type: str
+    data_base64: str
+    byte_size: int
+    source: str = "clipboard"
+    data_ref: str | None = None
+
+    @property
+    def display_label(self) -> str:
+        return f"[{self.label}]"
+
+    @property
+    def data_url(self) -> str:
+        return f"data:{self.mime_type};base64,{self.data_base64}"
+
+    def to_input_image_block(self) -> dict[str, str]:
+        return {"type": "input_image", "image_url": self.data_url}
+
+
+def supports_image_input(settings: Settings) -> bool:
+    return model_supports_image_input(settings.model.provider, settings.model.name)
+
+
+def model_supports_image_input(provider: str, model: str) -> bool:
+    normalized_provider = provider.strip().lower()
+    normalized_model = model.strip().lower()
+    if normalized_provider == "xiaomi":
+        return normalized_model == "mimo-v2.5"
+    if normalized_provider == "openrouter":
+        return normalized_model == "xiaomi/mimo-v2.5"
+    return False
+
+
+def validate_image_attachment(
+    *,
+    mime_type: str,
+    byte_size: int,
+    max_bytes: int = DEFAULT_MAX_IMAGE_BYTES,
+) -> None:
+    normalized_mime = mime_type.strip().lower()
+    if normalized_mime not in SUPPORTED_IMAGE_MIME_TYPES:
+        raise ImageAttachmentError(f"不支持的图片格式：{mime_type or 'unknown'}")
+    if byte_size <= 0:
+        raise ImageAttachmentError("图片为空，已忽略粘贴的图片。")
+    if byte_size > max_bytes:
+        raise ImageAttachmentError("图片过大，已忽略粘贴的图片。")
+
+
+def build_prompt_image_attachment(
+    *,
+    data: bytes,
+    mime_type: str,
+    index: int,
+    source: str = "clipboard",
+    max_bytes: int = DEFAULT_MAX_IMAGE_BYTES,
+) -> PromptImageAttachment:
+    normalized_mime = mime_type.strip().lower()
+    validate_image_attachment(
+        mime_type=normalized_mime,
+        byte_size=len(data),
+        max_bytes=max_bytes,
+    )
+    return PromptImageAttachment(
+        label=f"图片{index}",
+        mime_type=normalized_mime,
+        data_base64=base64.b64encode(data).decode("ascii"),
+        byte_size=len(data),
+        source=source,
+    )
+
+
+def image_attachment_labels(attachments: list[PromptImageAttachment]) -> str:
+    return " ".join(attachment.display_label for attachment in attachments)
+
+
+def format_user_prompt_display(prompt: str, attachments: list[PromptImageAttachment]) -> str:
+    labels = image_attachment_labels(attachments)
+    text = prompt.strip()
+    if text and labels:
+        return f"{text}\n{labels}"
+    return text or labels
+
+
+def build_user_input(
+    prompt: str,
+    attachments: list[PromptImageAttachment] | None = None,
+) -> str | list[dict[str, Any]]:
+    image_attachments = list(attachments or [])
+    if not image_attachments:
+        return prompt
+    content: list[dict[str, str]] = []
+    if prompt.strip():
+        content.append({"type": "input_text", "text": prompt})
+    content.extend(attachment.to_input_image_block() for attachment in image_attachments)
+    return [{"role": "user", "content": content}]
+
+
+def item_contains_image_content(item: Any) -> bool:
+    if isinstance(item, dict):
+        return _content_contains_image(item.get("content"))
+    return _content_contains_image(getattr(item, "content", None))
+
+
+def items_contain_image_content(items: list[Any]) -> bool:
+    return any(item_contains_image_content(item) for item in items)
+
+
+def strip_image_content_from_items(items: list[Any]) -> list[Any]:
+    stripped: list[Any] = []
+    for item in items:
+        cleaned = strip_image_content_from_item(item)
+        if cleaned is not None:
+            stripped.append(cleaned)
+    return stripped
+
+
+def strip_image_content_from_item(item: Any) -> Any | None:
+    if not isinstance(item, dict) or "content" not in item:
+        return item
+    cleaned = dict(item)
+    content = _strip_image_content(item.get("content"))
+    if content is None:
+        return None
+    cleaned["content"] = content
+    return cleaned
+
+
+def redacted_content_text(value: Any) -> str:
+    if isinstance(value, str):
+        return _redact_data_urls(value)
+    if isinstance(value, list):
+        parts: list[str] = []
+        image_index = 1
+        for part in value:
+            if not isinstance(part, dict):
+                continue
+            if _part_is_image(part):
+                parts.append(f"[图片{image_index}]")
+                image_index += 1
+                continue
+            text = _text_part(part)
+            if text:
+                parts.append(_redact_data_urls(text))
+        return "\n".join(parts)
+    if isinstance(value, dict):
+        if _part_is_image(value):
+            return "[图片1]"
+        text = _text_part(value)
+        return _redact_data_urls(text) if text else ""
+    return "" if value is None else _redact_data_urls(str(value))
+
+
+def redact_image_data_urls(value: Any) -> Any:
+    if isinstance(value, str):
+        return _redact_data_urls(value)
+    if isinstance(value, list):
+        return [redact_image_data_urls(item) for item in value]
+    if isinstance(value, dict):
+        redacted = {key: redact_image_data_urls(item) for key, item in value.items()}
+        if _part_is_image(redacted):
+            if isinstance(redacted.get("image_url"), str):
+                redacted["image_url"] = "[图片]"
+            elif isinstance(redacted.get("image_url"), dict):
+                image_url = dict(redacted["image_url"])
+                image_url["url"] = "[图片]"
+                redacted["image_url"] = image_url
+        return redacted
+    return value
+
+
+def normalize_multimodal_content_blocks(content: Any) -> Any:
+    if not isinstance(content, list):
+        return content
+    normalized: list[Any] = []
+    has_text = False
+    has_image = False
+    for part in content:
+        if not isinstance(part, dict):
+            normalized.append(part)
+            continue
+        part_type = part.get("type")
+        if part_type == "input_text":
+            text = part.get("text")
+            if text is None:
+                text = part.get("input_text")
+            text_value = text if isinstance(text, str) else ""
+            has_text = has_text or bool(text_value.strip())
+            normalized.append({"type": "text", "text": text_value})
+            continue
+        if part_type == "input_image":
+            image_url = part.get("image_url")
+            has_image = True
+            normalized.append(
+                {
+                    "type": "image_url",
+                    "image_url": {
+                        "url": image_url if isinstance(image_url, str) else "",
+                    },
+                }
+            )
+            continue
+        if _part_is_image(part):
+            has_image = True
+        elif (text := _text_part(part)).strip():
+            has_text = True
+        normalized.append(part)
+    if has_image and not has_text:
+        normalized.insert(0, {"type": "text", "text": IMAGE_ONLY_DEFAULT_TEXT})
+    return normalized
+
+
+def _content_contains_image(content: Any) -> bool:
+    if isinstance(content, list):
+        return any(isinstance(part, dict) and _part_is_image(part) for part in content)
+    return isinstance(content, dict) and _part_is_image(content)
+
+
+def _strip_image_content(content: Any) -> Any | None:
+    if isinstance(content, list):
+        parts = [
+            part
+            for part in content
+            if not (isinstance(part, dict) and _part_is_image(part))
+        ]
+        return parts or None
+    if isinstance(content, dict) and _part_is_image(content):
+        return None
+    return content
+
+
+def _part_is_image(part: dict[str, Any]) -> bool:
+    part_type = part.get("type")
+    return part_type in {"input_image", "image", "image_url"} or "image_url" in part
+
+
+def _text_part(part: dict[str, Any]) -> str:
+    for key in ("text", "input_text", "output_text", "refusal"):
+        value = part.get(key)
+        if isinstance(value, str):
+            return value
+    return ""
+
+
+def _redact_data_urls(text: str) -> str:
+    return IMAGE_DATA_URL_RE.sub("data:image/...;base64,", text)

{deepy_cli-0.2.25 → deepy_cli-0.2.26}/src/deepy/llm/provider.py

@@ -8,8 +8,14 @@ from agents import Model, ModelSettings
 from agents import OpenAIChatCompletionsModel
 
 from deepy.config import Settings
+from deepy.config.settings import infer_provider_from_base_url
 
 from .cache_context import capture_sdk_request_shape
+from .multimodal import (
+    items_contain_image_content,
+    model_supports_image_input,
+    strip_image_content_from_items,
+)
 from .replay import (
     sanitize_chat_completion_stream_event,
     sanitize_model_input_for_chat_completions,
@@ -43,10 +49,21 @@ class DeepyOpenAIChatCompletionsModel(OpenAIChatCompletionsModel):
         *args: Any,
         **kwargs: Any,
     ) -> Any:
+        model_name = str(getattr(self, "model", ""))
+        base_url = str(getattr(self._get_client(), "base_url", "") or "")
+        inferred_provider = infer_provider_from_base_url(base_url) or (
+            "openrouter" if _is_openrouter_base_url(base_url) else ""
+        )
+        if (
+            isinstance(input, list)
+            and items_contain_image_content(input)
+            and not model_supports_image_input(inferred_provider, model_name)
+        ):
+            input = strip_image_content_from_items(input)
         capture_sdk_request_shape(
             system_instructions=system_instructions,
             input=input,
-            model=str(getattr(self, "model", "")),
+            model=model_name,
             model_settings=args[0] if args else None,
             tools=args[1] if len(args) > 1 and isinstance(args[1], list) else None,
             mcp_servers=None,

{deepy_cli-0.2.25 → deepy_cli-0.2.26}/src/deepy/llm/replay.py

@@ -4,6 +4,8 @@ from collections import Counter
 from collections.abc import Iterable
 from typing import Any, cast
 
+from deepy.llm.multimodal import normalize_multimodal_content_blocks
+
 
 def sanitize_model_input_for_chat_completions(input_value: Any) -> Any:
     if not isinstance(input_value, list):
@@ -92,10 +94,18 @@ def _normalize_chat_tool_items(items: Iterable[Any]) -> list[Any]:
                     }
                 )
                 continue
-        normalized.append(item)
+        normalized.append(_normalize_multimodal_item(item))
     return normalized
 
 
+def _normalize_multimodal_item(item: Any) -> Any:
+    if isinstance(item, dict) and "content" in item:
+        normalized = dict(item)
+        normalized["content"] = normalize_multimodal_content_blocks(item.get("content"))
+        return normalized
+    return item
+
+
 def sanitize_chat_completion_stream_event(event: Any) -> Any | None:
     if getattr(event, "type", None) == "response.output_item.done" and _is_empty_assistant_message(
         getattr(event, "item", None)

{deepy_cli-0.2.25 → deepy_cli-0.2.26}/src/deepy/llm/runner.py

@@ -30,6 +30,11 @@ from .cache_context import (
 from .compaction import ContextCompactionError, ensure_context_ready
 from .context import build_session_input_callback
 from .events import DeepyStreamEvent, normalize_stream_event
+from .multimodal import (
+    PromptImageAttachment,
+    build_user_input,
+    supports_image_input,
+)
 from .provider import ProviderBundle, build_provider_bundle
 
 DEFAULT_MAX_TURNS = 100
@@ -73,6 +78,7 @@ async def run_prompt_once(
         list[ApprovalDecision] | Awaitable[list[ApprovalDecision]],
     ]
     | None = None,
+    image_attachments: list[PromptImageAttachment] | None = None,
 ) -> RunSummary:
     from agents import RunConfig, Runner
     from agents.exceptions import MaxTurnsExceeded, ModelBehaviorError
@@ -86,6 +92,9 @@ async def run_prompt_once(
     )
     audit_policy = AuditPolicy(lambda: audit_state.mode, resolved_settings.audit)
     session = DeepySession.open(root, session_id) if session_id else DeepySession.create(root)
+    effective_image_attachments = (
+        list(image_attachments or []) if supports_image_input(resolved_settings) else []
+    )
     initial_todos, _ = normalize_todo_items(session.todo_state())
     runtime = ToolRuntime(
         cwd=root,
@@ -134,7 +143,7 @@ async def run_prompt_once(
             prefix_snapshot=prefix_snapshot,
             prefix_tools=list(getattr(agent, "tools", []) or []),
             prefix_mcp_servers=list(getattr(agent, "mcp_servers", []) or []),
-            additional_input=prompt,
+            additional_input=build_user_input(prompt, effective_image_attachments),
         )
     except ContextCompactionError as exc:
         duration_ms = int((time.time() - started_at) * 1000) if "started_at" in locals() else 0
@@ -175,7 +184,7 @@ async def run_prompt_once(
     prefix_token: Any | None = None
     try:
         prefix_token = set_current_cache_prefix_snapshot(prefix_snapshot)
-        run_input: Any = prompt
+        run_input: Any = build_user_input(prompt, effective_image_attachments)
         while True:
             result = Runner.run_streamed(
                 agent,
@@ -527,7 +536,7 @@ def _approval_server_name(raw_item: Any, tool_name: str) -> str:
 def _approval_action_kind(tool_name: str) -> str:
     if tool_name in {"Write", "Update"}:
         return "text_write"
-    if tool_name == "shell":
+    if tool_name in {"shell", "test_shell"}:
         return "command"
     if tool_name == "task_stop":
         return "background_task_control"

{deepy_cli-0.2.25 → deepy_cli-0.2.26}/src/deepy/prompts/system.py

@@ -66,10 +66,9 @@ Core rules:
   `subagent_explore` for broad read-only investigation, `subagent_reviewer` for
   focused review, and `subagent_tester` for reproduction or verification. Keep
   Deepy responsible for final synthesis and do not delegate tiny one-step work.
-- If a subagent reports `test_shell` `approval_required`, ask the user through
-  `AskUserQuestion` with the exact command, policy reason, and approval token.
-  Retry only the same command through the constrained `test_shell` path after
-  the user approves; do not broaden access to raw shell.
+- If a subagent runs `test_shell`, let medium-risk command approvals surface
+  through Deepy's audit flow. Do not rerun a blocked tester command through raw
+  `shell`; keep verification inside the constrained `test_shell` path.
 
 Tool protocol:
 Tool results are JSON strings: ok, name, output, error, metadata, awaitUserResponse.

{deepy_cli-0.2.25 → deepy_cli-0.2.26}/src/deepy/sessions/session.py

@@ -454,6 +454,14 @@ class DeepySession:
         increment_cache_generation: bool = False,
         reset_cache_generation: bool = False,
     ) -> None:
+        clear_usage_state = (
+            active_tokens is not None
+            and usage is None
+            and latest_context_window_tokens is None
+            and last_usage_tokens is None
+            and last_usage_record_count is None
+            and cache_usage is None
+        )
         with self._transaction() as conn:
             self._update_session_metadata(
                 conn,
@@ -475,6 +483,19 @@ class DeepySession:
                 increment_cache_generation=increment_cache_generation,
                 reset_cache_generation=reset_cache_generation,
             )
+            if clear_usage_state:
+                conn.execute(
+                    """
+                    update sessions
+                    set usage_json = null,
+                        latest_context_window_tokens = null,
+                        last_usage_tokens = null,
+                        last_usage_record_count = null,
+                        cache_usage_json = null
+                    where id = ?
+                    """,
+                    (self.session_id,),
+                )
 
     @contextmanager
     def _transaction(self) -> Iterator[sqlite3.Connection]:

{deepy_cli-0.2.25 → deepy_cli-0.2.26}/src/deepy/sessions/store_helpers.py

@@ -5,6 +5,7 @@ from typing import Any
 
 from deepy.todos import todo_state_from_tool_output
 from deepy.utils import json as json_utils
+from deepy.llm.multimodal import redacted_content_text
 
 CONTEXT_UNDERCOUNT_REPAIR_RATIO = 2
 CONTEXT_UNDERCOUNT_REPAIR_MIN_DELTA = 128
@@ -173,16 +174,8 @@ def session_status(items: list[dict[str, Any]]) -> str:
 def item_text(item: dict[str, Any]) -> str:
     for key in ("content", "text", "output"):
         value = item.get(key)
-        if isinstance(value, str):
-            return value
-        if isinstance(value, list):
-            parts: list[str] = []
-            for part in value:
-                if isinstance(part, dict):
-                    text = part.get("text") or part.get("input_text")
-                    if isinstance(text, str):
-                        parts.append(text)
-            return "".join(parts)
+        if value is not None:
+            return redacted_content_text(value)
     return ""