"""deeppresenter 全局常量定义"""

import logging
import os
from pathlib import Path

# ============ Path ============
PACKAGE_DIR = Path(__file__).parent.parent

# ============ Logging ===========
LOGGING_LEVEL = int(os.getenv("DEEPPRESENTER_LOG_LEVEL", logging.INFO))
MAX_LOGGING_LENGTH = int(os.getenv("DEEPPRESENTER_MAX_LOGGING_LENGTH", 1024))

# ============ Agent  ============
RETRY_TIMES = int(os.getenv("RETRY_TIMES", 10))
MAX_TOOLCALL_PER_TURN = int(os.getenv("MAX_TOOLCALL_PER_TURN", 7))
MAX_RETRY_INTERVAL = int(os.getenv("MAX_RETRY_INTERVAL", 60))
# count in chars, this is about the first 4 page of a dual-column paper
TOOL_CUTOFF_LEN = int(os.getenv("TOOL_CUTOFF_LEN", 4096))
MAX_SUBAGENT_TURNS = int(os.getenv("MAX_SUBAGENT_TURNS", 10))
# count in tokens
CONTEXT_LENGTH_LIMIT = int(os.getenv("CONTEXT_LENGTH_LIMIT", 200_000))
CUTOFF_WARNING = "NOTE: Output truncated (showing first {line} lines). Use `read_file` with `offset` parameter to continue reading from {resource_id}."

# ============ Environment ============
PIXEL_MULTIPLE = int(os.getenv("PIXEL_MULTIPLE", 16))
MCP_CONNECT_TIMEOUT = int(os.getenv("MCP_CONNECT_TIMEOUT", 120))
MCP_CALL_TIMEOUT = int(os.getenv("MCP_CALL_TIMEOUT", 1800))
WORKSPACE_BASE = Path(
    os.getenv(
        "DEEPPRESENTER_WORKSPACE_BASE",
        str(Path.home() / ".cache/deeppresenter"),
    )
)
TOOL_CACHE = PACKAGE_DIR / ".tools.json"

GLOBAL_ENV_LIST = [
    "HTTP_PROXY",
    "HTTPS_PROXY",
    "NO_PROXY",
    "ALL_PROXY",
    "http_proxy",
    "https_proxy",
    "no_proxy",
    "all_proxy",
    "PYTHONWARNINGS",
]

# ============ Webview ============
PDF_OPTIONS = {
    "print_background": True,
    "landscape": False,
    "margin": {"top": "0mm", "right": "0mm", "bottom": "0mm", "left": "0mm"},
    "prefer_css_page_size": False,
    "display_header_footer": False,
    "scale": 1,
    "page_ranges": "1",
}

# ============ Additional Agent Prompt ===========

AGENT_PROMPT = """
<Environment>
Current time: {time}
Working directory: {workspace}
Platform: Debian Linux container

Pre-installed tools:
- Python 3.13, Node.js, imagemagic, mermaid-cli (mmdc), curl, wget, and other common utilities
- python-pptx, matplotlib, plotly, and other common packages
You can freely install any required tools, packages, or command-line utilities to complete the task
</Environment>

<Task Guidelines>
- Exploration Principle: A warning is issued at 10% remaining computation budget, Until then, explore thoroughly and give your best effort.
- Max Length: Tool Call Output exceeding {cutoff_len} characters will be truncated at the preceding line break. Full content is saved locally and accessible via `read_file` with `offset`.
- Tool Call Principle:
    1. Every response must include reasoning content and a valid tool call.
    2. All tool calls are processed in parallel; do not emit tool calls with interdependencies in the same turn.
- Toolcall Limit: You can calling up to {max_toolcall_per_turn} tools per turn.
</Task Guidelines>
"""

# Long-context understanding and multi-perspective retrieval
MA_RESEACHER_PROMPT = """
<Guide on Subagents>
You can use subagents to execute multiple complex tasks in parallel. They have the same capabilities as you, but start with empty context.
The subagent tool accepts a minimal `task` and a `context_file`.
Before calling the subagent, write the complete delegation brief to a local file yourself.
Put the complete background, source paths, constraints, expected deliverables, and handoff format into that file.
Keep `task` short and action-oriented.
In general, you should use subagents in scenarios that can be parallelized at scale without information loss. For example:
1. Long-document understanding: for a document with 20,000 lines, you can assign each agent 1,000 lines and launch 20 subagents in parallel.
2. Multi-perspective retrieval: analyze one subject from multiple aspects, such as a car's exterior design, configuration and pricing, and development history.
</Guide on Subagents>
"""

# Generate multiple pages in parallel after defining the global CSS
MA_RRESENTER_PROMPT = """
<Guide on Subagents>
You can use subagents to execute multiple complex tasks in parallel. They have the same capabilities as you, but start with empty context.
Therefore, you should first define a global visual theme, including a detailed design specification such as the background and accent colors.
Then, distribute the generation of each slide draft to different subagents.
The subagent tool accepts a minimal `task` and a `context_file`.
Before calling the subagent, write the shared visual system, manuscript excerpt, slide scope, constraints, and handoff requirements into a local file.
Keep `task` as a short action such as "Generate slide 1 according to the global visual system".
</Guide on Subagents>
"""


OFFLINE_PROMPT = """
<Offline Mode>
- You are operating in offline mode without internet access. All network-dependent tools have been removed.
- Focus on the available tools and adjust your plan accordingly.
</Offline Mode>
"""

CONTEXT_MODE_PROMPT = """
<Context Mode>
- You are operating in limited working context. When approaching the limit, you will be asked to compact history into a local summary, then continue.
- To minimize information loss, save files, images, and intermediate results immediately after generation or retrieval—do not defer.
- After compaction, only the first few messages, recent messages, and your saved summary will remain—all other context will be discarded.
</Context Mode>
"""


HALF_BUDGET_NOTICE_MSG = {
    "text": "<NOTICE>You have used about half of your working budget. Now focused on the core task and skipping unnecessary steps or explorations.</NOTICE>",
    "type": "text",
}
URGENT_BUDGET_NOTICE_MSG = {
    "text": "<URGENT>Working budget nearly exhausted. You must finish the core task and call `finalize` now, or your work will fail. Skip extras like inspection and validation.</URGENT>",
    "type": "text",
}
HIST_LOST_MSG = {
    "text": "<NOTICE>History between this point and the following message has been compacted into a summary</NOTICE>",
    "type": "text",
}

CONTINUE_MSG = {
    "text": "<NOTICE>History has been compacted. Refer to the saved summary and continue your work</NOTICE>",
    "type": "text",
}

LAST_ITER_MSG = {
    "text": "<URGENT>Working budget nearly exhausted. You must finish the core task and call `finalize` now, or your work will fail. Skip extras like inspection and validation.</URGENT>",
    "type": "text",
}

MEMORY_COMPACT_MSG = """
You have reached the context length limit for this conversation. Immediately extract key information from the tool interaction history, generate a complete state summary, and save it to the working directory to ensure seamless continuation in subsequent conversations.

<summary_requirements>
All information must be recorded with specific details. Do not use references like "as mentioned above" or "see previous section". Only extract information from tool interactions in the current session; do not record information provided by the user or system instructions.

1. Collected Information & Data
   - Factual data, evidence, research findings
   - Key source materials and references

2. Uncertainties & Open Issues
   - Information gaps, unverified assumptions, identified limitations

3. Generated Artifacts
   - Intermediate files, code, images/charts: path + purpose

4. Next Steps
   - Work completed so far and outcomes achieved
   - Remaining tasks and suggested execution order
   - Pending items that were planned/started but not yet finished

5. Lessons Learned (if applicable)
   - Issues encountered during tool calls and their solutions
   - Operations to avoid
</summary_requirements>

<important>
- Use {language} as the primary language. The summary must be detailed enough for any successor to fully understand current progress and continue work without reviewing history.
- Complete the summary generation in this turn. Do not plan multi-turn generation, or history will be lost. Save directly to the working directory.
</important>
"""