Paperfessor/.env.example at main · dunknowcoding/Paperfessor · GitHub

1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
# Paperfessor example environment file.
# Copy to .env and fill in. NEVER commit a real .env.

# Interface language (en, zh-CN, ja)
PAPERFESSOR_LANGUAGE=en

# LLM provider (minimax, openai, anthropic, google, ollama, llamacpp, custom)
PAPERFESSOR_PROVIDER=minimax

# Default model + per-agent overrides (all use the MiniMax-M3 default in dev)
PAPERFESSOR_MODEL=MiniMax-M3
PAPERFESSOR_PHD_MODEL=MiniMax-M3
PAPERFESSOR_MS_MODEL=MiniMax-M3
PAPERFESSOR_UG_MODEL=MiniMax-M3

# MiniMax direct API endpoint
PAPERFESSOR_BASE_URL=https://api.minimax.io/v1

# Research depth
PAPERFESSOR_DEPTH=normal

# Thinking mode (true to use MiniMax-M3's reasoning prefill)
PAPERFESSOR_THINKING_MODE=true

# Token budget
PAPERFESSOR_MAX_INPUT_TOKENS=1000000
PAPERFESSOR_DEFAULT_MAX_TOKENS=32768

# CLI display (minimal | default | vibrant)
PAPERFESSOR_DISPLAY_STYLE=default
PAPERFESSOR_DISPLAY_COLOR=auto

# Workspace root override (defaults: the project checkout when run
# from it, otherwise the per-user data directory)
#PAPERFESSOR_WORKSPACE=/path/to/your/project

# Optional REAL contact email, sent to OpenAlex for its "polite pool"
# (better API rate limits). Leave unset to stay anonymous.
#PAPERFESSOR_CONTACT_EMAIL=you@example.org

# ---- Coordination budgets (bound every agent loop) --------------------
# Raise for higher quality, lower for cheaper/faster runs.
PAPERFESSOR_MAX_METHOD_ROUNDS=3      # method-improvement rounds before abandoning
PAPERFESSOR_MAX_UG_ROUNDS=5          # UG implement -> verify -> fix attempts
PAPERFESSOR_MAX_SECTION_REDRAFTS=1   # supervisor-rejected section rewrites
PAPERFESSOR_MAX_INSPECTION_ROUNDS=3  # whole-paper self-inspection cycles
PAPERFESSOR_MAX_LLM_CALLS=85         # hard per-run LLM-call budget

# ---- Undergraduate permissions (DEFAULTS ARE PERMISSIVE) --------------
# The UG can install software and run local tools out of the box.
# Set any of these to false to lock the agent down. (LLM-generated
# model code always runs sandboxed regardless of these.)
PAPERFESSOR_UG_ALLOW_LOCAL_TOOLS=true  # run MATLAB/R/office/other local CLIs
PAPERFESSOR_UG_ALLOW_INSTALLS=true     # pip install packages (logged to src/tools)
PAPERFESSOR_UG_ALLOW_GPU=true          # use CUDA for heavy/speed-topic experiments
PAPERFESSOR_UG_TOOL_TIMEOUT_SECONDS=300
PAPERFESSOR_UG_SANDBOX_TIMEOUT_SECONDS=240

# ---- Paper output -----------------------------------------------------
PAPERFESSOR_PAPER_MAX_PAGES=9        # main-body page budget (venue may override)