# OGhidra Configuration File
# This file contains environment variables for configuring the Ollama-GhidraMCP Bridge
# COPY THIS FILE TO .env AND EDIT THE VALUES


# ===== NECESSARY CONFIGURATIONS =====


# Active LLM Provider ('ollama', 'external', or 'custom_api')
LLM_PROVIDER=external

# Generic External API Configuration (Google, OpenAI, etc.)
EXTERNAL_PROVIDER=google
EXTERNAL_API_KEY=YOUR_API_KEY_HERE
EXTERNAL_MODEL=your-model-name
EXTERNAL_EMBEDDING_MODEL=gemini-embedding-001
EXTERNAL_TIMEOUT=120
EXTERNAL_TEMPERATURE=0.7
EXTERNAL_MAX_TOKENS=8192

# Ollama Configuration
OLLAMA_BASE_URL=http://localhost:11434/
OLLAMA_MODEL=your-model-name
OLLAMA_EMBEDDING_MODEL=nomic-embed-text
OLLAMA_TIMEOUT=300
OLLAMA_USERNAME=foo
OLLAMA_PASSWORD=bar

# ═══════════════════════════════════════════════════════════════════════════════
# CUSTOM API CONFIGURATION (OpenAI-compatible APIs)
# ═══════════════════════════════════════════════════════════════════════════════
# Set LLM_PROVIDER=custom_api to use this configuration
#
# TOKEN LIMITS - Adjust based on YOUR model:
# - Different models have different output limits (e.g., 4K, 8K, 16K, 32K, 64K, 128K)
# - Start with 4096. Increase if responses are truncated.
# - Check your model's documentation for specific limits
#
CUSTOM_API_URL=https://api.example.com/v1/chat/completions
CUSTOM_API_KEY=your-api-key-here
CUSTOM_API_MODEL=your-model-name
CUSTOM_API_EMBEDDING_MODEL=your-embedding-model
CUSTOM_API_TIMEOUT=300
CUSTOM_API_TEMPERATURE=0.7
CUSTOM_API_MAX_TOKENS=4096
CUSTOM_API_VERIFY_SSL=false
CUSTOM_API_SYSTEM_PROMPT=
CUSTOM_API_REQUEST_DELAY=0.0
CUSTOM_API_MAX_RETRIES=3

# Global throttling / concurrency control (helps prevent 429 bursts)
# - Concurrency=1 prevents overlapping requests during agentic loops.
# - Global min interval spaces request starts (in seconds).
CUSTOM_API_MAX_CONCURRENCY=1
CUSTOM_API_GLOBAL_MIN_INTERVAL=0.0

# Respect provider Retry-After header on 429/503
CUSTOM_API_RESPECT_RETRY_AFTER=true
CUSTOM_API_RETRY_AFTER_MAX_SECONDS=60

# Adaptive throttling (auto-increase on 429/503 retries, slowly relax on success)
CUSTOM_API_ADAPTIVE_THROTTLE_ENABLED=true
CUSTOM_API_ADAPTIVE_MAX_INTERVAL=10.0
CUSTOM_API_ADAPTIVE_INCREASE_FACTOR=1.5
CUSTOM_API_ADAPTIVE_DECREASE_FACTOR=0.9
CUSTOM_API_ADAPTIVE_SUCCESS_STREAK_THRESHOLD=10
CUSTOM_API_ADAPTIVE_JITTER_SECONDS=0.25


# ===== OPTIONAL CONFIGURATIONS =====

# Review Agent's level of approval
#REVIEW_THOROUGHNESS=basic    # Quick pass/fail
REVIEW_THOROUGHNESS=standard # Balanced (default)
#REVIEW_THOROUGHNESS=thorough # Comprehensive


# LLM Logging Configuration
# Enable comprehensive logging of all LLM interactions (prompts, responses, timing, tokens)
LLM_LOGGING_ENABLED=true
LLM_LOG_FILE=logs/llm_interactions.log
LLM_LOG_PROMPTS=true
LLM_LOG_RESPONSES=true
LLM_LOG_TOKENS=true
LLM_LOG_TIMING=true
LLM_LOG_FORMAT=json

# Options for LLM_LOG_FORMAT:
# - json: Structured JSON format (recommended for parsing and analysis)
# - text: Human-readable text format

# To disable specific logging components while keeping LLM_LOGGING_ENABLED=true:
# - LLM_LOG_PROMPTS=false (don't log prompts)
# - LLM_LOG_RESPONSES=false (don't log responses)
# - LLM_LOG_TOKENS=false (don't log token counts)
# - LLM_LOG_TIMING=false (don't log timing information)

# GhidraMCP Configuration
GHIDRA_BASE_URL=http://localhost:8080/
GHIDRA_TIMEOUT=30
GHIDRA_MOCK_MODE=false

MAX_EXECUTION_STEPS=5    # Increased from 10 to 20
# ===== AGENTIC CYCLES (Outer Loop) =====
MAX_AGENTIC_CYCLES=2        # Investigation cycles
AGENTIC_LOOP_ENABLED=true     # Enable adaptive re-planning


# Delay between ollama calls (some servers auto reject too many requests especially cloud hosting)
OLLAMA_REQUEST_DELAY=2.0

# ===== CONTEXT MANAGEMENT =====
# These settings control how much data the AI can process per request.
# All char limits scale proportionally to CONTEXT_BUDGET (chars ≈ tokens × 4).
# Adjust CONTEXT_BUDGET based on your model's context window size.
#
# SCALING GUIDE (based on CONTEXT_BUDGET):
# ┌─────────────────────┬──────────┬──────────┬──────────┬──────────┬──────────┬──────────┐
# │ Variable            │ 25K      │ 50K      │ 100K     │ 200K     │ 500K     │ 1M       │
# ├─────────────────────┼──────────┼──────────┼──────────┼──────────┼──────────┼──────────┤
# │ CURRENT_LOOP_MAX    │ 500      │ 1000     │ 2000     │ 4000     │ 10000    │ 20000    │
# │ PREV_LOOP_MAX       │ 100      │ 200      │ 400      │ 800      │ 2000     │ 4000     │
# │ OLDER_LOOP_MAX      │ 25       │ 50       │ 100      │ 200      │ 500      │ 1000     │
# │ MAX_DETAILED_STEPS  │ 5        │ 8        │ 10       │ 15       │ 20       │ 25       │
# └─────────────────────┴──────────┴──────────┴──────────┴──────────┴──────────┴──────────┘
# Formula: CURRENT_LOOP ≈ CONTEXT_BUDGET × 0.02 (2% of budget)

# Maximum context budget in tokens (adjust to your LLM's context window limit)
# Examples: 8000 for smaller models, 100000 for large context models, 200000+ for frontier models
CONTEXT_BUDGET=100000
# Fraction of context budget allocated for execution results (0.1-0.8)
CONTEXT_BUDGET_EXECUTION=0.4
# Enable LLM-based summarization for large results
ENABLE_RESULT_SUMMARIZATION=true
# Enable result caching with get_cached_result() references
RESULT_CACHE_ENABLED=true
# Enable tiered context compression (recent=detailed, older=summarized)
TIERED_CONTEXT_ENABLED=true
# Tiered context char limits (based on how old the loop is):
# - Current loop: full details up to this limit
CURRENT_LOOP_MAX_CHARS=2000
# - Previous loop (1 cycle ago): bullet-point summaries
PREV_LOOP_MAX_CHARS=400
# - Older loops (2+ cycles ago): one-line references only
OLDER_LOOP_MAX_CHARS=100
# Sliding window: number of recent steps kept in full detail
# Older steps are compressed to one-liners with cache hints
MAX_DETAILED_STEPS=10
# Hybrid Context Management
TOP_N_PER_CATEGORY=10
ENABLE_CORRELATION_HINTS=true
MIN_CORRELATION_MENTIONS=2
ENABLE_CYCLE_CONCLUSIONS=true

# ===== INTERACTIVE EXECUTION GATE =====
# pause execution at critical moments
EXECUTION_GATE_ENABLED=true
GATE_ON_ARTIFACT=true           # Pause when critical artifact found (crypto keys, credentials, etc.)
GATE_ON_REPETITION=true         # Pause on N identical tool calls (doom-loop detection)
                                 # Note: Investigation tools (xrefs, decompile, list_*) are exempt
                                 # from repetition detection to allow thorough analysis
GATE_REPETITION_THRESHOLD=3     # How many identical calls before triggering
GATE_ON_HIGH_RISK_TOOL=false    # Pause before destructive tools (rename)
GATE_AUTO_RESUME_TIMEOUT=0      # 0 = wait forever, >0 = auto-resume after N seconds

# ===== SESSION COMPACTION =====
# Smart context pruning to prevent overflow (empty responses from context saturation)
COMPACTION_ENABLED=true
COMPACTION_THRESHOLD=0.75       # Trigger at 75% context usage
COMPACTION_AUTO=true            # Auto-compact between agentic cycles