// src/common/constants.ts

/**
 * Minimum thinking budget allowed by the Anthropic API.
 * Extended thinking requires at least 1024 budget_tokens.
 * @see https://docs.anthropic.com/en/docs/build-with-claude/extended-thinking
 */
export const MIN_THINKING_BUDGET = 1024;

/**
 * Reduced thinking budget for subsequent ReAct iterations (tool-result turns).
 *
 * In a ReAct agent loop, the first LLM call processes the user's query and
 * may need deep reasoning. Subsequent iterations (after tool results return)
 * typically only need to decide "call next tool" or "generate final response"
 * — 1024 tokens is sufficient for this routing logic.
 *
 * This reduces wall-clock time per iteration from ~20-30s to ~5-10s,
 * compounding across multi-tool conversations (e.g., 10 tool calls).
 */
export const TOOL_TURN_THINKING_BUDGET = 1024;

// ============================================================================
// CONTEXT OVERFLOW MANAGEMENT
//
// Context overflow is handled mechanically — no token budget numbers are
// exposed to the LLM. The system uses: pruning (Graph), summarization
// (summarizeCallback), and auto-continuation (client.js max_tokens detection).
//
// See: docs/context-overflow-architecture.md
// ============================================================================

/**
 * Minimum number of attached documents before the multi-document delegation
 * hint is injected. Below this threshold, the agent processes documents
 * directly within its own context.
 */
export const MULTI_DOCUMENT_THRESHOLD = 3;

/**
 * Context utilization safety buffer multiplier (0-1).
 * Applied as: effectiveMax = (maxContextTokens - maxOutputTokens) * CONTEXT_SAFETY_BUFFER
 *
 * Reserves headroom so the LLM doesn't hit hard token limits mid-generation.
 * 0.9 = 10% reserved for safety.
 */
export const CONTEXT_SAFETY_BUFFER = 0.9;

// ============================================================================
// SUMMARIZATION CONFIGURATION DEFAULTS
//
// These constants provide sensible defaults for the SummarizationConfig.
// They can be overridden per-agent via AgentInputs.summarizationConfig.
// ============================================================================

/**
 * Default context utilization percentage (0-100) at which summarization triggers.
 * When the context window is ≥80% full, pruning + summarization activates.
 */
export const SUMMARIZATION_CONTEXT_THRESHOLD = 80;

/**
 * Proactive summarization threshold (0-1 fraction of context window).
 * At this utilization level, background summarization fires BEFORE pruning is needed.
 * This gives the summary time to complete so it's ready when context actually fills up.
 *
 * Inspired by VS Code Copilot Chat's 3-tier strategy:
 *   80% → proactive background summary
 *   90% → pruning kicks in (with summary already cached)
 *  100% → graceful: use existing summary + recent messages, never block
 */
export const PROACTIVE_SUMMARY_THRESHOLD = 0.8;

/**
 * Number of recent conversation rounds (human+AI pairs) to keep in the
 * windowed view when a summary is available. Everything older is covered
 * by the summary. 2 rounds = last 2 user questions + 2 AI responses.
 *
 * This prevents wasting tokens on raw messages the summary already covers
 * and keeps context tight for the LLM.
 */
export const COMPACTION_RECENT_ROUNDS = 2;

/**
 * Default reserve ratio (0-1) — fraction of context window to preserve as recent messages.
 * 0.3 means 30% of the context budget is reserved for the most recent messages,
 * ensuring the model always has immediate conversation history even after aggressive pruning.
 */
export const SUMMARIZATION_RESERVE_RATIO = 0.3;

/**
 * Default EMA (Exponential Moving Average) alpha for pruning calibration.
 * Controls how quickly the calibration adapts to new token counts.
 * Higher α = faster adaptation (more responsive to recent changes).
 * Lower α = smoother adaptation (more stable across iterations).
 * 0.3 provides a balance between responsiveness and stability.
 */
export const PRUNING_EMA_ALPHA = 0.3;

/**
 * Default initial calibration ratio for EMA pruning.
 * 1.0 means no adjustment on the first iteration (trust the raw token counts).
 * Subsequent iterations will adjust based on actual vs. estimated token usage.
 */
export const PRUNING_INITIAL_CALIBRATION = 1.0;

// ============================================================================
// TOOL DISCOVERY CACHING
// ============================================================================

/**
 * Maximum number of tool discovery entries to cache per conversation.
 * Prevents unbounded memory growth in very long conversations.
 */
export const TOOL_DISCOVERY_CACHE_MAX_SIZE = 200;

// ============================================================================
// MESSAGE DEDUPLICATION
// ============================================================================

/**
 * Maximum length of system message content to hash for deduplication.
 * Messages longer than this are always considered unique (hashing would be expensive).
 */
export const DEDUP_MAX_CONTENT_LENGTH = 10000;