/**
 * Rolling summarization — when a conversation's estimated token count
 * approaches the model's context window, summarize older turns and
 * inject the summary as a single synthetic system-role message.
 */
import { getContextWindow } from './cost-estimator'

// ~4 chars per token, approximate
function estimateTokens(text: string): number { return Math.ceil((text || '').length / 4) }

export function estimateTotalTokens(messages: any[]): number {
  let total = 0
  for (const m of messages) {
    if (Array.isArray(m.content)) {
      for (const b of m.content) {
        if (b.text) total += estimateTokens(b.text)
        else if (b.image) total += 1500 // rough image token
        else if (b.toolUse) total += estimateTokens(JSON.stringify(b.toolUse.input))
        else if (b.toolResult?.content) total += estimateTokens(JSON.stringify(b.toolResult.content))
      }
    } else if (typeof m.content === 'string') total += estimateTokens(m.content)
  }
  return total
}

/**
 * If total exceeds threshold, return a new array where the first N
 * messages are collapsed into a single summary.
 * The summarization itself is delegated to the caller (e.g. via sub-agent).
 */
export function shouldSummarize(messages: any[], model: string, thresholdPct = 70): boolean {
  const tokens = estimateTotalTokens(messages)
  const window = getContextWindow(model)
  return (tokens * 100 / window) > thresholdPct
}

/** Create a plain-text summary of older messages for compression. */
export function summaryPrompt(olderMessages: any[]): string {
  let txt = 'Summarize this conversation concisely (under 500 words). Keep key facts, decisions, and unresolved questions.\n\n'
  for (const m of olderMessages) {
    const role = m.role === 'user' ? 'User' : 'Assistant'
    const content = Array.isArray(m.content)
      ? m.content.map((b: any) => b.text || (b.toolUse ? `[tool:${b.toolUse.name}]` : '')).filter(Boolean).join(' ')
      : String(m.content || '')
    txt += `${role}: ${content.slice(0, 1000)}\n\n`
  }
  return txt
}