/**
 * Pure logic for turning a chat history into a compressed "view":
 * canonical serialization (stable across turns, used for hashing and token
 * counting), rolling prefix hashes (content-addressed cache keys), round
 * boundaries (where history may be cut without orphaning tool results), and
 * chunk planning.
 */
import { createHash } from "crypto";

export interface CanonMessage {
  role: "system" | "user" | "assistant" | "tool";
  text: string;
  toolCalls?: { name: string; args: string }[];
  toolResults?: { content: string }[];
  /** Names of file/image attachments (content not included in the view). */
  files?: string[];
}

/** Prefix marking this plugin's own replies (e.g. /compress stats). */
export const SENTINEL = "[context-compressor]";

export function canon(m: CanonMessage): string {
  return JSON.stringify({
    role: m.role,
    text: m.text,
    toolCalls: (m.toolCalls ?? []).map((c) => ({ name: c.name, args: c.args })),
    toolResults: (m.toolResults ?? []).map((r) => ({ content: r.content })),
    files: m.files ?? [],
  });
}

/**
 * h[i] = hash chain covering msgs[0..i]. Same prefix => same hash. The seed
 * namespaces the chain: summaries produced under a different model, prompt
 * version, or settings must never be reused for one another.
 */
export function prefixHashes(msgs: CanonMessage[], seed = ""): string[] {
  const hashes: string[] = [];
  let prev = seed;
  for (const m of msgs) {
    prev = createHash("sha256").update(prev).update("\x00").update(canon(m)).digest("hex");
    hashes.push(prev);
  }
  return hashes;
}

/** Stable namespace seed derived from compaction-relevant settings. */
export function hashNamespace(parts: Record<string, unknown>): string {
  const stable = JSON.stringify(
    Object.keys(parts)
      .sort()
      .map((key) => [key, parts[key]]),
  );
  return createHash("sha256").update(stable).digest("hex");
}

/**
 * Valid cut indices c (meaning: summarize msgs[0..c), keep msgs[c..]).
 * A cut is valid before any non-tool message: cutting before a tool-result
 * message would orphan it from the assistant message that requested the
 * call. Agentic histories may contain a single user message followed by a
 * long assistant/tool loop, so user-only boundaries would be far too coarse.
 */
export function roundBoundaries(msgs: CanonMessage[]): number[] {
  const boundaries: number[] = [];
  for (let i = 1; i < msgs.length; i++) {
    if (msgs[i].role !== "tool") boundaries.push(i);
  }
  return boundaries;
}

/**
 * Largest valid cut that still keeps at least keepRecentTokens verbatim in
 * the tail. Returns 0 when no cut is allowed.
 */
export function maxAllowedCut(
  boundaries: number[],
  tokens: number[],
  keepRecentTokens: number,
): number {
  const suffix: number[] = new Array(tokens.length + 1).fill(0);
  for (let i = tokens.length - 1; i >= 0; i--) {
    suffix[i] = suffix[i + 1] + tokens[i];
  }
  let best = 0;
  for (const b of boundaries) {
    if (suffix[b] >= keepRecentTokens && b > best) best = b;
  }
  return best;
}

/**
 * Group rounds into summarization chunks of at most chunkTokens each
 * (a single round may exceed chunkTokens; it then forms its own chunk).
 * Returns ascending cut indices, each the end of one chunk, from `from`
 * (exclusive) up to `maxCut` (inclusive, if reachable).
 */
export function chunkEnds(
  boundaries: number[],
  tokens: number[],
  from: number,
  chunkTokens: number,
  maxCut: number,
): number[] {
  const cuts: number[] = [];
  const candidates = boundaries.filter((b) => b > from && b <= maxCut);
  let chunkStart = from;
  let chunkSize = 0;
  const roundTokens = (start: number, end: number) => {
    let sum = 0;
    for (let i = start; i < end; i++) sum += tokens[i];
    return sum;
  };
  let prev = from;
  for (const b of candidates) {
    const size = roundTokens(prev, b);
    if (chunkSize > 0 && chunkSize + size > chunkTokens) {
      cuts.push(prev);
      chunkStart = prev;
      chunkSize = 0;
    }
    chunkSize += size;
    if (chunkSize >= chunkTokens) {
      cuts.push(b);
      chunkStart = b;
      chunkSize = 0;
    }
    prev = b;
  }
  if (chunkSize > 0 && prev > chunkStart) cuts.push(prev);
  return cuts;
}

/**
 * Transform one message for the model's view: drop this plugin's own
 * replies (SENTINEL) and pure command messages entirely; for messages that
 * mix a command with real content, strip only the command line so the
 * user's actual instruction survives. Returns null to drop the message.
 */
export function transformForView(m: CanonMessage): CanonMessage | null {
  if (m.role === "assistant" && m.text.startsWith(SENTINEL)) return null;
  // Canceled/dead replies leave behind empty assistant messages — pure
  // noise that pollutes the history and the hash chain.
  if (
    m.role === "assistant" &&
    m.text.trim() === "" &&
    (m.toolCalls ?? []).length === 0 &&
    (m.toolResults ?? []).length === 0 &&
    (m.files ?? []).length === 0
  ) {
    return null;
  }
  if (!isCompressCommand(m) && !isUsageCommand(m)) return m;
  const allCommands = [...COMPRESS_COMMANDS, ...USAGE_COMMANDS];
  const stripped = m.text
    .split(/\r?\n/)
    .filter((line) => !allCommands.includes(line.trim().toLowerCase()))
    .join("\n");
  if (stripped.trim() === "") return null;
  return { ...m, text: stripped };
}

/**
 * Drop this plugin's own exchanges from the model's view; messages mixing a
 * command with real content keep the content (command line removed).
 */
export function filterCompressorExchanges(msgs: CanonMessage[]): CanonMessage[] {
  const out: CanonMessage[] = [];
  for (const m of msgs) {
    const transformed = transformForView(m);
    if (transformed !== null) out.push(transformed);
  }
  return out;
}

/** Commands that force compaction. */
export const COMPRESS_COMMANDS = ["/compress", "/compact"];

/** Commands that report context/token usage. */
export const USAGE_COMMANDS = ["/usage"];

/**
 * Other plugins' prompt preprocessors may inject context around what the
 * user typed, so a command counts when it stands alone on any line of a
 * user message.
 */
function matchesCommand(m: CanonMessage, commands: string[]): boolean {
  if (m.role !== "user") return false;
  return m.text
    .split(/\r?\n/)
    .some((line) => commands.includes(line.trim().toLowerCase()));
}

/** True when the message is a compress command. */
export function isCompressCommand(m: CanonMessage): boolean {
  return matchesCommand(m, COMPRESS_COMMANDS);
}

/** True when the message is the /usage command. */
export function isUsageCommand(m: CanonMessage): boolean {
  return matchesCommand(m, USAGE_COMMANDS);
}

/**
 * Chat templates (qwen's tool template among them) refuse to render a
 * conversation with zero user messages ("No user query found in messages").
 * Compaction can produce exactly that in tool-heavy chats, so the view must
 * check and inject a synthetic user message when needed.
 */
export function hasUserMessage(msgs: CanonMessage[]): boolean {
  return msgs.some((m) => m.role === "user");
}

/** Byte-stable synthetic user message for user-less compacted tails. */
export const SYNTHETIC_USER_MESSAGE =
  "Continue the task described in the summary above from where it left off.";

/**
 * Pure logic for turning a chat history into a compressed "view":
 * canonical serialization (stable across turns, used for hashing and token
 * counting), rolling prefix hashes (content-addressed cache keys), round
 * boundaries (where history may be cut without orphaning tool results), and
 * chunk planning.
 */
import { createHash } from "crypto";

export interface CanonMessage {
  role: "system" | "user" | "assistant" | "tool";
  text: string;
  toolCalls?: { name: string; args: string }[];
  toolResults?: { content: string }[];
  /** Names of file/image attachments (content not included in the view). */
  files?: string[];
}

/** Prefix marking this plugin's own replies (e.g. /compress stats). */
export const SENTINEL = "[context-compressor]";

export function canon(m: CanonMessage): string {
  return JSON.stringify({
    role: m.role,
    text: m.text,
    toolCalls: (m.toolCalls ?? []).map((c) => ({ name: c.name, args: c.args })),
    toolResults: (m.toolResults ?? []).map((r) => ({ content: r.content })),
    files: m.files ?? [],
  });
}

/**
 * h[i] = hash chain covering msgs[0..i]. Same prefix => same hash. The seed
 * namespaces the chain: summaries produced under a different model, prompt
 * version, or settings must never be reused for one another.
 */
export function prefixHashes(msgs: CanonMessage[], seed = ""): string[] {
  const hashes: string[] = [];
  let prev = seed;
  for (const m of msgs) {
    prev = createHash("sha256").update(prev).update("\x00").update(canon(m)).digest("hex");
    hashes.push(prev);
  }
  return hashes;
}

/** Stable namespace seed derived from compaction-relevant settings. */
export function hashNamespace(parts: Record<string, unknown>): string {
  const stable = JSON.stringify(
    Object.keys(parts)
      .sort()
      .map((key) => [key, parts[key]]),
  );
  return createHash("sha256").update(stable).digest("hex");
}

/**
 * Valid cut indices c (meaning: summarize msgs[0..c), keep msgs[c..]).
 * A cut is valid before any non-tool message: cutting before a tool-result
 * message would orphan it from the assistant message that requested the
 * call. Agentic histories may contain a single user message followed by a
 * long assistant/tool loop, so user-only boundaries would be far too coarse.
 */
export function roundBoundaries(msgs: CanonMessage[]): number[] {
  const boundaries: number[] = [];
  for (let i = 1; i < msgs.length; i++) {
    if (msgs[i].role !== "tool") boundaries.push(i);
  }
  return boundaries;
}

/**
 * Largest valid cut that still keeps at least keepRecentTokens verbatim in
 * the tail. Returns 0 when no cut is allowed.
 */
export function maxAllowedCut(
  boundaries: number[],
  tokens: number[],
  keepRecentTokens: number,
): number {
  const suffix: number[] = new Array(tokens.length + 1).fill(0);
  for (let i = tokens.length - 1; i >= 0; i--) {
    suffix[i] = suffix[i + 1] + tokens[i];
  }
  let best = 0;
  for (const b of boundaries) {
    if (suffix[b] >= keepRecentTokens && b > best) best = b;
  }
  return best;
}

/**
 * Group rounds into summarization chunks of at most chunkTokens each
 * (a single round may exceed chunkTokens; it then forms its own chunk).
 * Returns ascending cut indices, each the end of one chunk, from `from`
 * (exclusive) up to `maxCut` (inclusive, if reachable).
 */
export function chunkEnds(
  boundaries: number[],
  tokens: number[],
  from: number,
  chunkTokens: number,
  maxCut: number,
): number[] {
  const cuts: number[] = [];
  const candidates = boundaries.filter((b) => b > from && b <= maxCut);
  let chunkStart = from;
  let chunkSize = 0;
  const roundTokens = (start: number, end: number) => {
    let sum = 0;
    for (let i = start; i < end; i++) sum += tokens[i];
    return sum;
  };
  let prev = from;
  for (const b of candidates) {
    const size = roundTokens(prev, b);
    if (chunkSize > 0 && chunkSize + size > chunkTokens) {
      cuts.push(prev);
      chunkStart = prev;
      chunkSize = 0;
    }
    chunkSize += size;
    if (chunkSize >= chunkTokens) {
      cuts.push(b);
      chunkStart = b;
      chunkSize = 0;
    }
    prev = b;
  }
  if (chunkSize > 0 && prev > chunkStart) cuts.push(prev);
  return cuts;
}

/**
 * Transform one message for the model's view: drop this plugin's own
 * replies (SENTINEL) and pure command messages entirely; for messages that
 * mix a command with real content, strip only the command line so the
 * user's actual instruction survives. Returns null to drop the message.
 */
export function transformForView(m: CanonMessage): CanonMessage | null {
  if (m.role === "assistant" && m.text.startsWith(SENTINEL)) return null;
  // Canceled/dead replies leave behind empty assistant messages — pure
  // noise that pollutes the history and the hash chain.
  if (
    m.role === "assistant" &&
    m.text.trim() === "" &&
    (m.toolCalls ?? []).length === 0 &&
    (m.toolResults ?? []).length === 0 &&
    (m.files ?? []).length === 0
  ) {
    return null;
  }
  if (!isCompressCommand(m) && !isUsageCommand(m)) return m;
  const allCommands = [...COMPRESS_COMMANDS, ...USAGE_COMMANDS];
  const stripped = m.text
    .split(/\r?\n/)
    .filter((line) => !allCommands.includes(line.trim().toLowerCase()))
    .join("\n");
  if (stripped.trim() === "") return null;
  return { ...m, text: stripped };
}

/**
 * Drop this plugin's own exchanges from the model's view; messages mixing a
 * command with real content keep the content (command line removed).
 */
export function filterCompressorExchanges(msgs: CanonMessage[]): CanonMessage[] {
  const out: CanonMessage[] = [];
  for (const m of msgs) {
    const transformed = transformForView(m);
    if (transformed !== null) out.push(transformed);
  }
  return out;
}

/** Commands that force compaction. */
export const COMPRESS_COMMANDS = ["/compress", "/compact"];

/** Commands that report context/token usage. */
export const USAGE_COMMANDS = ["/usage"];

/**
 * Other plugins' prompt preprocessors may inject context around what the
 * user typed, so a command counts when it stands alone on any line of a
 * user message.
 */
function matchesCommand(m: CanonMessage, commands: string[]): boolean {
  if (m.role !== "user") return false;
  return m.text
    .split(/\r?\n/)
    .some((line) => commands.includes(line.trim().toLowerCase()));
}

/** True when the message is a compress command. */
export function isCompressCommand(m: CanonMessage): boolean {
  return matchesCommand(m, COMPRESS_COMMANDS);
}

/** True when the message is the /usage command. */
export function isUsageCommand(m: CanonMessage): boolean {
  return matchesCommand(m, USAGE_COMMANDS);
}

/**
 * Chat templates (qwen's tool template among them) refuse to render a
 * conversation with zero user messages ("No user query found in messages").
 * Compaction can produce exactly that in tool-heavy chats, so the view must
 * check and inject a synthetic user message when needed.
 */
export function hasUserMessage(msgs: CanonMessage[]): boolean {
  return msgs.some((m) => m.role === "user");
}

/** Byte-stable synthetic user message for user-less compacted tails. */
export const SYNTHETIC_USER_MESSAGE =
  "Continue the task described in the summary above from where it left off.";