muonroi-cli 1.4.0 → 1.4.1

package/src/orchestrator/compaction.ts

@@ -1,626 +0,0 @@
-import { generateText, type ModelMessage } from "ai";
-import { isMetaAnalysisPrompt } from "../pil/layer6-output.js";
-import { getProviderCapabilities } from "../providers/capabilities.js";
-import type { ProviderFactory as LegacyProvider } from "../providers/runtime.js";
-import { requireRuntimeProvider, resolveModelRuntime } from "../providers/runtime.js";
-import { containsEncryptedReasoning } from "./reasoning";
-import { countTokens } from "./token-counter.js";
-
-export interface CompactionSettings {
-  reserveTokens: number;
-  keepRecentTokens: number;
-}
-
-export interface CutPointResult {
-  firstKeptIndex: number;
-  turnStartIndex: number;
-  isSplitTurn: boolean;
-}
-
-export interface PreparedCompaction {
-  previousSummary?: string;
-  messagesToSummarize: ModelMessage[];
-  turnPrefixMessages: ModelMessage[];
-  keptMessages: ModelMessage[];
-  firstKeptIndex: number;
-  isSplitTurn: boolean;
-  tokensBefore: number;
-  settings: CompactionSettings;
-}
-
-const TOOL_RESULT_MAX_CHARS = 8000;
-const MIN_KEPT_TOKENS_ON_RETRY = 4000;
-
-export const DEFAULT_RESERVE_TOKENS = 16_384;
-export const DEFAULT_KEEP_RECENT_TOKENS = 20_000;
-export const POST_TURN_MIN_TOKENS = 2_000;
-export const COMPACTION_MAX_OUTPUT_TOKENS = 4_096;
-export const TOOL_RESULT_MAX_CHARS_CONFIGURABLE = 8000;
-export const COMPACTION_SUMMARY_HEADER = "[Context checkpoint summary]";
-
-const SUMMARIZATION_SYSTEM_PROMPT = `You are a context summarization assistant.
-
-Do not continue the conversation. Do not answer any questions from the conversation.
-Only output a structured checkpoint summary that another coding agent can use to continue the work.
-
-CRITICAL: The agent reading this summary must NEVER redo completed work. Mark completed items clearly.`;
-
-const SUMMARIZATION_PROMPT = `The messages above are a conversation to summarize. Create a structured context checkpoint summary that another LLM will use to continue the work.
-
-Use this exact format:
-
-## Goal
-[What the user is trying to accomplish]
-
-## Active Plan
-- **Plan file**: [path to PLAN.md or plan document being followed, or "(none)" if no explicit plan]
-- **Plan step**: [current step number and description, e.g. "Step 3/7: implement layer6-output.ts", or "(none)"]
-- **Plan source**: [e.g. "/gsd:execute-phase", "/gsd:quick", "manual", or "(none)"]
-
-## Constraints & Preferences
-- [Requirements, preferences, or constraints]
-- [(none) if none were mentioned]
-
-## Progress
-### ✅ DONE — DO NOT REDO
-- [x] [Completed work — be specific: file changed, test passed, command ran]
-
-### 🔄 In Progress (resume here)
-- [ ] [Current work with exact state: e.g. "editing src/foo.ts — halfway through refactor"]
-
-### ❌ Blocked
-- [Any active blockers with root cause if known]
-
-## Session Notes
-- [Important observations, warnings, or gotchas discovered during the session]
-- [(none) if not applicable]
-
-## Key Decisions
-- **[Decision]**: [Rationale]
-
-## Next Steps
-1. [The very next action to take — be specific enough that the agent can act without re-reading the conversation]
-
-## Critical Context
-- [Important details needed to continue]
-- [(none) if not applicable]
-
-## Critical Data
-- **File paths mentioned**: [list all file paths exactly as they appear]
-- **Function/method names**: [list all function/method names]
-- **Error messages**: [copy verbatim any error messages]
-- **Key numbers/results**: [list important numbers, query results, or metrics]
-
-## Kept Artifacts (high-fidelity tool outputs preserved verbatim across B3/B4 compaction per anti-mù; agent can query the ee_query tool with "tool-artifact id=..." or "full tool result id=..." for full content later)
-- toolCallId | toolName | first120chars | reason (high-value|KEEP_TOOL_IDS|recent-turn|error)
-
-Keep it concise, but preserve exact file paths, function names, and error messages.`;
-
-const UPDATE_SUMMARIZATION_PROMPT = `The messages above are new conversation messages to incorporate into the existing summary provided below.
-
-Update the existing structured summary with new information. Rules:
-- Preserve still-relevant information from the previous summary
-- Add new progress, decisions, and critical context
-- Move completed items from "In Progress" to "Done" (✅ DONE — DO NOT REDO) when appropriate
-- Update "Active Plan" step if the current step has advanced
-- Update "Next Steps" based on the current state
-- Append new observations to "Session Notes"
-- Preserve exact file paths, function names, and error messages
-
-## Critical Data
-- Preserve exact file paths, function names, error messages (verbatim), and key numbers/results
-
-## Kept Artifacts
-- (carry forward or add new high-fidelity toolCallId | tool | reason from this increment)
-
-Use the exact same section structure as the existing summary format.`;
-
-const TURN_PREFIX_SUMMARIZATION_PROMPT = `This is the early prefix of a single turn that was too large to keep in full. The recent suffix is still available.
-
-Summarize only what is needed so another coding agent can understand the retained suffix.
-CRITICAL: List all completed steps explicitly so the agent does not redo them.
-
-Use this exact format:
-
-## Original Request
-[What the user asked for in this turn]
-
-## Early Progress — COMPLETED (DO NOT REDO)
-- [x] [Each completed step before the suffix — be specific: file edited, test ran, etc.]
-
-## Current State
-- [Exact state at the cut point: what was just done, what variable/file is mid-edit]
-
-## Context For Suffix
-- [Information needed to understand the kept recent messages]`;
-
-function isRecord(value: unknown): value is Record<string, unknown> {
-  return typeof value === "object" && value !== null;
-}
-
-function getTextParts(content: unknown): string[] {
-  if (typeof content === "string") return [content];
-  if (!Array.isArray(content)) return [];
-
-  const parts: string[] = [];
-  for (const part of content) {
-    if (!isRecord(part)) continue;
-    if (part.type === "text" && typeof part.text === "string") {
-      parts.push(part.text);
-      continue;
-    }
-    if (part.type === "reasoning" && typeof part.text === "string") {
-      if (!containsEncryptedReasoning(part.text)) {
-        parts.push(part.text);
-      }
-      continue;
-    }
-    if (part.type === "reasoning" && typeof part.reasoning === "string") {
-      if (!containsEncryptedReasoning(part.reasoning)) {
-        parts.push(part.reasoning);
-      }
-    }
-  }
-  return parts;
-}
-
-function stringifyForSummary(value: unknown): string {
-  if (typeof value === "string") return value;
-  try {
-    return JSON.stringify(value, null, 2);
-  } catch {
-    return String(value);
-  }
-}
-
-function truncateForSummary(text: string, maxChars = TOOL_RESULT_MAX_CHARS): string {
-  if (text.length <= maxChars) return text;
-  // Keep head + tail. Errors, exit codes, and summaries usually live at the end of long
-  // tool outputs; head-only truncation routinely drops the most diagnostic lines.
-  const headChars = Math.floor(maxChars * 0.6);
-  const tailChars = maxChars - headChars;
-  const omitted = text.length - maxChars;
-  return `${text.slice(0, headChars)}\n\n[... ${omitted} more characters truncated ...]\n\n${text.slice(-tailChars)}`;
-}
-
-function hashString(s: string): string {
-  let h = 0;
-  for (let i = 0; i < s.length; i++) {
-    h = ((h << 5) - h + s.charCodeAt(i)) | 0;
-  }
-  return h.toString(36);
-}
-
-const DEDUP_MIN_CHARS = 500;
-
-/**
- * Dedup identical tool-result payloads inside a slice destined for summarization.
- * Repeated reads of an unchanged file inflate the summarize prompt with
- * (truncated) duplicates that cost tokens and add zero new context.
- */
-function dedupToolResultsWithState(messages: ModelMessage[], seenHashes: Set<string>): ModelMessage[] {
-  return messages.map((msg) => {
-    if (msg.role !== "tool" || !Array.isArray(msg.content)) return msg;
-    const parts = msg.content as unknown[];
-    const newContent = parts.map((part) => {
-      if (!isRecord(part) || part.type !== "tool-result") return part;
-      const serialized = stringifyForSummary((part as Record<string, unknown>).output);
-      if (serialized.length < DEDUP_MIN_CHARS) return part;
-      const hash = hashString(serialized);
-      if (seenHashes.has(hash)) {
-        const toolName =
-          typeof (part as Record<string, unknown>).toolName === "string"
-            ? ((part as Record<string, unknown>).toolName as string)
-            : "tool";
-        return {
-          ...(part as Record<string, unknown>),
-          output: `[Identical to earlier ${toolName} result; ${serialized.length} chars elided]`,
-        };
-      }
-      seenHashes.add(hash);
-      return part;
-    });
-    return { ...msg, content: newContent as typeof msg.content };
-  });
-}
-
-export function extractUserContent(content: unknown): string {
-  if (typeof content === "string") return content;
-  if (!Array.isArray(content)) return "";
-
-  const parts: string[] = [];
-  for (const part of content) {
-    if (!isRecord(part)) continue;
-    if (part.type === "text" && typeof part.text === "string") {
-      parts.push(part.text);
-      continue;
-    }
-    if (part.type === "image") {
-      parts.push("[Image]");
-      continue;
-    }
-    if (part.type === "file") {
-      const filename = typeof part.filename === "string" ? part.filename : null;
-      parts.push(filename ? `[File: ${filename}]` : "[File]");
-    }
-  }
-  return parts.join("\n");
-}
-
-function extractAssistantText(content: unknown): string {
-  return getTextParts(content).join("\n");
-}
-
-function extractToolCallText(content: unknown): string[] {
-  if (!Array.isArray(content)) return [];
-
-  const toolCalls: string[] = [];
-  for (const part of content) {
-    if (!isRecord(part) || part.type !== "tool-call") continue;
-    const toolName = typeof part.toolName === "string" ? part.toolName : "tool";
-    const input = isRecord(part.input) ? part.input : {};
-    const args = Object.entries(input)
-      .map(([key, value]) => `${key}=${JSON.stringify(value)}`)
-      .join(", ");
-    toolCalls.push(`${toolName}(${args})`);
-  }
-  return toolCalls;
-}
-
-function extractToolResultText(content: unknown): string[] {
-  if (!Array.isArray(content)) return [];
-
-  const toolResults: string[] = [];
-  for (const part of content) {
-    if (!isRecord(part) || part.type !== "tool-result") continue;
-    toolResults.push(truncateForSummary(stringifyForSummary(part.output)));
-  }
-  return toolResults;
-}
-
-export function createCompactionSummaryMessage(summary: string): ModelMessage {
-  return {
-    role: "system",
-    content: `${COMPACTION_SUMMARY_HEADER}\n${summary.trim()}`,
-  };
-}
-
-export function isCompactionSummaryMessage(message: ModelMessage | undefined): boolean {
-  return message?.role === "system" && typeof message.content === "string"
-    ? message.content.startsWith(COMPACTION_SUMMARY_HEADER)
-    : false;
-}
-
-export function getCompactionSummaryText(message: ModelMessage | undefined): string | null {
-  if (!isCompactionSummaryMessage(message) || typeof message?.content !== "string") {
-    return null;
-  }
-  return message.content.slice(COMPACTION_SUMMARY_HEADER.length).trim();
-}
-
-function messageToString(message: ModelMessage): string {
-  switch (message.role) {
-    case "user":
-      return extractUserContent(message.content);
-    case "assistant": {
-      const text = extractAssistantText(message.content);
-      const toolCalls = extractToolCallText(message.content).join("; ");
-      return toolCalls ? `${text}\n${toolCalls}` : text;
-    }
-    case "tool":
-      return extractToolResultText(message.content).join("\n");
-    case "system":
-      return typeof message.content === "string" ? message.content : getTextParts(message.content).join("\n");
-    default:
-      return stringifyForSummary((message as { content?: unknown }).content);
-  }
-}
-
-export function estimateMessageTokens(message: ModelMessage): number {
-  return countTokens(messageToString(message));
-}
-
-export function estimateConversationTokens(systemPrompt: string, messages: ModelMessage[], inFlightText = ""): number {
-  const systemTokens = countTokens(systemPrompt) + (inFlightText ? countTokens(inFlightText) : 0);
-  return systemTokens + messages.reduce((sum, message) => sum + estimateMessageTokens(message), 0);
-}
-
-export function shouldCompactContext(
-  contextTokens: number,
-  contextWindow: number,
-  settings: CompactionSettings,
-): boolean {
-  return contextTokens > contextWindow - settings.reserveTokens;
-}
-
-function isValidCutPoint(message: ModelMessage): boolean {
-  return message.role !== "tool";
-}
-
-function findTurnStartIndex(messages: ModelMessage[], entryIndex: number, startIndex: number): number {
-  for (let i = entryIndex; i >= startIndex; i--) {
-    if (messages[i]?.role === "user") {
-      return i;
-    }
-  }
-  return -1;
-}
-
-export function findCutPoint(messages: ModelMessage[], startIndex: number, keepRecentTokens: number): CutPointResult {
-  const cutPoints: number[] = [];
-  for (let i = startIndex; i < messages.length; i++) {
-    if (isValidCutPoint(messages[i])) {
-      cutPoints.push(i);
-    }
-  }
-
-  if (cutPoints.length === 0) {
-    return { firstKeptIndex: startIndex, turnStartIndex: -1, isSplitTurn: false };
-  }
-
-  let accumulatedTokens = 0;
-  let cutIndex = cutPoints[0];
-
-  for (let i = messages.length - 1; i >= startIndex; i--) {
-    accumulatedTokens += estimateMessageTokens(messages[i]);
-    if (accumulatedTokens >= keepRecentTokens) {
-      cutIndex = cutPoints.find((index) => index >= i) ?? cutPoints[cutPoints.length - 1];
-      break;
-    }
-  }
-
-  const cutMessage = messages[cutIndex];
-  const isUserMessage = cutMessage?.role === "user";
-  const turnStartIndex = isUserMessage ? -1 : findTurnStartIndex(messages, cutIndex, startIndex);
-
-  return {
-    firstKeptIndex: cutIndex,
-    turnStartIndex,
-    isSplitTurn: !isUserMessage && turnStartIndex !== -1,
-  };
-}
-
-export function prepareCompaction(
-  messages: ModelMessage[],
-  systemPrompt: string,
-  settings: CompactionSettings,
-): PreparedCompaction | null {
-  const previousSummary = getCompactionSummaryText(messages[0]) ?? undefined;
-  const boundaryStart = previousSummary ? 1 : 0;
-  if (boundaryStart >= messages.length) {
-    return null;
-  }
-
-  const cutPoint = findCutPoint(messages, boundaryStart, settings.keepRecentTokens);
-  const historyEnd = cutPoint.isSplitTurn ? cutPoint.turnStartIndex : cutPoint.firstKeptIndex;
-  const rawHistory = messages.slice(boundaryStart, Math.max(boundaryStart, historyEnd));
-  const rawTurnPrefix = cutPoint.isSplitTurn ? messages.slice(cutPoint.turnStartIndex, cutPoint.firstKeptIndex) : [];
-  // Shared dedup state — second occurrence of a duplicate tool-result gets stubbed
-  // even when the first occurrence is in the history slice and the second in the prefix.
-  const dedupState = new Set<string>();
-  const messagesToSummarize = dedupToolResultsWithState(rawHistory, dedupState);
-  const turnPrefixMessages = dedupToolResultsWithState(rawTurnPrefix, dedupState);
-  // keptMessages stay verbatim — they live in the active conversation, not the summarize prompt.
-  const keptMessages = messages.slice(cutPoint.firstKeptIndex);
-  const tokensBefore = estimateConversationTokens(systemPrompt, messages);
-
-  if (keptMessages.length === 0) {
-    return null;
-  }
-
-  if (messagesToSummarize.length === 0 && turnPrefixMessages.length === 0) {
-    return null;
-  }
-
-  return {
-    previousSummary,
-    messagesToSummarize,
-    turnPrefixMessages,
-    keptMessages,
-    firstKeptIndex: cutPoint.firstKeptIndex,
-    isSplitTurn: cutPoint.isSplitTurn,
-    tokensBefore,
-    settings,
-  };
-}
-
-export function relaxCompactionSettings(settings: CompactionSettings): CompactionSettings {
-  return {
-    ...settings,
-    keepRecentTokens: Math.max(MIN_KEPT_TOKENS_ON_RETRY, Math.floor(settings.keepRecentTokens / 2)),
-  };
-}
-
-export function serializeConversation(messages: ModelMessage[]): string {
-  const parts: string[] = [];
-
-  for (const message of messages) {
-    if (isCompactionSummaryMessage(message)) {
-      const summary = getCompactionSummaryText(message);
-      if (summary) {
-        parts.push(`[Previous summary]: ${summary}`);
-      }
-      continue;
-    }
-
-    if (message.role === "user") {
-      const content = extractUserContent(message.content).trim();
-      if (content) parts.push(`[User]: ${content}`);
-      continue;
-    }
-
-    if (message.role === "assistant") {
-      const text = extractAssistantText(message.content).trim();
-      const toolCalls = extractToolCallText(message.content);
-      if (text) parts.push(`[Assistant response]: ${text}`);
-      if (toolCalls.length > 0) parts.push(`[Tool call]: ${toolCalls.join("; ")}`);
-      continue;
-    }
-
-    if (message.role === "tool") {
-      if (Array.isArray(message.content)) {
-        for (const part of message.content) {
-          if (!isRecord(part)) continue;
-          const r = part as Record<string, unknown>;
-          if (r.type !== "tool-result") continue;
-          const toolName = typeof r.toolName === "string" ? r.toolName : "unknown";
-          const result = truncateForSummary(stringifyForSummary(r.output));
-          if (result.trim()) parts.push(`[Tool result from ${toolName}]: ${result}`);
-        }
-      } else {
-        // Fallback for string content
-        const text = String(message.content).trim();
-        if (text) parts.push(`[Tool result]: ${text}`);
-      }
-      continue;
-    }
-
-    if (message.role === "system") {
-      const content =
-        typeof message.content === "string" ? message.content.trim() : getTextParts(message.content).join("\n").trim();
-      if (content) parts.push(`[System]: ${content}`);
-    }
-  }
-
-  return parts.join("\n\n");
-}
-
-export interface CompactionUsage {
-  promptTokens: number;
-  completionTokens: number;
-}
-
-export interface CompactionSummaryResult {
-  summary: string;
-  usage: CompactionUsage;
-}
-
-function readUsage(usage: unknown): CompactionUsage {
-  if (!isRecord(usage)) return { promptTokens: 0, completionTokens: 0 };
-  const prompt =
-    typeof usage.inputTokens === "number"
-      ? usage.inputTokens
-      : typeof (usage as Record<string, unknown>).promptTokens === "number"
-        ? (usage as { promptTokens: number }).promptTokens
-        : 0;
-  const completion =
-    typeof usage.outputTokens === "number"
-      ? usage.outputTokens
-      : typeof (usage as Record<string, unknown>).completionTokens === "number"
-        ? (usage as { completionTokens: number }).completionTokens
-        : 0;
-  return { promptTokens: prompt, completionTokens: completion };
-}
-
-async function summarizeConversation(
-  provider: LegacyProvider,
-  modelId: string,
-  messages: ModelMessage[],
-  reserveTokens: number,
-  customInstructions?: string,
-  previousSummary?: string,
-  promptOverride?: string,
-  signal?: AbortSignal,
-): Promise<CompactionSummaryResult> {
-  const serialized = serializeConversation(messages);
-  const promptParts = [serialized];
-
-  // Improvement from session df2dbb878984 analysis: detect meta-analysis (self-evaluation of agent/PIL)
-  // and force tighter maxOutputTokens to prevent 14k+ char summaries (see DB compaction at seq 52, 73k tokens -> 14k chars).
-  const userText = messages.map((m) => extractUserContent(m.content)).join("\n");
-  const isMeta = isMetaAnalysisPrompt(userText);
-  const effectiveMax = isMeta
-    ? Math.min(1024, Math.max(512, Math.floor(reserveTokens * 0.5)))
-    : Math.min(COMPACTION_MAX_OUTPUT_TOKENS, Math.max(512, Math.floor(reserveTokens * 0.8)));
-
-  if (previousSummary) {
-    promptParts.push(`Existing summary:\n${previousSummary}`);
-  }
-
-  const basePrompt = promptOverride ?? (previousSummary ? UPDATE_SUMMARIZATION_PROMPT : SUMMARIZATION_PROMPT);
-  promptParts.push(basePrompt);
-
-  if (customInstructions?.trim()) {
-    promptParts.push(`Additional focus: ${customInstructions.trim()}`);
-  }
-
-  const runtime = resolveModelRuntime(provider, modelId);
-  const compactCaps = getProviderCapabilities(requireRuntimeProvider(runtime));
-  const result = await generateText({
-    model: runtime.model,
-    system: SUMMARIZATION_SYSTEM_PROMPT,
-    prompt: promptParts.filter(Boolean).join("\n\n"),
-    abortSignal: signal,
-    maxRetries: 0,
-    temperature: 0.2,
-    ...(!compactCaps.acceptsParam("maxOutputTokens", runtime.modelInfo) ? {} : { maxOutputTokens: effectiveMax }),
-    ...(runtime.providerOptions ? { providerOptions: runtime.providerOptions } : {}),
-  });
-
-  return { summary: result.text.trim(), usage: readUsage(result.usage) };
-}
-
-function emptyUsage(): CompactionUsage {
-  return { promptTokens: 0, completionTokens: 0 };
-}
-
-function addUsage(a: CompactionUsage, b: CompactionUsage): CompactionUsage {
-  return {
-    promptTokens: a.promptTokens + b.promptTokens,
-    completionTokens: a.completionTokens + b.completionTokens,
-  };
-}
-
-export async function generateCompactionSummary(
-  provider: LegacyProvider,
-  modelId: string,
-  preparation: PreparedCompaction,
-  customInstructions?: string,
-  signal?: AbortSignal,
-): Promise<CompactionSummaryResult> {
-  const { messagesToSummarize, turnPrefixMessages, isSplitTurn, previousSummary, settings } = preparation;
-
-  if (isSplitTurn && turnPrefixMessages.length > 0) {
-    const [historyResult, prefixResult] = await Promise.all([
-      messagesToSummarize.length > 0
-        ? summarizeConversation(
-            provider,
-            modelId,
-            messagesToSummarize,
-            settings.reserveTokens,
-            customInstructions,
-            previousSummary,
-            undefined,
-            signal,
-          )
-        : Promise.resolve<CompactionSummaryResult>({ summary: previousSummary?.trim() || "", usage: emptyUsage() }),
-      summarizeConversation(
-        provider,
-        modelId,
-        turnPrefixMessages,
-        settings.reserveTokens,
-        undefined,
-        undefined,
-        TURN_PREFIX_SUMMARIZATION_PROMPT,
-        signal,
-      ),
-    ]);
-
-    const usage = addUsage(historyResult.usage, prefixResult.usage);
-    if (historyResult.summary && prefixResult.summary) {
-      return { summary: `${historyResult.summary}\n\n---\n\n${prefixResult.summary}`, usage };
-    }
-    return { summary: (historyResult.summary || prefixResult.summary).trim(), usage };
-  }
-
-  return summarizeConversation(
-    provider,
-    modelId,
-    messagesToSummarize,
-    settings.reserveTokens,
-    customInstructions,
-    previousSummary,
-    undefined,
-    signal,
-  );
-}