@haystackeditor/cli 0.7.2 → 0.8.1

package/dist/assets/hooks/agent-context/parsers/gemini.ts

@@ -0,0 +1,155 @@
+import * as fs from 'fs';
+import type {
+  AgentContext,
+  AgentParser,
+  DetectionResult,
+  InteractionTurn,
+  ModifiedFile,
+  TokenUsage,
+  ToolUseInfo,
+} from '../types.js';
+import { detectAgent } from '../detect.js';
+
+interface GeminiToolCall {
+  id?: string;
+  name?: string;
+  args?: Record<string, unknown>;
+  status?: string;
+}
+
+interface GeminiTokens {
+  input?: number;
+  output?: number;
+  cached?: number;
+  thought?: number;
+}
+
+interface GeminiMessage {
+  id?: string;
+  type: 'user' | 'gemini';
+  content: string | Array<{ text: string }>;
+  toolCalls?: GeminiToolCall[];
+  tokens?: GeminiTokens;
+}
+
+interface GeminiSession {
+  messages: GeminiMessage[];
+}
+
+const FILE_PATH_KEYS = ['file_path', 'path', 'filename'];
+
+function extractContent(content: string | Array<{ text: string }>): string {
+  if (typeof content === 'string') return content;
+  if (Array.isArray(content)) {
+    return content.map((c) => c.text).join('\n');
+  }
+  return '';
+}
+
+function extractModifiedFiles(toolCalls?: GeminiToolCall[]): ModifiedFile[] {
+  if (!toolCalls) return [];
+  const files: ModifiedFile[] = [];
+
+  for (const call of toolCalls) {
+    if (!call.args) continue;
+    for (const key of FILE_PATH_KEYS) {
+      const val = call.args[key];
+      if (typeof val === 'string') {
+        files.push({
+          filePath: val,
+          toolName: call.name || 'unknown',
+          toolUseId: call.id,
+        });
+        break;
+      }
+    }
+  }
+  return files;
+}
+
+function extractToolUses(toolCalls?: GeminiToolCall[]): ToolUseInfo[] {
+  if (!toolCalls) return [];
+  return toolCalls.map((call) => ({
+    toolName: call.name || 'unknown',
+    toolUseId: call.id || '',
+    input: (call.args as Record<string, unknown>) || {},
+  }));
+}
+
+export class GeminiParser implements AgentParser {
+  async detect(repoPath: string): Promise<DetectionResult> {
+    return detectAgent(repoPath);
+  }
+
+  async parse(
+    sessionFilePath: string,
+    _repoPath: string,
+  ): Promise<AgentContext> {
+    const raw = fs.readFileSync(sessionFilePath, 'utf-8');
+    const session: GeminiSession = JSON.parse(raw);
+    const messages = session.messages || [];
+
+    const transcript: InteractionTurn[] = [];
+    const totalUsage: TokenUsage = {
+      inputTokens: 0,
+      outputTokens: 0,
+    };
+
+    let taskPrompt = '';
+
+    for (let i = 0; i < messages.length; i++) {
+      const msg = messages[i];
+      const role = msg.type === 'user' ? 'user' : 'assistant';
+      const content = extractContent(msg.content);
+
+      if (role === 'user' && !taskPrompt && content) {
+        taskPrompt = content;
+      }
+
+      if (msg.tokens) {
+        totalUsage.inputTokens += msg.tokens.input || 0;
+        totalUsage.outputTokens += msg.tokens.output || 0;
+      }
+
+      transcript.push({
+        turnIndex: i,
+        role,
+        content,
+        modifiedFiles: role === 'assistant' ? extractModifiedFiles(msg.toolCalls) : [],
+        tokenUsage: msg.tokens
+          ? {
+              inputTokens: msg.tokens.input || 0,
+              outputTokens: msg.tokens.output || 0,
+            }
+          : undefined,
+        toolUses: role === 'assistant' ? extractToolUses(msg.toolCalls) : [],
+      });
+    }
+
+    const allModifiedFiles = new Set<string>();
+    for (const turn of transcript) {
+      for (const f of turn.modifiedFiles) {
+        allModifiedFiles.add(f.filePath);
+      }
+    }
+
+    const stat = fs.statSync(sessionFilePath);
+    const sessionId = sessionFilePath
+      .split('/')
+      .pop()
+      ?.replace(/\.json$/, '') || '';
+
+    return {
+      agent: 'gemini-cli',
+      sessionId,
+      sessionFilePath,
+      taskPrompt,
+      modifiedFiles: Array.from(allModifiedFiles),
+      tokenUsage: totalUsage,
+      transcript,
+      metadata: {
+        sessionFileModifiedAt: stat.mtime.toISOString(),
+      },
+    };
+  }
+}

package/dist/assets/hooks/agent-context/parsers/opencode.ts

@@ -0,0 +1,174 @@
+import * as fs from 'fs';
+import type {
+  AgentContext,
+  AgentParser,
+  DetectionResult,
+  InteractionTurn,
+  ModifiedFile,
+  TokenUsage,
+  ToolUseInfo,
+} from '../types.js';
+import { detectAgent } from '../detect.js';
+
+const FILE_MOD_TOOLS = new Set(['edit', 'write', 'patch']);
+
+interface OpenCodeTokens {
+  input?: number;
+  output?: number;
+  cache?: {
+    read?: number;
+    write?: number;
+  };
+}
+
+interface OpenCodePart {
+  type: 'text' | 'tool';
+  text?: string;
+  tool?: string;
+  state?: {
+    input?: Record<string, unknown>;
+  };
+}
+
+interface OpenCodeMessage {
+  info: {
+    id?: string;
+    role: 'user' | 'assistant';
+    createdAt?: string;
+    tokens?: OpenCodeTokens;
+  };
+  parts: OpenCodePart[];
+}
+
+interface OpenCodeSession {
+  info: {
+    id: string;
+    title?: string;
+    createdAt?: string;
+    updatedAt?: string;
+  };
+  messages: OpenCodeMessage[];
+}
+
+function extractTextContent(parts: OpenCodePart[]): string {
+  return parts
+    .filter((p) => p.type === 'text' && p.text)
+    .map((p) => p.text!)
+    .join('\n');
+}
+
+function extractModifiedFiles(parts: OpenCodePart[]): ModifiedFile[] {
+  const files: ModifiedFile[] = [];
+  for (const part of parts) {
+    if (part.type !== 'tool' || !part.tool) continue;
+    if (!FILE_MOD_TOOLS.has(part.tool)) continue;
+
+    const input = part.state?.input;
+    if (!input) continue;
+
+    const filePath = (input.filePath as string) || (input.path as string);
+    if (filePath) {
+      files.push({
+        filePath,
+        toolName: part.tool,
+      });
+    }
+  }
+  return files;
+}
+
+function extractToolUses(parts: OpenCodePart[]): ToolUseInfo[] {
+  const tools: ToolUseInfo[] = [];
+  for (const part of parts) {
+    if (part.type !== 'tool' || !part.tool) continue;
+    tools.push({
+      toolName: part.tool,
+      toolUseId: '',
+      input: (part.state?.input as Record<string, unknown>) || {},
+    });
+  }
+  return tools;
+}
+
+export class OpenCodeParser implements AgentParser {
+  async detect(repoPath: string): Promise<DetectionResult> {
+    return detectAgent(repoPath);
+  }
+
+  async parse(
+    sessionFilePath: string,
+    _repoPath: string,
+  ): Promise<AgentContext> {
+    const raw = fs.readFileSync(sessionFilePath, 'utf-8');
+    const session: OpenCodeSession = JSON.parse(raw);
+    const messages = session.messages || [];
+
+    const transcript: InteractionTurn[] = [];
+    const totalUsage: TokenUsage = {
+      inputTokens: 0,
+      outputTokens: 0,
+      cacheCreationTokens: 0,
+      cacheReadTokens: 0,
+    };
+
+    let taskPrompt = '';
+
+    for (let i = 0; i < messages.length; i++) {
+      const msg = messages[i];
+      const role = msg.info.role;
+      const content = extractTextContent(msg.parts);
+
+      if (role === 'user' && !taskPrompt && content) {
+        taskPrompt = content;
+      }
+
+      const tokens = msg.info.tokens;
+      if (tokens) {
+        totalUsage.inputTokens += tokens.input || 0;
+        totalUsage.outputTokens += tokens.output || 0;
+        totalUsage.cacheReadTokens! += tokens.cache?.read || 0;
+        totalUsage.cacheCreationTokens! += tokens.cache?.write || 0;
+      }
+
+      transcript.push({
+        turnIndex: i,
+        role,
+        content,
+        modifiedFiles:
+          role === 'assistant' ? extractModifiedFiles(msg.parts) : [],
+        tokenUsage: tokens
+          ? {
+              inputTokens: tokens.input || 0,
+              outputTokens: tokens.output || 0,
+              cacheReadTokens: tokens.cache?.read,
+              cacheCreationTokens: tokens.cache?.write,
+            }
+          : undefined,
+        toolUses: role === 'assistant' ? extractToolUses(msg.parts) : [],
+      });
+    }
+
+    const allModifiedFiles = new Set<string>();
+    for (const turn of transcript) {
+      for (const f of turn.modifiedFiles) {
+        allModifiedFiles.add(f.filePath);
+      }
+    }
+
+    const stat = fs.statSync(sessionFilePath);
+
+    return {
+      agent: 'opencode',
+      sessionId: session.info.id,
+      sessionFilePath,
+      taskPrompt,
+      modifiedFiles: Array.from(allModifiedFiles),
+      tokenUsage: totalUsage,
+      transcript,
+      metadata: {
+        startedAt: session.info.createdAt,
+        sessionFileModifiedAt: stat.mtime.toISOString(),
+      },
+    };
+  }
+}

package/dist/assets/hooks/agent-context/tsconfig.json

@@ -0,0 +1,13 @@
+{
+  "compilerOptions": {
+    "target": "ES2022",
+    "module": "Node16",
+    "moduleResolution": "Node16",
+    "strict": true,
+    "esModuleInterop": true,
+    "resolveJsonModule": true,
+    "noEmit": true,
+    "skipLibCheck": true
+  },
+  "include": ["./**/*.ts"]
+}

package/dist/assets/hooks/agent-context/types.ts

@@ -0,0 +1,58 @@
+export type AgentType = 'claude-code' | 'gemini-cli' | 'opencode';
+
+export interface ModifiedFile {
+  filePath: string;
+  toolName: string;
+  toolUseId?: string;
+}
+
+export interface TokenUsage {
+  inputTokens: number;
+  outputTokens: number;
+  cacheCreationTokens?: number;
+  cacheReadTokens?: number;
+}
+
+export interface ToolUseInfo {
+  toolName: string;
+  toolUseId: string;
+  input: Record<string, unknown>;
+}
+
+export interface InteractionTurn {
+  turnIndex: number;
+  role: 'user' | 'assistant';
+  content: string;
+  modifiedFiles: ModifiedFile[];
+  tokenUsage?: TokenUsage;
+  toolUses: ToolUseInfo[];
+}
+
+export interface AgentContext {
+  agent: AgentType;
+  sessionId: string;
+  sessionFilePath: string;
+  taskPrompt: string;
+  modifiedFiles: string[];
+  tokenUsage: TokenUsage;
+  transcript: InteractionTurn[];
+  metadata: {
+    model?: string;
+    version?: string;
+    gitBranch?: string;
+    startedAt?: string;
+    sessionFileModifiedAt: string;
+  };
+}
+
+export interface DetectionResult {
+  detected: boolean;
+  agent?: AgentType;
+  sessionFilePath?: string;
+  sessionId?: string;
+}
+
+export interface AgentParser {
+  detect(repoPath: string): Promise<DetectionResult>;
+  parse(sessionFilePath: string, repoPath: string): Promise<AgentContext>;
+}

package/dist/assets/hooks/llm-rules-template.md

@@ -0,0 +1,35 @@
+# LLM Rules
+
+These rules govern how LLM agents should interact with this codebase. They are enforced by a pre-commit hook.
+
+- When you write or edit LLM prompts in response to a failure running against some test data, be certain that your edit is not specific to the test data, and is actually addressing the more generic problem that caused the failure.
+- If the user specifies a model, follow it. Your understanding of current LLM models is outdated.
+- Do not add silent fallbacks.
+- Do not add backwards compatibility, unless you confirmed with the user first that it's desired.
+- Do not add TODOs and incomplete code unless they have been explicitly flagged to the user.
+
+## No Truncation Without Permission
+
+When writing or modifying agent code (tool calls, prompts, LLM pipelines):
+
+- **Never truncate, slice, or omit content** from tool call inputs, tool call outputs, prompt context, search results, file contents, or any data passed to/from LLMs without explicit user permission.
+- **Never summarize in place of full content** (e.g., "... [content omitted]", "showing first 50 lines", "[truncated for brevity]").
+- **No silent length limits** - do not add code that silently cuts off text at N characters/lines/tokens.
+
+If truncation is technically necessary (e.g., context limits):
+1. **Disclose explicitly** that truncation will occur
+2. **State what and how much** will be omitted
+3. **Ask for user permission** before proceeding
+4. **Offer alternatives** (e.g., chunking, pagination with user control)
+
+This applies to: grep/search results, file reads, API responses, prompt templates, tool definitions, agent memory, conversation history, and any other text flowing through agent pipelines.
+
+## No Hardcoded Repo Knowledge in Prompts
+
+When writing LLM prompts for generic tools:
+
+- **Never hardcode repo-specific examples** like component names, file paths, or internal terminology.
+- **Use generic examples** that would work for any codebase (e.g., "Sidebar", "Card", "Modal").
+- **Let the LLM infer from context** rather than baking in assumptions about your specific codebase.
+
+This keeps tools portable and prevents prompts from becoming stale when the codebase changes.

package/dist/assets/hooks/package.json

@@ -0,0 +1,11 @@
+{
+  "name": "haystack-hooks",
+  "version": "1.0.0",
+  "private": true,
+  "type": "module",
+  "description": "Git hooks for AI agent quality checks (installed by @haystackeditor/cli)",
+  "dependencies": {
+    "tree-sitter": "^0.22.4",
+    "tree-sitter-typescript": "^0.23.2"
+  }
+}

package/dist/assets/hooks/scripts/commit-msg.sh

@@ -0,0 +1,4 @@
+#!/bin/sh
+# Haystack hooks (powered by Entire)
+# Commit-msg hook: strip trailer if no user content (allows aborting empty commits)
+"$HOME/.haystack/bin/entire" hooks git commit-msg "$1" || exit 1

package/dist/assets/hooks/scripts/post-commit.sh

@@ -0,0 +1,4 @@
+#!/bin/sh
+# Haystack hooks (powered by Entire)
+# Post-commit hook: condense session data if commit has Entire-Checkpoint trailer
+"$HOME/.haystack/bin/entire" hooks git post-commit 2>/dev/null || true

package/dist/assets/hooks/scripts/pre-commit.sh

@@ -0,0 +1,92 @@
+#!/usr/bin/env bash
+HOOKS_DIR="$(cd "$(dirname "$0")" && pwd)"
+REPO_ROOT="$(cd "$HOOKS_DIR/.." && pwd)"
+
+# Extract agent context if running under an AI agent
+# Capture stderr to detect whether an AI agent is present
+AGENT_OUTPUT=$(npx tsx "$HOOKS_DIR/agent-context/index.ts" "$REPO_ROOT" 2>&1) || true
+
+# Print the agent context to stderr so it still appears in the commit output
+if [ -n "$AGENT_OUTPUT" ]; then
+  echo "$AGENT_OUTPUT" >&2
+fi
+
+# If no agent detected, this is a human commit — pass through
+if ! echo "$AGENT_OUTPUT" | grep -q "AGENT CONTEXT DETECTED"; then
+  exit 0
+fi
+
+# --- AI agent detected: run automated checks ---
+
+# Check for truncation violations in staged changes
+TRUNCATION_OUTPUT=$(npx tsx "$HOOKS_DIR/truncation-checker/index.ts" 2>&1)
+TRUNCATION_EXIT=$?
+
+if [ -n "$TRUNCATION_OUTPUT" ]; then
+  echo "$TRUNCATION_OUTPUT" >&2
+fi
+
+if [ $TRUNCATION_EXIT -ne 0 ]; then
+  echo "" >&2
+  echo "========================================" >&2
+  echo "  COMMIT BLOCKED: TRUNCATION VIOLATIONS" >&2
+  echo "========================================" >&2
+  echo "" >&2
+  echo "Your staged changes contain truncation patterns that violate LLM_RULES.md." >&2
+  echo "Review the violations above and either:" >&2
+  echo "  1. Remove the truncation code" >&2
+  echo "  2. Add a comment with 'intentional' explaining why it's needed" >&2
+  echo "  3. Get explicit user permission for the truncation" >&2
+  echo "" >&2
+  exit 1
+fi
+
+# --- Enforce LLM rules review ---
+
+GIT_DIR=$(git rev-parse --git-dir)
+BYPASS_TOKEN_FILE="$GIT_DIR/llm-rules-reviewed"
+RULES_FILE="$REPO_ROOT/LLM_RULES.md"
+
+# Get the SHA of the current staged tree
+STAGED_TREE_SHA=$(git write-tree)
+
+# Check for a valid bypass token (written on a previous failed attempt)
+if [ -f "$BYPASS_TOKEN_FILE" ]; then
+  STORED_SHA=$(cat "$BYPASS_TOKEN_FILE")
+  if [ "$STORED_SHA" = "$STAGED_TREE_SHA" ]; then
+    # Agent has already been shown the rules for this exact set of changes
+    rm -f "$BYPASS_TOKEN_FILE"
+    exit 0
+  fi
+fi
+
+# First attempt — block the commit and inject the rules into the agent's context
+
+echo "" >&2
+echo "========================================" >&2
+echo "  LLM RULES — REVIEW BEFORE COMMITTING" >&2
+echo "========================================" >&2
+echo "" >&2
+
+if [ -f "$RULES_FILE" ]; then
+  cat "$RULES_FILE" >&2
+else
+  echo "(LLM_RULES.md not found at $RULES_FILE)" >&2
+fi
+
+echo "" >&2
+echo "========================================" >&2
+echo "  STAGED CHANGES" >&2
+echo "========================================" >&2
+echo "" >&2
+git diff --cached --stat >&2
+
+echo "" >&2
+echo "========================================" >&2
+echo "PRE-COMMIT CHECK: You must confirm your changes comply with the above rules. The rules have been re-injected into your context. Review your staged changes against each rule, then retry the commit. A bypass token has been written." >&2
+echo "========================================" >&2
+
+# Write the bypass token so the next attempt with the same staged tree passes
+echo "$STAGED_TREE_SHA" > "$BYPASS_TOKEN_FILE"
+
+exit 1

package/dist/assets/hooks/scripts/pre-push.sh

@@ -0,0 +1,5 @@
+#!/bin/sh
+# Haystack hooks (powered by Entire)
+# Pre-push hook: push session logs alongside user's push
+# $1 is the remote name (e.g., "origin")
+"$HOME/.haystack/bin/entire" hooks git pre-push "$1" || true

package/dist/assets/hooks/scripts/prepare-commit-msg.sh

@@ -0,0 +1,3 @@
+#!/bin/sh
+# Haystack hooks (powered by Entire)
+"$HOME/.haystack/bin/entire" hooks git prepare-commit-msg "$1" "$2" 2>/dev/null || true