Naive Coding Agent (Context Stuffing & Broken Patches):
1. Ingests all 500 files (1.2M tokens!) ──► High Cost ($2.50 / query) + Attention Confusion
2. Outputs raw diff with wrong line numbers ──► `git apply` FAILS! 💥

State-of-the-Art Coding Agent Architecture (Aider + SWE-agent):
1. [ Tree-Sitter AST Repo Map ]: Extracts only signatures, classes & call-graphs (compact 2k tokens!)
2. [ Search & Navigation Tools ]: Explores code iteratively with `grep_search` and `view_file`.
3. [ Precise Surgical Editing ]: Uses Exact-Match Unified Diffs or ACI file replacement commands! ✅

                              [ Entire 500,000-Line Codebase ]
                                             │
                                             ▼ (Tree-Sitter AST Parser)
                        [ Extract Class, Function & Struct Signatures ]
                                             │
                                             ▼ (Build Call-Graph Dependency Matrix)
                        [ Apply Graph PageRank / Personalized PageRank ]
                                             │
                                             ▼
                 [ Compact 2,048-Token Semantic Repo Map Injected into LLM Prompt! ]

# Sample Aider Repo Map Output (Injected into Context):
# src/compiler/parser.py:
# │ class ASTParser:
# │     def parse_tokens(tokens: List[Token]) -> ASTNode
# │     def evaluate_expressions(node: ASTNode) -> Value
#
# src/runtime/engine.py:
# │ class VMRuntime:
# │     def execute_bytecode(code: Bytecode) -> ExecutionResult

┌──────────────────┬───────────────────────────────────────────────────────┐
│ ACI Custom Tool  │ Purpose & LLM Guardrail Design                        │
├──────────────────┼───────────────────────────────────────────────────────┤
│ `open_file`      │ Opens file and displays first 100 lines with explicit │
│                  │ line numbers. (Prevents dumping 10k lines into context)│
├─────────────────┼───────────────────────────────────────────────────────┤
│ `scroll_up/down` │ Paginates smoothly through long files.                │
├─────────────────┼───────────────────────────────────────────────────────┤
│ `edit_file`      │ Replaces a contiguous block of text using EXACT-MATCH │
│                  │ search strings rather than hallucinated line numbers. │
├─────────────────┼───────────────────────────────────────────────────────┤
│ `find_symbol`    │ Locates class/function definitions across files.      │
└─────────────────┴───────────────────────────────────────────────────────┘

# coding_agent_loop.py - Production Autonomous Coding Loop
import subprocess
from typing import Dict, Any

class CodingAgentHarness:
    def __init__(self, workspace_root: str):
        self.workspace = workspace_root

    def apply_exact_replacement(self, file_path: str, target_block: str, replacement_block: str) -> Dict[str, Any]:
        with open(file_path, "r", encoding="utf-8") as f:
            content = f.read()

        # Exact match verification
        if target_block not in content:
            return {"success": False, "error": "Target block not found in file. Ensure exact whitespace matching."}
        
        if content.count(target_block) > 1:
            return {"success": False, "error": "Multiple occurrences found. Please include more surrounding context lines."}

        new_content = content.replace(target_block, replacement_block, 1)
        with open(file_path, "w", encoding="utf-8") as f:
            f.write(new_content)

        return {"success": True, "message": "Replacement applied successfully."}

    def run_validation_test(self, test_command: str) -> Dict[str, Any]:
        res = subprocess.run(test_command, shell=True, capture_output=True, text=True, cwd=self.workspace)
        return {
            "passed": res.returncode == 0,
            "stdout": res.stdout[:2000], # Cap output length
            "stderr": res.stderr[:2000]
        }

SWE-bench Verified Resolution Rate (%):
┌─────────────────────────────────────────────────────────┐
│ Zero-Shot Raw Files:  ██ 8.2%                           │
│ ReAct Standard Bash:  ███████ 28.4%                     │
│ Aider Repo Maps:      ███████████ 44.8%                 │
│ SWE-agent ACI:        ████████████ 49.2%                │
│ o1 Tree Search Agent: ████████████████ 64.8%!           │
└─────────────────────────────────────────────────────────┘