AICode/ai_software_factory
3
0
Fork 0
Code Issues Pull Requests Actions Packages Projects Releases 30 Wiki Activity

fix: more file generation improvements, refs NOISSUE

Browse Source
This commit is contained in:
simon
2026-04-11 14:33:45 +02:00
parent 7e4066c609
commit 1944e2a9cf
1 changed files with 123 additions and 24 deletions
Download Patch File Download Diff File Expand all files Collapse all files

147
ai_software_factory/agents/orchestrator.py
View File

@@ -7,6 +7,7 @@ import json
import py_compile
import re
import subprocess
from pathlib import PurePosixPath
from typing import Optional
from datetime import datetime
@@ -31,6 +32,10 @@ class AgentOrchestrator:
REMOTE_READY_REPOSITORY_MODES = {'project', 'onboarded'}
REMOTE_READY_REPOSITORY_STATUSES = {'created', 'exists', 'ready', 'onboarded'}
GENERATED_TEXT_FILE_SUFFIXES = {'.py', '.md', '.txt', '.toml', '.yaml', '.yml', '.json', '.ini', '.cfg', '.sh', '.html', '.css', '.js', '.ts'}
GENERATED_TEXT_FILE_NAMES = {'README', 'README.md', '.gitignore', 'requirements.txt', 'pyproject.toml', 'Dockerfile', 'Containerfile', 'Makefile'}
MAX_WORKSPACE_CONTEXT_FILES = 20
MAX_WORKSPACE_CONTEXT_CHARS = 24000
def __init__(
self,
@@ -240,6 +245,59 @@ class AgentOrchestrator:
fallback_used=False,
)
def _is_safe_relative_path(self, path: str) -> bool:
"""Return whether a generated file path is safe to write under the project root."""
normalized = str(PurePosixPath((path or '').strip()))
if not normalized or normalized in {'.', '..'}:
return False
if normalized.startswith('/') or normalized.startswith('../') or '/../' in normalized:
return False
if normalized.startswith('.git/'):
return False
return True
def _is_supported_generated_text_file(self, path: str) -> bool:
"""Return whether the generated path is a supported text artifact."""
normalized = PurePosixPath(path)
if normalized.name in self.GENERATED_TEXT_FILE_NAMES:
return True
return normalized.suffix.lower() in self.GENERATED_TEXT_FILE_SUFFIXES
def _collect_workspace_context(self) -> dict:
"""Collect a compact, text-only snapshot of the current project workspace."""
if not self.project_root.exists():
return {'has_existing_files': False, 'files': []}
files: list[dict] = []
total_chars = 0
for path in sorted(self.project_root.rglob('*')):
if not path.is_file():
continue
relative_path = path.relative_to(self.project_root).as_posix()
if relative_path == '.gitignore':
continue
if not self._is_safe_relative_path(relative_path) or not self._is_supported_generated_text_file(relative_path):
continue
try:
content = path.read_text(encoding='utf-8')
except (UnicodeDecodeError, OSError):
continue
remaining_chars = self.MAX_WORKSPACE_CONTEXT_CHARS - total_chars
if remaining_chars <= 0:
break
snippet = content[:remaining_chars]
files.append(
{
'path': relative_path,
'content': snippet,
'truncated': len(snippet) < len(content),
}
)
total_chars += len(snippet)
if len(files) >= self.MAX_WORKSPACE_CONTEXT_FILES:
break
return {'has_existing_files': bool(files), 'files': files}
def _parse_generated_files(self, content: str | None) -> dict[str, str]:
"""Parse an LLM file bundle response into relative-path/content pairs."""
if not content:
@@ -248,7 +306,6 @@ class AgentOrchestrator:
parsed = json.loads(content)
except Exception:
return {}
allowed_paths = set(self._fallback_generated_files().keys())
generated: dict[str, str] = {}
if isinstance(parsed, dict) and isinstance(parsed.get('files'), list):
for item in parsed['files']:
@@ -256,34 +313,68 @@ class AgentOrchestrator:
continue
path = str(item.get('path') or '').strip()
file_content = item.get('content')
if path in allowed_paths and isinstance(file_content, str) and file_content.strip():
if (
self._is_safe_relative_path(path)
and self._is_supported_generated_text_file(path)
and isinstance(file_content, str)
and file_content.strip()
):
generated[path] = file_content.rstrip() + "\n"
elif isinstance(parsed, dict):
for path, file_content in parsed.items():
if path in allowed_paths and isinstance(file_content, str) and file_content.strip():
generated[str(path)] = file_content.rstrip() + "\n"
normalized_path = str(path).strip()
if (
self._is_safe_relative_path(normalized_path)
and self._is_supported_generated_text_file(normalized_path)
and isinstance(file_content, str)
and file_content.strip()
):
generated[normalized_path] = file_content.rstrip() + "\n"
return generated
async def _generate_prompt_driven_files(self) -> tuple[dict[str, str], dict | None]:
async def _generate_prompt_driven_files(self) -> tuple[dict[str, str], dict | None, bool]:
"""Use the configured LLM to generate prompt-specific project files."""
fallback_files = self._fallback_generated_files()
system_prompt = (
'You generate small but concrete starter projects. '
'Return only JSON. Provide production-like but compact code that directly reflects the user request. '
'Include the files README.md, requirements.txt, main.py, and tests/test_app.py. '
'Use FastAPI for Python web requests unless the prompt clearly demands something else. '
'The test must verify a real behavior from main.py. '
'Do not wrap the JSON in markdown fences.'
)
user_prompt = (
f"Project name: {self.project_name}\n"
f"Description: {self.description}\n"
f"Original prompt: {self.prompt_text or self.description}\n"
f"Requested features: {json.dumps(self.features)}\n"
f"Preferred tech stack: {json.dumps(self.tech_stack)}\n"
f"Related issue: {json.dumps(self.related_issue) if self.related_issue else 'null'}\n\n"
"Return JSON shaped as {\"files\": [{\"path\": \"README.md\", \"content\": \"...\"}, ...]}."
)
workspace_context = self._collect_workspace_context()
has_existing_files = bool(workspace_context.get('has_existing_files'))
if has_existing_files:
system_prompt = (
'You modify an existing software repository. '
'Return only JSON. Update the smallest necessary set of files to satisfy the new prompt. '
'Prefer editing existing files over inventing a new starter app. '
'Only return files that should be written. Omit unchanged files. '
'Use repository-relative paths and do not wrap the JSON in markdown fences.'
)
user_prompt = (
f"Project name: {self.project_name}\n"
f"Description: {self.description}\n"
f"Original prompt: {self.prompt_text or self.description}\n"
f"Requested features: {json.dumps(self.features)}\n"
f"Preferred tech stack: {json.dumps(self.tech_stack)}\n"
f"Related issue: {json.dumps(self.related_issue) if self.related_issue else 'null'}\n\n"
f"Current workspace snapshot:\n{json.dumps(workspace_context['files'], indent=2)}\n\n"
'Return JSON shaped as {"files": [{"path": "relative/path.py", "content": "..."}, ...]}. '
'Each file path must be relative to the repository root.'
)
else:
system_prompt = (
'You generate small but concrete starter projects. '
'Return only JSON. Provide production-like but compact code that directly reflects the user request. '
'Include the files README.md, requirements.txt, main.py, and tests/test_app.py. '
'Use FastAPI for Python web requests unless the prompt clearly demands something else. '
'The test must verify a real behavior from main.py. '
'Do not wrap the JSON in markdown fences.'
)
user_prompt = (
f"Project name: {self.project_name}\n"
f"Description: {self.description}\n"
f"Original prompt: {self.prompt_text or self.description}\n"
f"Requested features: {json.dumps(self.features)}\n"
f"Preferred tech stack: {json.dumps(self.tech_stack)}\n"
f"Related issue: {json.dumps(self.related_issue) if self.related_issue else 'null'}\n\n"
'Return JSON shaped as {"files": [{"path": "README.md", "content": "..."}, ...]}. '
'At minimum include README.md, requirements.txt, main.py, and tests/test_app.py.'
)
content, trace = await LLMServiceClient().chat_with_trace(
stage='generation_plan',
system_prompt=system_prompt,
@@ -293,12 +384,15 @@ class AgentOrchestrator:
'project_name': self.project_name,
'repository': self.ui_manager.ui_data.get('repository'),
'related_issue': self.related_issue,
'workspace_files': workspace_context.get('files', []),
},
expect_json=True,
)
generated_files = self._parse_generated_files(content)
if has_existing_files:
return generated_files, trace, True
merged_files = {**fallback_files, **generated_files}
return merged_files, trace
return merged_files, trace, False
async def _sync_issue_context(self) -> None:
"""Sync repository issues and resolve a linked issue from the prompt when present."""
@@ -571,6 +665,8 @@ class AgentOrchestrator:
target.parent.mkdir(parents=True, exist_ok=True)
change_type = "UPDATE" if target.exists() else "CREATE"
previous_content = target.read_text(encoding="utf-8") if target.exists() else ""
if previous_content == content:
return
diff_text = self._build_diff_text(relative_path, previous_content, content)
target.write_text(content, encoding="utf-8")
self.changed_files.append(relative_path)
@@ -679,9 +775,12 @@ class AgentOrchestrator:
async def _generate_code(self) -> None:
"""Generate code using Ollama."""
generated_files, trace = await self._generate_prompt_driven_files()
change_count_before = len(self.pending_code_changes)
generated_files, trace, editing_existing_workspace = await self._generate_prompt_driven_files()
for relative_path, content in generated_files.items():
self._write_file(relative_path, content)
if editing_existing_workspace and len(self.pending_code_changes) == change_count_before:
raise RuntimeError('The LLM response did not produce any file changes for the existing project.')
fallback_used = bool(trace and trace.get('fallback_used')) or trace is None
if self.db_manager and self.history and self.prompt_audit and trace:
self.db_manager.log_llm_trace(