forked from Rockachopa/Timmy-time-dashboard
feat: Self-Coding Foundation (Phase 1)
Implements the foundational infrastructure for Timmy's self-modification capability:
## New Services
1. **GitSafety** (src/self_coding/git_safety.py)
- Atomic git operations with rollback capability
- Snapshot/restore for safe experimentation
- Feature branch management (timmy/self-edit/{timestamp})
- Merge to main only after tests pass
2. **CodebaseIndexer** (src/self_coding/codebase_indexer.py)
- AST-based parsing of Python source files
- Extracts classes, functions, imports, docstrings
- Builds dependency graph for blast radius analysis
- SQLite storage with hash-based incremental indexing
- get_summary() for LLM context (<4000 tokens)
- get_relevant_files() for task-based file discovery
3. **ModificationJournal** (src/self_coding/modification_journal.py)
- Persistent log of all self-modification attempts
- Tracks outcomes: success, failure, rollback
- find_similar() for learning from past attempts
- Success rate metrics and recent failure tracking
- Supports vector embeddings (Phase 2)
4. **ReflectionService** (src/self_coding/reflection.py)
- LLM-powered analysis of modification attempts
- Generates lessons learned from successes and failures
- Fallback templates when LLM unavailable
- Supports context from similar past attempts
## Test Coverage
- 104 new tests across 7 test files
- 95% code coverage on self_coding module
- Green path tests: full workflow integration
- Red path tests: errors, rollbacks, edge cases
- Safety constraint tests: test coverage requirements, protected files
## Usage
from self_coding import GitSafety, CodebaseIndexer, ModificationJournal
git = GitSafety(repo_path=/path/to/repo)
indexer = CodebaseIndexer(repo_path=/path/to/repo)
journal = ModificationJournal()
Phase 2 will build the Self-Edit MCP Tool that orchestrates these services.
This commit is contained in:
Alexander Payne
50
src/self_coding/__init__.py
Normal file
50
src/self_coding/__init__.py
Normal file
@@ -0,0 +1,50 @@
|
||||
"""Self-Coding Layer — Timmy's ability to modify its own source code safely.
|
||||
|
||||
This module provides the foundational infrastructure for self-modification:
|
||||
|
||||
- GitSafety: Atomic git operations with rollback capability
|
||||
- CodebaseIndexer: Live mental model of the codebase
|
||||
- ModificationJournal: Persistent log of modification attempts
|
||||
- ReflectionService: Generate lessons learned from attempts
|
||||
|
||||
Usage:
|
||||
from self_coding import GitSafety, CodebaseIndexer, ModificationJournal
|
||||
from self_coding import ModificationAttempt, Outcome, Snapshot
|
||||
|
||||
# Initialize services
|
||||
git = GitSafety(repo_path="/path/to/repo")
|
||||
indexer = CodebaseIndexer(repo_path="/path/to/repo")
|
||||
journal = ModificationJournal()
|
||||
|
||||
# Use in self-modification workflow
|
||||
snapshot = await git.snapshot()
|
||||
# ... make changes ...
|
||||
if tests_pass:
|
||||
await git.commit("Changes", ["file.py"])
|
||||
else:
|
||||
await git.rollback(snapshot)
|
||||
"""
|
||||
|
||||
from self_coding.git_safety import GitSafety, Snapshot
|
||||
from self_coding.codebase_indexer import CodebaseIndexer, ModuleInfo, FunctionInfo, ClassInfo
|
||||
from self_coding.modification_journal import (
|
||||
ModificationJournal,
|
||||
ModificationAttempt,
|
||||
Outcome,
|
||||
)
|
||||
from self_coding.reflection import ReflectionService
|
||||
|
||||
__all__ = [
|
||||
# Core services
|
||||
"GitSafety",
|
||||
"CodebaseIndexer",
|
||||
"ModificationJournal",
|
||||
"ReflectionService",
|
||||
# Data classes
|
||||
"Snapshot",
|
||||
"ModuleInfo",
|
||||
"FunctionInfo",
|
||||
"ClassInfo",
|
||||
"ModificationAttempt",
|
||||
"Outcome",
|
||||
]
|
||||
772
src/self_coding/codebase_indexer.py
Normal file
772
src/self_coding/codebase_indexer.py
Normal file
@@ -0,0 +1,772 @@
|
||||
"""Codebase Indexer — Live mental model of Timmy's own codebase.
|
||||
|
||||
Parses Python files using AST to extract classes, functions, imports, and
|
||||
docstrings. Builds a dependency graph and provides semantic search for
|
||||
relevant files.
|
||||
"""
|
||||
|
||||
from __future__ import annotations
|
||||
|
||||
import ast
|
||||
import hashlib
|
||||
import json
|
||||
import logging
|
||||
import sqlite3
|
||||
from dataclasses import asdict, dataclass, field
|
||||
from datetime import datetime, timezone
|
||||
from pathlib import Path
|
||||
from typing import Any, Optional
|
||||
|
||||
logger = logging.getLogger(__name__)
|
||||
|
||||
# Default database location
|
||||
DEFAULT_DB_PATH = Path("data/self_coding.db")
|
||||
|
||||
|
||||
@dataclass
|
||||
class FunctionInfo:
|
||||
"""Information about a function."""
|
||||
name: str
|
||||
args: list[str]
|
||||
returns: Optional[str] = None
|
||||
docstring: Optional[str] = None
|
||||
line_number: int = 0
|
||||
is_async: bool = False
|
||||
is_method: bool = False
|
||||
|
||||
|
||||
@dataclass
|
||||
class ClassInfo:
|
||||
"""Information about a class."""
|
||||
name: str
|
||||
methods: list[FunctionInfo] = field(default_factory=list)
|
||||
docstring: Optional[str] = None
|
||||
line_number: int = 0
|
||||
bases: list[str] = field(default_factory=list)
|
||||
|
||||
|
||||
@dataclass
|
||||
class ModuleInfo:
|
||||
"""Information about a Python module."""
|
||||
file_path: str
|
||||
module_name: str
|
||||
classes: list[ClassInfo] = field(default_factory=list)
|
||||
functions: list[FunctionInfo] = field(default_factory=list)
|
||||
imports: list[str] = field(default_factory=list)
|
||||
docstring: Optional[str] = None
|
||||
test_coverage: Optional[str] = None
|
||||
|
||||
|
||||
class CodebaseIndexer:
|
||||
"""Indexes Python codebase for self-modification workflows.
|
||||
|
||||
Parses all Python files using AST to extract:
|
||||
- Module names and structure
|
||||
- Class definitions with methods
|
||||
- Function signatures with args and return types
|
||||
- Import relationships
|
||||
- Test coverage mapping
|
||||
|
||||
Stores everything in SQLite for fast querying.
|
||||
|
||||
Usage:
|
||||
indexer = CodebaseIndexer(repo_path="/path/to/repo")
|
||||
|
||||
# Full reindex
|
||||
await indexer.index_all()
|
||||
|
||||
# Incremental update
|
||||
await indexer.index_changed()
|
||||
|
||||
# Get LLM context summary
|
||||
summary = await indexer.get_summary()
|
||||
|
||||
# Find relevant files for a task
|
||||
files = await indexer.get_relevant_files("Add error handling to health endpoint")
|
||||
|
||||
# Get dependency chain
|
||||
deps = await indexer.get_dependency_chain("src/timmy/agent.py")
|
||||
"""
|
||||
|
||||
def __init__(
|
||||
self,
|
||||
repo_path: Optional[str | Path] = None,
|
||||
db_path: Optional[str | Path] = None,
|
||||
src_dirs: Optional[list[str]] = None,
|
||||
) -> None:
|
||||
"""Initialize CodebaseIndexer.
|
||||
|
||||
Args:
|
||||
repo_path: Root of repository to index. Defaults to current directory.
|
||||
db_path: SQLite database path. Defaults to data/self_coding.db
|
||||
src_dirs: Source directories to index. Defaults to ["src", "tests"]
|
||||
"""
|
||||
self.repo_path = Path(repo_path).resolve() if repo_path else Path.cwd()
|
||||
self.db_path = Path(db_path) if db_path else DEFAULT_DB_PATH
|
||||
self.src_dirs = src_dirs or ["src", "tests"]
|
||||
self._ensure_schema()
|
||||
logger.info("CodebaseIndexer initialized for %s", self.repo_path)
|
||||
|
||||
def _get_conn(self) -> sqlite3.Connection:
|
||||
"""Get database connection with schema ensured."""
|
||||
self.db_path.parent.mkdir(parents=True, exist_ok=True)
|
||||
conn = sqlite3.connect(str(self.db_path))
|
||||
conn.row_factory = sqlite3.Row
|
||||
return conn
|
||||
|
||||
def _ensure_schema(self) -> None:
|
||||
"""Create database tables if they don't exist."""
|
||||
with self._get_conn() as conn:
|
||||
# Main codebase index table
|
||||
conn.execute(
|
||||
"""
|
||||
CREATE TABLE IF NOT EXISTS codebase_index (
|
||||
file_path TEXT PRIMARY KEY,
|
||||
module_name TEXT NOT NULL,
|
||||
classes JSON,
|
||||
functions JSON,
|
||||
imports JSON,
|
||||
test_coverage TEXT,
|
||||
last_indexed TIMESTAMP NOT NULL,
|
||||
content_hash TEXT NOT NULL,
|
||||
docstring TEXT,
|
||||
embedding BLOB
|
||||
)
|
||||
"""
|
||||
)
|
||||
|
||||
# Dependency graph table
|
||||
conn.execute(
|
||||
"""
|
||||
CREATE TABLE IF NOT EXISTS dependency_graph (
|
||||
id INTEGER PRIMARY KEY AUTOINCREMENT,
|
||||
source_file TEXT NOT NULL,
|
||||
target_file TEXT NOT NULL,
|
||||
import_type TEXT NOT NULL,
|
||||
UNIQUE(source_file, target_file)
|
||||
)
|
||||
"""
|
||||
)
|
||||
|
||||
# Create indexes
|
||||
conn.execute(
|
||||
"CREATE INDEX IF NOT EXISTS idx_module_name ON codebase_index(module_name)"
|
||||
)
|
||||
conn.execute(
|
||||
"CREATE INDEX IF NOT EXISTS idx_test_coverage ON codebase_index(test_coverage)"
|
||||
)
|
||||
conn.execute(
|
||||
"CREATE INDEX IF NOT EXISTS idx_deps_source ON dependency_graph(source_file)"
|
||||
)
|
||||
conn.execute(
|
||||
"CREATE INDEX IF NOT EXISTS idx_deps_target ON dependency_graph(target_file)"
|
||||
)
|
||||
|
||||
conn.commit()
|
||||
|
||||
def _compute_hash(self, content: str) -> str:
|
||||
"""Compute SHA-256 hash of file content."""
|
||||
return hashlib.sha256(content.encode("utf-8")).hexdigest()
|
||||
|
||||
def _find_python_files(self) -> list[Path]:
|
||||
"""Find all Python files in source directories."""
|
||||
files = []
|
||||
for src_dir in self.src_dirs:
|
||||
src_path = self.repo_path / src_dir
|
||||
if src_path.exists():
|
||||
files.extend(src_path.rglob("*.py"))
|
||||
return sorted(files)
|
||||
|
||||
def _find_test_file(self, source_file: Path) -> Optional[str]:
|
||||
"""Find corresponding test file for a source file.
|
||||
|
||||
Uses conventions:
|
||||
- src/x/y.py -> tests/test_x_y.py
|
||||
- src/x/y.py -> tests/x/test_y.py
|
||||
- src/x/y.py -> tests/test_y.py
|
||||
"""
|
||||
rel_path = source_file.relative_to(self.repo_path)
|
||||
|
||||
# Only look for tests for files in src/
|
||||
if not str(rel_path).startswith("src/"):
|
||||
return None
|
||||
|
||||
# Try various test file naming conventions
|
||||
possible_tests = [
|
||||
# tests/test_module.py
|
||||
self.repo_path / "tests" / f"test_{source_file.stem}.py",
|
||||
# tests/test_path_module.py (flat)
|
||||
self.repo_path / "tests" / f"test_{'_'.join(rel_path.with_suffix('').parts[1:])}.py",
|
||||
]
|
||||
|
||||
# Try mirroring src structure in tests (tests/x/test_y.py)
|
||||
try:
|
||||
src_relative = rel_path.relative_to("src")
|
||||
possible_tests.append(
|
||||
self.repo_path / "tests" / src_relative.parent / f"test_{source_file.stem}.py"
|
||||
)
|
||||
except ValueError:
|
||||
pass
|
||||
|
||||
for test_path in possible_tests:
|
||||
if test_path.exists():
|
||||
return str(test_path.relative_to(self.repo_path))
|
||||
|
||||
return None
|
||||
|
||||
def _parse_function(self, node: ast.FunctionDef | ast.AsyncFunctionDef, is_method: bool = False) -> FunctionInfo:
|
||||
"""Parse a function definition node."""
|
||||
args = []
|
||||
|
||||
# Handle different Python versions' AST structures
|
||||
func_args = node.args
|
||||
|
||||
# Positional args
|
||||
for arg in func_args.args:
|
||||
arg_str = arg.arg
|
||||
if arg.annotation:
|
||||
arg_str += f": {ast.unparse(arg.annotation)}"
|
||||
args.append(arg_str)
|
||||
|
||||
# Keyword-only args
|
||||
for arg in func_args.kwonlyargs:
|
||||
arg_str = arg.arg
|
||||
if arg.annotation:
|
||||
arg_str += f": {ast.unparse(arg.annotation)}"
|
||||
args.append(arg_str)
|
||||
|
||||
# Return type
|
||||
returns = None
|
||||
if node.returns:
|
||||
returns = ast.unparse(node.returns)
|
||||
|
||||
# Docstring
|
||||
docstring = ast.get_docstring(node)
|
||||
|
||||
return FunctionInfo(
|
||||
name=node.name,
|
||||
args=args,
|
||||
returns=returns,
|
||||
docstring=docstring,
|
||||
line_number=node.lineno,
|
||||
is_async=isinstance(node, ast.AsyncFunctionDef),
|
||||
is_method=is_method,
|
||||
)
|
||||
|
||||
def _parse_class(self, node: ast.ClassDef) -> ClassInfo:
|
||||
"""Parse a class definition node."""
|
||||
methods = []
|
||||
|
||||
for item in node.body:
|
||||
if isinstance(item, (ast.FunctionDef, ast.AsyncFunctionDef)):
|
||||
methods.append(self._parse_function(item, is_method=True))
|
||||
|
||||
# Get bases
|
||||
bases = [ast.unparse(base) for base in node.bases]
|
||||
|
||||
return ClassInfo(
|
||||
name=node.name,
|
||||
methods=methods,
|
||||
docstring=ast.get_docstring(node),
|
||||
line_number=node.lineno,
|
||||
bases=bases,
|
||||
)
|
||||
|
||||
def _parse_module(self, file_path: Path) -> Optional[ModuleInfo]:
|
||||
"""Parse a Python module file.
|
||||
|
||||
Args:
|
||||
file_path: Path to Python file
|
||||
|
||||
Returns:
|
||||
ModuleInfo or None if parsing fails
|
||||
"""
|
||||
try:
|
||||
content = file_path.read_text(encoding="utf-8")
|
||||
tree = ast.parse(content)
|
||||
|
||||
# Compute module name from file path
|
||||
rel_path = file_path.relative_to(self.repo_path)
|
||||
module_name = str(rel_path.with_suffix("")).replace("/", ".")
|
||||
|
||||
classes = []
|
||||
functions = []
|
||||
imports = []
|
||||
|
||||
for node in ast.walk(tree):
|
||||
if isinstance(node, ast.Import):
|
||||
for alias in node.names:
|
||||
imports.append(alias.name)
|
||||
elif isinstance(node, ast.ImportFrom):
|
||||
module = node.module or ""
|
||||
for alias in node.names:
|
||||
imports.append(f"{module}.{alias.name}")
|
||||
|
||||
# Get top-level definitions (not in classes)
|
||||
for node in tree.body:
|
||||
if isinstance(node, ast.ClassDef):
|
||||
classes.append(self._parse_class(node))
|
||||
elif isinstance(node, (ast.FunctionDef, ast.AsyncFunctionDef)):
|
||||
functions.append(self._parse_function(node))
|
||||
|
||||
# Get module docstring
|
||||
docstring = ast.get_docstring(tree)
|
||||
|
||||
# Find test coverage
|
||||
test_coverage = self._find_test_file(file_path)
|
||||
|
||||
return ModuleInfo(
|
||||
file_path=str(rel_path),
|
||||
module_name=module_name,
|
||||
classes=classes,
|
||||
functions=functions,
|
||||
imports=imports,
|
||||
docstring=docstring,
|
||||
test_coverage=test_coverage,
|
||||
)
|
||||
|
||||
except SyntaxError as e:
|
||||
logger.warning("Syntax error in %s: %s", file_path, e)
|
||||
return None
|
||||
except Exception as e:
|
||||
logger.error("Failed to parse %s: %s", file_path, e)
|
||||
return None
|
||||
|
||||
def _store_module(self, conn: sqlite3.Connection, module: ModuleInfo, content_hash: str) -> None:
|
||||
"""Store module info in database."""
|
||||
conn.execute(
|
||||
"""
|
||||
INSERT OR REPLACE INTO codebase_index
|
||||
(file_path, module_name, classes, functions, imports, test_coverage,
|
||||
last_indexed, content_hash, docstring)
|
||||
VALUES (?, ?, ?, ?, ?, ?, ?, ?, ?)
|
||||
""",
|
||||
(
|
||||
module.file_path,
|
||||
module.module_name,
|
||||
json.dumps([asdict(c) for c in module.classes]),
|
||||
json.dumps([asdict(f) for f in module.functions]),
|
||||
json.dumps(module.imports),
|
||||
module.test_coverage,
|
||||
datetime.now(timezone.utc).isoformat(),
|
||||
content_hash,
|
||||
module.docstring,
|
||||
),
|
||||
)
|
||||
|
||||
def _build_dependency_graph(self, conn: sqlite3.Connection) -> None:
|
||||
"""Build and store dependency graph from imports."""
|
||||
# Clear existing graph
|
||||
conn.execute("DELETE FROM dependency_graph")
|
||||
|
||||
# Get all modules
|
||||
rows = conn.execute("SELECT file_path, module_name, imports FROM codebase_index").fetchall()
|
||||
|
||||
# Map module names to file paths
|
||||
module_to_file = {row["module_name"]: row["file_path"] for row in rows}
|
||||
|
||||
# Also map without src/ prefix for package imports like myproject.utils
|
||||
module_to_file_alt = {}
|
||||
for row in rows:
|
||||
module_name = row["module_name"]
|
||||
if module_name.startswith("src."):
|
||||
alt_name = module_name[4:] # Remove "src." prefix
|
||||
module_to_file_alt[alt_name] = row["file_path"]
|
||||
|
||||
# Build dependencies
|
||||
for row in rows:
|
||||
source_file = row["file_path"]
|
||||
imports = json.loads(row["imports"])
|
||||
|
||||
for imp in imports:
|
||||
# Try to resolve import to a file
|
||||
# Handle both "module.name" and "module.name.Class" forms
|
||||
|
||||
# First try exact match
|
||||
if imp in module_to_file:
|
||||
conn.execute(
|
||||
"""
|
||||
INSERT OR IGNORE INTO dependency_graph
|
||||
(source_file, target_file, import_type)
|
||||
VALUES (?, ?, ?)
|
||||
""",
|
||||
(source_file, module_to_file[imp], "import"),
|
||||
)
|
||||
continue
|
||||
|
||||
# Try alternative name (without src/ prefix)
|
||||
if imp in module_to_file_alt:
|
||||
conn.execute(
|
||||
"""
|
||||
INSERT OR IGNORE INTO dependency_graph
|
||||
(source_file, target_file, import_type)
|
||||
VALUES (?, ?, ?)
|
||||
""",
|
||||
(source_file, module_to_file_alt[imp], "import"),
|
||||
)
|
||||
continue
|
||||
|
||||
# Try prefix match (import myproject.utils.Helper -> myproject.utils)
|
||||
imp_parts = imp.split(".")
|
||||
for i in range(len(imp_parts), 0, -1):
|
||||
prefix = ".".join(imp_parts[:i])
|
||||
|
||||
# Try original module name
|
||||
if prefix in module_to_file:
|
||||
conn.execute(
|
||||
"""
|
||||
INSERT OR IGNORE INTO dependency_graph
|
||||
(source_file, target_file, import_type)
|
||||
VALUES (?, ?, ?)
|
||||
""",
|
||||
(source_file, module_to_file[prefix], "import"),
|
||||
)
|
||||
break
|
||||
|
||||
# Try alternative name (without src/ prefix)
|
||||
if prefix in module_to_file_alt:
|
||||
conn.execute(
|
||||
"""
|
||||
INSERT OR IGNORE INTO dependency_graph
|
||||
(source_file, target_file, import_type)
|
||||
VALUES (?, ?, ?)
|
||||
""",
|
||||
(source_file, module_to_file_alt[prefix], "import"),
|
||||
)
|
||||
break
|
||||
|
||||
conn.commit()
|
||||
|
||||
async def index_all(self) -> dict[str, int]:
|
||||
"""Perform full reindex of all Python files.
|
||||
|
||||
Returns:
|
||||
Dict with stats: {"indexed": int, "failed": int, "skipped": int}
|
||||
"""
|
||||
logger.info("Starting full codebase index")
|
||||
|
||||
files = self._find_python_files()
|
||||
stats = {"indexed": 0, "failed": 0, "skipped": 0}
|
||||
|
||||
with self._get_conn() as conn:
|
||||
for file_path in files:
|
||||
try:
|
||||
content = file_path.read_text(encoding="utf-8")
|
||||
content_hash = self._compute_hash(content)
|
||||
|
||||
# Check if file needs reindexing
|
||||
existing = conn.execute(
|
||||
"SELECT content_hash FROM codebase_index WHERE file_path = ?",
|
||||
(str(file_path.relative_to(self.repo_path)),),
|
||||
).fetchone()
|
||||
|
||||
if existing and existing["content_hash"] == content_hash:
|
||||
stats["skipped"] += 1
|
||||
continue
|
||||
|
||||
module = self._parse_module(file_path)
|
||||
if module:
|
||||
self._store_module(conn, module, content_hash)
|
||||
stats["indexed"] += 1
|
||||
else:
|
||||
stats["failed"] += 1
|
||||
|
||||
except Exception as e:
|
||||
logger.error("Failed to index %s: %s", file_path, e)
|
||||
stats["failed"] += 1
|
||||
|
||||
# Build dependency graph
|
||||
self._build_dependency_graph(conn)
|
||||
conn.commit()
|
||||
|
||||
logger.info(
|
||||
"Indexing complete: %(indexed)d indexed, %(failed)d failed, %(skipped)d skipped",
|
||||
stats,
|
||||
)
|
||||
return stats
|
||||
|
||||
async def index_changed(self) -> dict[str, int]:
|
||||
"""Perform incremental index of only changed files.
|
||||
|
||||
Compares content hashes to detect changes.
|
||||
|
||||
Returns:
|
||||
Dict with stats: {"indexed": int, "failed": int, "skipped": int}
|
||||
"""
|
||||
logger.info("Starting incremental codebase index")
|
||||
|
||||
files = self._find_python_files()
|
||||
stats = {"indexed": 0, "failed": 0, "skipped": 0}
|
||||
|
||||
with self._get_conn() as conn:
|
||||
for file_path in files:
|
||||
try:
|
||||
rel_path = str(file_path.relative_to(self.repo_path))
|
||||
content = file_path.read_text(encoding="utf-8")
|
||||
content_hash = self._compute_hash(content)
|
||||
|
||||
# Check if changed
|
||||
existing = conn.execute(
|
||||
"SELECT content_hash FROM codebase_index WHERE file_path = ?",
|
||||
(rel_path,),
|
||||
).fetchone()
|
||||
|
||||
if existing and existing["content_hash"] == content_hash:
|
||||
stats["skipped"] += 1
|
||||
continue
|
||||
|
||||
module = self._parse_module(file_path)
|
||||
if module:
|
||||
self._store_module(conn, module, content_hash)
|
||||
stats["indexed"] += 1
|
||||
else:
|
||||
stats["failed"] += 1
|
||||
|
||||
except Exception as e:
|
||||
logger.error("Failed to index %s: %s", file_path, e)
|
||||
stats["failed"] += 1
|
||||
|
||||
# Rebuild dependency graph (some imports may have changed)
|
||||
self._build_dependency_graph(conn)
|
||||
conn.commit()
|
||||
|
||||
logger.info(
|
||||
"Incremental indexing complete: %(indexed)d indexed, %(failed)d failed, %(skipped)d skipped",
|
||||
stats,
|
||||
)
|
||||
return stats
|
||||
|
||||
async def get_summary(self, max_tokens: int = 4000) -> str:
|
||||
"""Generate compressed codebase summary for LLM context.
|
||||
|
||||
Lists modules, their purposes, key classes/functions, and test coverage.
|
||||
Keeps output under max_tokens (approximate).
|
||||
|
||||
Args:
|
||||
max_tokens: Maximum approximate tokens for summary
|
||||
|
||||
Returns:
|
||||
Summary string suitable for LLM context
|
||||
"""
|
||||
with self._get_conn() as conn:
|
||||
rows = conn.execute(
|
||||
"""
|
||||
SELECT file_path, module_name, classes, functions, test_coverage, docstring
|
||||
FROM codebase_index
|
||||
ORDER BY module_name
|
||||
"""
|
||||
).fetchall()
|
||||
|
||||
lines = ["# Codebase Summary\n"]
|
||||
lines.append(f"Total modules: {len(rows)}\n")
|
||||
lines.append("---\n")
|
||||
|
||||
for row in rows:
|
||||
module_name = row["module_name"]
|
||||
file_path = row["file_path"]
|
||||
docstring = row["docstring"]
|
||||
test_coverage = row["test_coverage"]
|
||||
|
||||
lines.append(f"\n## {module_name}")
|
||||
lines.append(f"File: `{file_path}`")
|
||||
|
||||
if test_coverage:
|
||||
lines.append(f"Tests: `{test_coverage}`")
|
||||
else:
|
||||
lines.append("Tests: None")
|
||||
|
||||
if docstring:
|
||||
# Take first line of docstring
|
||||
first_line = docstring.split("\n")[0][:100]
|
||||
lines.append(f"Purpose: {first_line}")
|
||||
|
||||
# Classes
|
||||
classes = json.loads(row["classes"])
|
||||
if classes:
|
||||
lines.append("Classes:")
|
||||
for cls in classes[:5]: # Limit to 5 classes
|
||||
methods = [m["name"] for m in cls["methods"][:3]]
|
||||
method_str = ", ".join(methods) + ("..." if len(cls["methods"]) > 3 else "")
|
||||
lines.append(f" - {cls['name']}({method_str})")
|
||||
if len(classes) > 5:
|
||||
lines.append(f" ... and {len(classes) - 5} more")
|
||||
|
||||
# Functions
|
||||
functions = json.loads(row["functions"])
|
||||
if functions:
|
||||
func_names = [f["name"] for f in functions[:5]]
|
||||
func_str = ", ".join(func_names)
|
||||
if len(functions) > 5:
|
||||
func_str += f"... and {len(functions) - 5} more"
|
||||
lines.append(f"Functions: {func_str}")
|
||||
|
||||
lines.append("")
|
||||
|
||||
summary = "\n".join(lines)
|
||||
|
||||
# Rough token estimation (1 token ≈ 4 characters)
|
||||
if len(summary) > max_tokens * 4:
|
||||
# Truncate with note
|
||||
summary = summary[:max_tokens * 4]
|
||||
summary += "\n\n[Summary truncated due to length]"
|
||||
|
||||
return summary
|
||||
|
||||
async def get_relevant_files(self, task_description: str, limit: int = 5) -> list[str]:
|
||||
"""Find files relevant to a task description.
|
||||
|
||||
Uses keyword matching and import relationships. In Phase 2,
|
||||
this will use semantic search with vector embeddings.
|
||||
|
||||
Args:
|
||||
task_description: Natural language description of the task
|
||||
limit: Maximum number of files to return
|
||||
|
||||
Returns:
|
||||
List of file paths sorted by relevance
|
||||
"""
|
||||
# Simple keyword extraction for now
|
||||
keywords = set(task_description.lower().split())
|
||||
# Remove common words
|
||||
keywords -= {"the", "a", "an", "to", "in", "on", "at", "for", "with", "and", "or", "of", "is", "are"}
|
||||
|
||||
with self._get_conn() as conn:
|
||||
rows = conn.execute(
|
||||
"""
|
||||
SELECT file_path, module_name, classes, functions, docstring, test_coverage
|
||||
FROM codebase_index
|
||||
"""
|
||||
).fetchall()
|
||||
|
||||
scored_files = []
|
||||
|
||||
for row in rows:
|
||||
score = 0
|
||||
file_path = row["file_path"].lower()
|
||||
module_name = row["module_name"].lower()
|
||||
docstring = (row["docstring"] or "").lower()
|
||||
|
||||
classes = json.loads(row["classes"])
|
||||
functions = json.loads(row["functions"])
|
||||
|
||||
# Score based on keyword matches
|
||||
for keyword in keywords:
|
||||
if keyword in file_path:
|
||||
score += 3
|
||||
if keyword in module_name:
|
||||
score += 2
|
||||
if keyword in docstring:
|
||||
score += 2
|
||||
|
||||
# Check class/function names
|
||||
for cls in classes:
|
||||
if keyword in cls["name"].lower():
|
||||
score += 2
|
||||
for method in cls["methods"]:
|
||||
if keyword in method["name"].lower():
|
||||
score += 1
|
||||
|
||||
for func in functions:
|
||||
if keyword in func["name"].lower():
|
||||
score += 1
|
||||
|
||||
# Boost files with test coverage (only if already matched)
|
||||
if score > 0 and row["test_coverage"]:
|
||||
score += 1
|
||||
|
||||
if score > 0:
|
||||
scored_files.append((score, row["file_path"]))
|
||||
|
||||
# Sort by score descending, return top N
|
||||
scored_files.sort(reverse=True, key=lambda x: x[0])
|
||||
return [f[1] for f in scored_files[:limit]]
|
||||
|
||||
async def get_dependency_chain(self, file_path: str) -> list[str]:
|
||||
"""Get all files that import the given file.
|
||||
|
||||
Useful for understanding blast radius of changes.
|
||||
|
||||
Args:
|
||||
file_path: Path to file (relative to repo root)
|
||||
|
||||
Returns:
|
||||
List of file paths that import this file
|
||||
"""
|
||||
with self._get_conn() as conn:
|
||||
rows = conn.execute(
|
||||
"""
|
||||
SELECT source_file FROM dependency_graph
|
||||
WHERE target_file = ?
|
||||
""",
|
||||
(file_path,),
|
||||
).fetchall()
|
||||
|
||||
return [row["source_file"] for row in rows]
|
||||
|
||||
async def has_test_coverage(self, file_path: str) -> bool:
|
||||
"""Check if a file has corresponding test coverage.
|
||||
|
||||
Args:
|
||||
file_path: Path to file (relative to repo root)
|
||||
|
||||
Returns:
|
||||
True if test file exists, False otherwise
|
||||
"""
|
||||
with self._get_conn() as conn:
|
||||
row = conn.execute(
|
||||
"SELECT test_coverage FROM codebase_index WHERE file_path = ?",
|
||||
(file_path,),
|
||||
).fetchone()
|
||||
|
||||
return row is not None and row["test_coverage"] is not None
|
||||
|
||||
async def get_module_info(self, file_path: str) -> Optional[ModuleInfo]:
|
||||
"""Get detailed info for a specific module.
|
||||
|
||||
Args:
|
||||
file_path: Path to file (relative to repo root)
|
||||
|
||||
Returns:
|
||||
ModuleInfo or None if not indexed
|
||||
"""
|
||||
with self._get_conn() as conn:
|
||||
row = conn.execute(
|
||||
"""
|
||||
SELECT file_path, module_name, classes, functions, imports,
|
||||
test_coverage, docstring
|
||||
FROM codebase_index
|
||||
WHERE file_path = ?
|
||||
""",
|
||||
(file_path,),
|
||||
).fetchone()
|
||||
|
||||
if not row:
|
||||
return None
|
||||
|
||||
# Parse classes - convert dict methods to FunctionInfo objects
|
||||
classes_data = json.loads(row["classes"])
|
||||
classes = []
|
||||
for cls_data in classes_data:
|
||||
methods = [FunctionInfo(**m) for m in cls_data.get("methods", [])]
|
||||
cls_info = ClassInfo(
|
||||
name=cls_data["name"],
|
||||
methods=methods,
|
||||
docstring=cls_data.get("docstring"),
|
||||
line_number=cls_data.get("line_number", 0),
|
||||
bases=cls_data.get("bases", []),
|
||||
)
|
||||
classes.append(cls_info)
|
||||
|
||||
# Parse functions
|
||||
functions_data = json.loads(row["functions"])
|
||||
functions = [FunctionInfo(**f) for f in functions_data]
|
||||
|
||||
return ModuleInfo(
|
||||
file_path=row["file_path"],
|
||||
module_name=row["module_name"],
|
||||
classes=classes,
|
||||
functions=functions,
|
||||
imports=json.loads(row["imports"]),
|
||||
docstring=row["docstring"],
|
||||
test_coverage=row["test_coverage"],
|
||||
)
|
||||
505
src/self_coding/git_safety.py
Normal file
505
src/self_coding/git_safety.py
Normal file
@@ -0,0 +1,505 @@
|
||||
"""Git Safety Layer — Atomic git operations with rollback capability.
|
||||
|
||||
All self-modifications happen on feature branches. Only merge to main after
|
||||
full test suite passes. Snapshots enable rollback on failure.
|
||||
"""
|
||||
|
||||
from __future__ import annotations
|
||||
|
||||
import asyncio
|
||||
import hashlib
|
||||
import logging
|
||||
import subprocess
|
||||
from dataclasses import dataclass
|
||||
from datetime import datetime, timezone
|
||||
from pathlib import Path
|
||||
from typing import Optional
|
||||
|
||||
logger = logging.getLogger(__name__)
|
||||
|
||||
|
||||
@dataclass(frozen=True)
|
||||
class Snapshot:
|
||||
"""Immutable snapshot of repository state before modification.
|
||||
|
||||
Attributes:
|
||||
commit_hash: Git commit hash at snapshot time
|
||||
branch: Current branch name
|
||||
timestamp: When snapshot was taken
|
||||
test_status: Whether tests were passing at snapshot time
|
||||
test_output: Pytest output from test run
|
||||
clean: Whether working directory was clean
|
||||
"""
|
||||
commit_hash: str
|
||||
branch: str
|
||||
timestamp: datetime
|
||||
test_status: bool
|
||||
test_output: str
|
||||
clean: bool
|
||||
|
||||
|
||||
class GitSafetyError(Exception):
|
||||
"""Base exception for git safety operations."""
|
||||
pass
|
||||
|
||||
|
||||
class GitNotRepositoryError(GitSafetyError):
|
||||
"""Raised when operation is attempted outside a git repository."""
|
||||
pass
|
||||
|
||||
|
||||
class GitDirtyWorkingDirectoryError(GitSafetyError):
|
||||
"""Raised when working directory is not clean and clean_required=True."""
|
||||
pass
|
||||
|
||||
|
||||
class GitOperationError(GitSafetyError):
|
||||
"""Raised when a git operation fails."""
|
||||
pass
|
||||
|
||||
|
||||
class GitSafety:
|
||||
"""Safe git operations for self-modification workflows.
|
||||
|
||||
All operations are atomic and support rollback. Self-modifications happen
|
||||
on feature branches named 'timmy/self-edit/{timestamp}'. Only merged to
|
||||
main after tests pass.
|
||||
|
||||
Usage:
|
||||
safety = GitSafety(repo_path="/path/to/repo")
|
||||
|
||||
# Take snapshot before changes
|
||||
snapshot = await safety.snapshot()
|
||||
|
||||
# Create feature branch
|
||||
branch = await safety.create_branch(f"timmy/self-edit/{timestamp}")
|
||||
|
||||
# Make changes, commit them
|
||||
await safety.commit("Add error handling", ["src/file.py"])
|
||||
|
||||
# Run tests, merge if pass
|
||||
if tests_pass:
|
||||
await safety.merge_to_main(branch)
|
||||
else:
|
||||
await safety.rollback(snapshot)
|
||||
"""
|
||||
|
||||
def __init__(
|
||||
self,
|
||||
repo_path: Optional[str | Path] = None,
|
||||
main_branch: str = "main",
|
||||
test_command: str = "python -m pytest --tb=short -q",
|
||||
) -> None:
|
||||
"""Initialize GitSafety with repository path.
|
||||
|
||||
Args:
|
||||
repo_path: Path to git repository. Defaults to current working directory.
|
||||
main_branch: Name of main branch (main, master, etc.)
|
||||
test_command: Command to run tests for snapshot validation
|
||||
"""
|
||||
self.repo_path = Path(repo_path).resolve() if repo_path else Path.cwd()
|
||||
self.main_branch = main_branch
|
||||
self.test_command = test_command
|
||||
self._verify_git_repo()
|
||||
logger.info("GitSafety initialized for %s", self.repo_path)
|
||||
|
||||
def _verify_git_repo(self) -> None:
|
||||
"""Verify that repo_path is a git repository."""
|
||||
git_dir = self.repo_path / ".git"
|
||||
if not git_dir.exists():
|
||||
raise GitNotRepositoryError(
|
||||
f"{self.repo_path} is not a git repository"
|
||||
)
|
||||
|
||||
async def _run_git(
|
||||
self,
|
||||
*args: str,
|
||||
check: bool = True,
|
||||
capture_output: bool = True,
|
||||
timeout: float = 30.0,
|
||||
) -> subprocess.CompletedProcess:
|
||||
"""Run a git command asynchronously.
|
||||
|
||||
Args:
|
||||
*args: Git command arguments
|
||||
check: Whether to raise on non-zero exit
|
||||
capture_output: Whether to capture stdout/stderr
|
||||
timeout: Maximum time to wait for command
|
||||
|
||||
Returns:
|
||||
CompletedProcess with returncode, stdout, stderr
|
||||
|
||||
Raises:
|
||||
GitOperationError: If git command fails and check=True
|
||||
"""
|
||||
cmd = ["git", *args]
|
||||
logger.debug("Running: %s", " ".join(cmd))
|
||||
|
||||
try:
|
||||
proc = await asyncio.create_subprocess_exec(
|
||||
*cmd,
|
||||
cwd=self.repo_path,
|
||||
stdout=asyncio.subprocess.PIPE if capture_output else None,
|
||||
stderr=asyncio.subprocess.PIPE if capture_output else None,
|
||||
)
|
||||
|
||||
stdout, stderr = await asyncio.wait_for(
|
||||
proc.communicate(),
|
||||
timeout=timeout,
|
||||
)
|
||||
|
||||
result = subprocess.CompletedProcess(
|
||||
args=cmd,
|
||||
returncode=proc.returncode or 0,
|
||||
stdout=stdout.decode() if stdout else "",
|
||||
stderr=stderr.decode() if stderr else "",
|
||||
)
|
||||
|
||||
if check and result.returncode != 0:
|
||||
raise GitOperationError(
|
||||
f"Git command failed: {' '.join(args)}\n"
|
||||
f"stdout: {result.stdout}\nstderr: {result.stderr}"
|
||||
)
|
||||
|
||||
return result
|
||||
|
||||
except asyncio.TimeoutError as e:
|
||||
proc.kill()
|
||||
raise GitOperationError(f"Git command timed out after {timeout}s: {' '.join(args)}") from e
|
||||
|
||||
async def _run_shell(
|
||||
self,
|
||||
command: str,
|
||||
timeout: float = 120.0,
|
||||
) -> subprocess.CompletedProcess:
|
||||
"""Run a shell command asynchronously.
|
||||
|
||||
Args:
|
||||
command: Shell command to run
|
||||
timeout: Maximum time to wait
|
||||
|
||||
Returns:
|
||||
CompletedProcess with returncode, stdout, stderr
|
||||
"""
|
||||
logger.debug("Running shell: %s", command)
|
||||
|
||||
proc = await asyncio.create_subprocess_shell(
|
||||
command,
|
||||
cwd=self.repo_path,
|
||||
stdout=asyncio.subprocess.PIPE,
|
||||
stderr=asyncio.subprocess.PIPE,
|
||||
)
|
||||
|
||||
stdout, stderr = await asyncio.wait_for(
|
||||
proc.communicate(),
|
||||
timeout=timeout,
|
||||
)
|
||||
|
||||
return subprocess.CompletedProcess(
|
||||
args=command,
|
||||
returncode=proc.returncode or 0,
|
||||
stdout=stdout.decode(),
|
||||
stderr=stderr.decode(),
|
||||
)
|
||||
|
||||
async def is_clean(self) -> bool:
|
||||
"""Check if working directory is clean (no uncommitted changes).
|
||||
|
||||
Returns:
|
||||
True if clean, False if there are uncommitted changes
|
||||
"""
|
||||
result = await self._run_git("status", "--porcelain", check=False)
|
||||
return result.stdout.strip() == ""
|
||||
|
||||
async def get_current_branch(self) -> str:
|
||||
"""Get current git branch name.
|
||||
|
||||
Returns:
|
||||
Current branch name
|
||||
"""
|
||||
result = await self._run_git("branch", "--show-current")
|
||||
return result.stdout.strip()
|
||||
|
||||
async def get_current_commit(self) -> str:
|
||||
"""Get current commit hash.
|
||||
|
||||
Returns:
|
||||
Full commit hash
|
||||
"""
|
||||
result = await self._run_git("rev-parse", "HEAD")
|
||||
return result.stdout.strip()
|
||||
|
||||
async def _run_tests(self) -> tuple[bool, str]:
|
||||
"""Run test suite and return results.
|
||||
|
||||
Returns:
|
||||
Tuple of (all_passed, test_output)
|
||||
"""
|
||||
logger.info("Running tests: %s", self.test_command)
|
||||
result = await self._run_shell(self.test_command, timeout=300.0)
|
||||
passed = result.returncode == 0
|
||||
output = result.stdout + "\n" + result.stderr
|
||||
|
||||
if passed:
|
||||
logger.info("Tests passed")
|
||||
else:
|
||||
logger.warning("Tests failed with returncode %d", result.returncode)
|
||||
|
||||
return passed, output
|
||||
|
||||
async def snapshot(self, run_tests: bool = True) -> Snapshot:
|
||||
"""Take a snapshot of current repository state.
|
||||
|
||||
Captures commit hash, branch, test status. Used for rollback if
|
||||
modifications fail.
|
||||
|
||||
Args:
|
||||
run_tests: Whether to run tests as part of snapshot
|
||||
|
||||
Returns:
|
||||
Snapshot object with current state
|
||||
|
||||
Raises:
|
||||
GitOperationError: If git operations fail
|
||||
"""
|
||||
logger.info("Taking snapshot of repository state")
|
||||
|
||||
commit_hash = await self.get_current_commit()
|
||||
branch = await self.get_current_branch()
|
||||
clean = await self.is_clean()
|
||||
timestamp = datetime.now(timezone.utc)
|
||||
|
||||
test_status = False
|
||||
test_output = ""
|
||||
|
||||
if run_tests:
|
||||
test_status, test_output = await self._run_tests()
|
||||
else:
|
||||
test_status = True # Assume OK if not running tests
|
||||
test_output = "Tests skipped"
|
||||
|
||||
snapshot = Snapshot(
|
||||
commit_hash=commit_hash,
|
||||
branch=branch,
|
||||
timestamp=timestamp,
|
||||
test_status=test_status,
|
||||
test_output=test_output,
|
||||
clean=clean,
|
||||
)
|
||||
|
||||
logger.info(
|
||||
"Snapshot taken: %s@%s (clean=%s, tests=%s)",
|
||||
branch,
|
||||
commit_hash[:8],
|
||||
clean,
|
||||
test_status,
|
||||
)
|
||||
|
||||
return snapshot
|
||||
|
||||
async def create_branch(self, name: str, base: Optional[str] = None) -> str:
|
||||
"""Create and checkout a new feature branch.
|
||||
|
||||
Args:
|
||||
name: Branch name (e.g., 'timmy/self-edit/20260226-143022')
|
||||
base: Base branch to create from (defaults to main_branch)
|
||||
|
||||
Returns:
|
||||
Name of created branch
|
||||
|
||||
Raises:
|
||||
GitOperationError: If branch creation fails
|
||||
"""
|
||||
base = base or self.main_branch
|
||||
|
||||
# Ensure we're on base branch and it's up to date
|
||||
await self._run_git("checkout", base)
|
||||
await self._run_git("pull", "origin", base, check=False) # May fail if no remote
|
||||
|
||||
# Create and checkout new branch
|
||||
await self._run_git("checkout", "-b", name)
|
||||
|
||||
logger.info("Created branch %s from %s", name, base)
|
||||
return name
|
||||
|
||||
async def commit(
|
||||
self,
|
||||
message: str,
|
||||
files: Optional[list[str | Path]] = None,
|
||||
allow_empty: bool = False,
|
||||
) -> str:
|
||||
"""Commit changes to current branch.
|
||||
|
||||
Args:
|
||||
message: Commit message
|
||||
files: Specific files to commit (None = all changes)
|
||||
allow_empty: Whether to allow empty commits
|
||||
|
||||
Returns:
|
||||
Commit hash of new commit
|
||||
|
||||
Raises:
|
||||
GitOperationError: If commit fails
|
||||
"""
|
||||
# Add files
|
||||
if files:
|
||||
for file_path in files:
|
||||
full_path = self.repo_path / file_path
|
||||
if not full_path.exists():
|
||||
logger.warning("File does not exist: %s", file_path)
|
||||
await self._run_git("add", str(file_path))
|
||||
else:
|
||||
await self._run_git("add", "-A")
|
||||
|
||||
# Check if there's anything to commit
|
||||
if not allow_empty:
|
||||
diff_result = await self._run_git(
|
||||
"diff", "--cached", "--quiet", check=False
|
||||
)
|
||||
if diff_result.returncode == 0:
|
||||
logger.warning("No changes to commit")
|
||||
return await self.get_current_commit()
|
||||
|
||||
# Commit
|
||||
commit_args = ["commit", "-m", message]
|
||||
if allow_empty:
|
||||
commit_args.append("--allow-empty")
|
||||
|
||||
await self._run_git(*commit_args)
|
||||
|
||||
commit_hash = await self.get_current_commit()
|
||||
logger.info("Committed %s: %s", commit_hash[:8], message)
|
||||
|
||||
return commit_hash
|
||||
|
||||
async def get_diff(self, from_hash: str, to_hash: Optional[str] = None) -> str:
|
||||
"""Get diff between commits.
|
||||
|
||||
Args:
|
||||
from_hash: Starting commit hash (or Snapshot object hash)
|
||||
to_hash: Ending commit hash (None = current)
|
||||
|
||||
Returns:
|
||||
Git diff as string
|
||||
"""
|
||||
args = ["diff", from_hash]
|
||||
if to_hash:
|
||||
args.append(to_hash)
|
||||
|
||||
result = await self._run_git(*args)
|
||||
return result.stdout
|
||||
|
||||
async def rollback(self, snapshot: Snapshot | str) -> str:
|
||||
"""Rollback to a previous snapshot.
|
||||
|
||||
Hard resets to the snapshot commit and deletes any uncommitted changes.
|
||||
Use with caution — this is destructive.
|
||||
|
||||
Args:
|
||||
snapshot: Snapshot object or commit hash to rollback to
|
||||
|
||||
Returns:
|
||||
Commit hash after rollback
|
||||
|
||||
Raises:
|
||||
GitOperationError: If rollback fails
|
||||
"""
|
||||
if isinstance(snapshot, Snapshot):
|
||||
target_hash = snapshot.commit_hash
|
||||
target_branch = snapshot.branch
|
||||
else:
|
||||
target_hash = snapshot
|
||||
target_branch = None
|
||||
|
||||
logger.warning("Rolling back to %s", target_hash[:8])
|
||||
|
||||
# Reset to target commit
|
||||
await self._run_git("reset", "--hard", target_hash)
|
||||
|
||||
# Clean any untracked files
|
||||
await self._run_git("clean", "-fd")
|
||||
|
||||
# If we know the original branch, switch back to it
|
||||
if target_branch:
|
||||
branch_exists = await self._run_git(
|
||||
"branch", "--list", target_branch, check=False
|
||||
)
|
||||
if branch_exists.stdout.strip():
|
||||
await self._run_git("checkout", target_branch)
|
||||
logger.info("Switched back to branch %s", target_branch)
|
||||
|
||||
current = await self.get_current_commit()
|
||||
logger.info("Rolled back to %s", current[:8])
|
||||
|
||||
return current
|
||||
|
||||
async def merge_to_main(
|
||||
self,
|
||||
branch: str,
|
||||
require_tests: bool = True,
|
||||
) -> str:
|
||||
"""Merge a feature branch into main after tests pass.
|
||||
|
||||
Args:
|
||||
branch: Feature branch to merge
|
||||
require_tests: Whether to require tests to pass before merging
|
||||
|
||||
Returns:
|
||||
Merge commit hash
|
||||
|
||||
Raises:
|
||||
GitOperationError: If merge fails or tests don't pass
|
||||
"""
|
||||
logger.info("Preparing to merge %s into %s", branch, self.main_branch)
|
||||
|
||||
# Checkout the feature branch and run tests
|
||||
await self._run_git("checkout", branch)
|
||||
|
||||
if require_tests:
|
||||
passed, output = await self._run_tests()
|
||||
if not passed:
|
||||
raise GitOperationError(
|
||||
f"Cannot merge {branch}: tests failed\n{output}"
|
||||
)
|
||||
|
||||
# Checkout main and merge
|
||||
await self._run_git("checkout", self.main_branch)
|
||||
await self._run_git("merge", "--no-ff", "-m", f"Merge {branch}", branch)
|
||||
|
||||
# Optionally delete the feature branch
|
||||
await self._run_git("branch", "-d", branch, check=False)
|
||||
|
||||
merge_hash = await self.get_current_commit()
|
||||
logger.info("Merged %s into %s: %s", branch, self.main_branch, merge_hash[:8])
|
||||
|
||||
return merge_hash
|
||||
|
||||
async def get_modified_files(self, since_hash: Optional[str] = None) -> list[str]:
|
||||
"""Get list of files modified since a commit.
|
||||
|
||||
Args:
|
||||
since_hash: Commit to compare against (None = uncommitted changes)
|
||||
|
||||
Returns:
|
||||
List of modified file paths
|
||||
"""
|
||||
if since_hash:
|
||||
result = await self._run_git(
|
||||
"diff", "--name-only", since_hash, "HEAD"
|
||||
)
|
||||
else:
|
||||
result = await self._run_git(
|
||||
"diff", "--name-only", "HEAD"
|
||||
)
|
||||
|
||||
files = [f.strip() for f in result.stdout.split("\n") if f.strip()]
|
||||
return files
|
||||
|
||||
async def stage_file(self, file_path: str | Path) -> None:
|
||||
"""Stage a single file for commit.
|
||||
|
||||
Args:
|
||||
file_path: Path to file relative to repo root
|
||||
"""
|
||||
await self._run_git("add", str(file_path))
|
||||
logger.debug("Staged %s", file_path)
|
||||
425
src/self_coding/modification_journal.py
Normal file
425
src/self_coding/modification_journal.py
Normal file
@@ -0,0 +1,425 @@
|
||||
"""Modification Journal — Persistent log of self-modification attempts.
|
||||
|
||||
Tracks successes and failures so Timmy can learn from experience.
|
||||
Supports semantic search for similar past attempts.
|
||||
"""
|
||||
|
||||
from __future__ import annotations
|
||||
|
||||
import json
|
||||
import logging
|
||||
import sqlite3
|
||||
from dataclasses import dataclass, field
|
||||
from datetime import datetime, timezone
|
||||
from enum import Enum
|
||||
from pathlib import Path
|
||||
from typing import Optional
|
||||
|
||||
logger = logging.getLogger(__name__)
|
||||
|
||||
# Default database location
|
||||
DEFAULT_DB_PATH = Path("data/self_coding.db")
|
||||
|
||||
|
||||
class Outcome(str, Enum):
|
||||
"""Possible outcomes of a modification attempt."""
|
||||
SUCCESS = "success"
|
||||
FAILURE = "failure"
|
||||
ROLLBACK = "rollback"
|
||||
|
||||
|
||||
@dataclass
|
||||
class ModificationAttempt:
|
||||
"""A single self-modification attempt.
|
||||
|
||||
Attributes:
|
||||
id: Unique identifier (auto-generated by database)
|
||||
timestamp: When the attempt was made
|
||||
task_description: What was Timmy trying to do
|
||||
approach: Strategy/approach planned
|
||||
files_modified: List of file paths that were modified
|
||||
diff: The actual git diff of changes
|
||||
test_results: Pytest output
|
||||
outcome: success, failure, or rollback
|
||||
failure_analysis: LLM-generated analysis of why it failed
|
||||
reflection: LLM-generated lessons learned
|
||||
retry_count: Number of retry attempts
|
||||
embedding: Vector embedding of task_description (for semantic search)
|
||||
"""
|
||||
task_description: str
|
||||
approach: str = ""
|
||||
files_modified: list[str] = field(default_factory=list)
|
||||
diff: str = ""
|
||||
test_results: str = ""
|
||||
outcome: Outcome = Outcome.FAILURE
|
||||
failure_analysis: str = ""
|
||||
reflection: str = ""
|
||||
retry_count: int = 0
|
||||
id: Optional[int] = None
|
||||
timestamp: Optional[datetime] = None
|
||||
embedding: Optional[bytes] = None
|
||||
|
||||
|
||||
class ModificationJournal:
|
||||
"""Persistent log of self-modification attempts.
|
||||
|
||||
Before any self-modification, Timmy should query the journal for
|
||||
similar past attempts and include relevant ones in the LLM context.
|
||||
|
||||
Usage:
|
||||
journal = ModificationJournal()
|
||||
|
||||
# Log an attempt
|
||||
attempt = ModificationAttempt(
|
||||
task_description="Add error handling",
|
||||
files_modified=["src/app.py"],
|
||||
outcome=Outcome.SUCCESS,
|
||||
)
|
||||
await journal.log_attempt(attempt)
|
||||
|
||||
# Find similar past attempts
|
||||
similar = await journal.find_similar("Add error handling to endpoints")
|
||||
|
||||
# Get success metrics
|
||||
metrics = await journal.get_success_rate()
|
||||
"""
|
||||
|
||||
def __init__(
|
||||
self,
|
||||
db_path: Optional[str | Path] = None,
|
||||
) -> None:
|
||||
"""Initialize ModificationJournal.
|
||||
|
||||
Args:
|
||||
db_path: SQLite database path. Defaults to data/self_coding.db
|
||||
"""
|
||||
self.db_path = Path(db_path) if db_path else DEFAULT_DB_PATH
|
||||
self._ensure_schema()
|
||||
logger.info("ModificationJournal initialized at %s", self.db_path)
|
||||
|
||||
def _get_conn(self) -> sqlite3.Connection:
|
||||
"""Get database connection with schema ensured."""
|
||||
self.db_path.parent.mkdir(parents=True, exist_ok=True)
|
||||
conn = sqlite3.connect(str(self.db_path))
|
||||
conn.row_factory = sqlite3.Row
|
||||
return conn
|
||||
|
||||
def _ensure_schema(self) -> None:
|
||||
"""Create database tables if they don't exist."""
|
||||
with self._get_conn() as conn:
|
||||
conn.execute(
|
||||
"""
|
||||
CREATE TABLE IF NOT EXISTS modification_journal (
|
||||
id INTEGER PRIMARY KEY AUTOINCREMENT,
|
||||
timestamp DATETIME DEFAULT CURRENT_TIMESTAMP,
|
||||
task_description TEXT NOT NULL,
|
||||
approach TEXT,
|
||||
files_modified JSON,
|
||||
diff TEXT,
|
||||
test_results TEXT,
|
||||
outcome TEXT CHECK(outcome IN ('success', 'failure', 'rollback')),
|
||||
failure_analysis TEXT,
|
||||
reflection TEXT,
|
||||
retry_count INTEGER DEFAULT 0,
|
||||
embedding BLOB
|
||||
)
|
||||
"""
|
||||
)
|
||||
|
||||
# Create indexes for common queries
|
||||
conn.execute(
|
||||
"CREATE INDEX IF NOT EXISTS idx_journal_outcome ON modification_journal(outcome)"
|
||||
)
|
||||
conn.execute(
|
||||
"CREATE INDEX IF NOT EXISTS idx_journal_timestamp ON modification_journal(timestamp)"
|
||||
)
|
||||
conn.execute(
|
||||
"CREATE INDEX IF NOT EXISTS idx_journal_task ON modification_journal(task_description)"
|
||||
)
|
||||
|
||||
conn.commit()
|
||||
|
||||
async def log_attempt(self, attempt: ModificationAttempt) -> int:
|
||||
"""Log a modification attempt to the journal.
|
||||
|
||||
Args:
|
||||
attempt: The modification attempt to log
|
||||
|
||||
Returns:
|
||||
ID of the logged entry
|
||||
"""
|
||||
with self._get_conn() as conn:
|
||||
cursor = conn.execute(
|
||||
"""
|
||||
INSERT INTO modification_journal
|
||||
(task_description, approach, files_modified, diff, test_results,
|
||||
outcome, failure_analysis, reflection, retry_count, embedding)
|
||||
VALUES (?, ?, ?, ?, ?, ?, ?, ?, ?, ?)
|
||||
""",
|
||||
(
|
||||
attempt.task_description,
|
||||
attempt.approach,
|
||||
json.dumps(attempt.files_modified),
|
||||
attempt.diff,
|
||||
attempt.test_results,
|
||||
attempt.outcome.value,
|
||||
attempt.failure_analysis,
|
||||
attempt.reflection,
|
||||
attempt.retry_count,
|
||||
attempt.embedding,
|
||||
),
|
||||
)
|
||||
conn.commit()
|
||||
|
||||
attempt_id = cursor.lastrowid
|
||||
logger.info(
|
||||
"Logged modification attempt %d: %s (%s)",
|
||||
attempt_id,
|
||||
attempt.task_description[:50],
|
||||
attempt.outcome.value,
|
||||
)
|
||||
return attempt_id
|
||||
|
||||
async def find_similar(
|
||||
self,
|
||||
task_description: str,
|
||||
limit: int = 5,
|
||||
include_outcomes: Optional[list[Outcome]] = None,
|
||||
) -> list[ModificationAttempt]:
|
||||
"""Find similar past modification attempts.
|
||||
|
||||
Uses keyword matching for now. In Phase 2, will use vector embeddings
|
||||
for semantic search.
|
||||
|
||||
Args:
|
||||
task_description: Task to find similar attempts for
|
||||
limit: Maximum number of results
|
||||
include_outcomes: Filter by outcomes (None = all)
|
||||
|
||||
Returns:
|
||||
List of similar modification attempts
|
||||
"""
|
||||
# Extract keywords from task description
|
||||
keywords = set(task_description.lower().split())
|
||||
keywords -= {"the", "a", "an", "to", "in", "on", "at", "for", "with", "and", "or", "of", "is", "are"}
|
||||
|
||||
with self._get_conn() as conn:
|
||||
# Build query
|
||||
if include_outcomes:
|
||||
outcome_filter = "AND outcome IN ({})".format(
|
||||
",".join("?" * len(include_outcomes))
|
||||
)
|
||||
outcome_values = [o.value for o in include_outcomes]
|
||||
else:
|
||||
outcome_filter = ""
|
||||
outcome_values = []
|
||||
|
||||
rows = conn.execute(
|
||||
f"""
|
||||
SELECT id, timestamp, task_description, approach, files_modified,
|
||||
diff, test_results, outcome, failure_analysis, reflection,
|
||||
retry_count
|
||||
FROM modification_journal
|
||||
WHERE 1=1 {outcome_filter}
|
||||
ORDER BY timestamp DESC
|
||||
LIMIT ?
|
||||
""",
|
||||
outcome_values + [limit * 3], # Get more for scoring
|
||||
).fetchall()
|
||||
|
||||
# Score by keyword match
|
||||
scored = []
|
||||
for row in rows:
|
||||
score = 0
|
||||
task = row["task_description"].lower()
|
||||
approach = (row["approach"] or "").lower()
|
||||
|
||||
for kw in keywords:
|
||||
if kw in task:
|
||||
score += 3
|
||||
if kw in approach:
|
||||
score += 1
|
||||
|
||||
# Boost recent attempts (only if already matched)
|
||||
if score > 0:
|
||||
timestamp = datetime.fromisoformat(row["timestamp"])
|
||||
if timestamp.tzinfo is None:
|
||||
timestamp = timestamp.replace(tzinfo=timezone.utc)
|
||||
age_days = (datetime.now(timezone.utc) - timestamp).days
|
||||
if age_days < 7:
|
||||
score += 2
|
||||
elif age_days < 30:
|
||||
score += 1
|
||||
|
||||
if score > 0:
|
||||
scored.append((score, row))
|
||||
|
||||
# Sort by score, take top N
|
||||
scored.sort(reverse=True, key=lambda x: x[0])
|
||||
top_rows = scored[:limit]
|
||||
|
||||
# Convert to ModificationAttempt objects
|
||||
return [self._row_to_attempt(row) for _, row in top_rows]
|
||||
|
||||
async def get_success_rate(self) -> dict[str, float]:
|
||||
"""Get success rate metrics.
|
||||
|
||||
Returns:
|
||||
Dict with overall and per-category success rates:
|
||||
{
|
||||
"overall": float, # 0.0 to 1.0
|
||||
"success": int, # count
|
||||
"failure": int, # count
|
||||
"rollback": int, # count
|
||||
"total": int, # total attempts
|
||||
}
|
||||
"""
|
||||
with self._get_conn() as conn:
|
||||
rows = conn.execute(
|
||||
"""
|
||||
SELECT outcome, COUNT(*) as count
|
||||
FROM modification_journal
|
||||
GROUP BY outcome
|
||||
"""
|
||||
).fetchall()
|
||||
|
||||
counts = {row["outcome"]: row["count"] for row in rows}
|
||||
|
||||
success = counts.get("success", 0)
|
||||
failure = counts.get("failure", 0)
|
||||
rollback = counts.get("rollback", 0)
|
||||
total = success + failure + rollback
|
||||
|
||||
overall = success / total if total > 0 else 0.0
|
||||
|
||||
return {
|
||||
"overall": overall,
|
||||
"success": success,
|
||||
"failure": failure,
|
||||
"rollback": rollback,
|
||||
"total": total,
|
||||
}
|
||||
|
||||
async def get_recent_failures(self, limit: int = 10) -> list[ModificationAttempt]:
|
||||
"""Get recent failed attempts with their analyses.
|
||||
|
||||
Args:
|
||||
limit: Maximum number of failures to return
|
||||
|
||||
Returns:
|
||||
List of failed modification attempts
|
||||
"""
|
||||
with self._get_conn() as conn:
|
||||
rows = conn.execute(
|
||||
"""
|
||||
SELECT id, timestamp, task_description, approach, files_modified,
|
||||
diff, test_results, outcome, failure_analysis, reflection,
|
||||
retry_count
|
||||
FROM modification_journal
|
||||
WHERE outcome IN ('failure', 'rollback')
|
||||
ORDER BY timestamp DESC
|
||||
LIMIT ?
|
||||
""",
|
||||
(limit,),
|
||||
).fetchall()
|
||||
|
||||
return [self._row_to_attempt(row) for row in rows]
|
||||
|
||||
async def get_by_id(self, attempt_id: int) -> Optional[ModificationAttempt]:
|
||||
"""Get a specific modification attempt by ID.
|
||||
|
||||
Args:
|
||||
attempt_id: ID of the attempt
|
||||
|
||||
Returns:
|
||||
ModificationAttempt or None if not found
|
||||
"""
|
||||
with self._get_conn() as conn:
|
||||
row = conn.execute(
|
||||
"""
|
||||
SELECT id, timestamp, task_description, approach, files_modified,
|
||||
diff, test_results, outcome, failure_analysis, reflection,
|
||||
retry_count
|
||||
FROM modification_journal
|
||||
WHERE id = ?
|
||||
""",
|
||||
(attempt_id,),
|
||||
).fetchone()
|
||||
|
||||
if not row:
|
||||
return None
|
||||
|
||||
return self._row_to_attempt(row)
|
||||
|
||||
async def update_reflection(self, attempt_id: int, reflection: str) -> bool:
|
||||
"""Update the reflection for a modification attempt.
|
||||
|
||||
Args:
|
||||
attempt_id: ID of the attempt
|
||||
reflection: New reflection text
|
||||
|
||||
Returns:
|
||||
True if updated, False if not found
|
||||
"""
|
||||
with self._get_conn() as conn:
|
||||
cursor = conn.execute(
|
||||
"""
|
||||
UPDATE modification_journal
|
||||
SET reflection = ?
|
||||
WHERE id = ?
|
||||
""",
|
||||
(reflection, attempt_id),
|
||||
)
|
||||
conn.commit()
|
||||
|
||||
if cursor.rowcount > 0:
|
||||
logger.info("Updated reflection for attempt %d", attempt_id)
|
||||
return True
|
||||
return False
|
||||
|
||||
async def get_attempts_for_file(
|
||||
self,
|
||||
file_path: str,
|
||||
limit: int = 10,
|
||||
) -> list[ModificationAttempt]:
|
||||
"""Get all attempts that modified a specific file.
|
||||
|
||||
Args:
|
||||
file_path: Path to file (relative to repo root)
|
||||
limit: Maximum number of attempts
|
||||
|
||||
Returns:
|
||||
List of modification attempts affecting this file
|
||||
"""
|
||||
with self._get_conn() as conn:
|
||||
# Try exact match first, then partial match
|
||||
rows = conn.execute(
|
||||
"""
|
||||
SELECT id, timestamp, task_description, approach, files_modified,
|
||||
diff, test_results, outcome, failure_analysis, reflection,
|
||||
retry_count
|
||||
FROM modification_journal
|
||||
WHERE files_modified LIKE ? OR files_modified LIKE ?
|
||||
ORDER BY timestamp DESC
|
||||
LIMIT ?
|
||||
""",
|
||||
(f'%"{file_path}"%', f'%{file_path}%', limit),
|
||||
).fetchall()
|
||||
|
||||
return [self._row_to_attempt(row) for row in rows]
|
||||
|
||||
def _row_to_attempt(self, row: sqlite3.Row) -> ModificationAttempt:
|
||||
"""Convert a database row to ModificationAttempt."""
|
||||
return ModificationAttempt(
|
||||
id=row["id"],
|
||||
timestamp=datetime.fromisoformat(row["timestamp"]),
|
||||
task_description=row["task_description"],
|
||||
approach=row["approach"] or "",
|
||||
files_modified=json.loads(row["files_modified"] or "[]"),
|
||||
diff=row["diff"] or "",
|
||||
test_results=row["test_results"] or "",
|
||||
outcome=Outcome(row["outcome"]),
|
||||
failure_analysis=row["failure_analysis"] or "",
|
||||
reflection=row["reflection"] or "",
|
||||
retry_count=row["retry_count"] or 0,
|
||||
)
|
||||
259
src/self_coding/reflection.py
Normal file
259
src/self_coding/reflection.py
Normal file
@@ -0,0 +1,259 @@
|
||||
"""Reflection Service — Generate lessons learned from modification attempts.
|
||||
|
||||
After every self-modification (success or failure), the Reflection Service
|
||||
prompts an LLM to analyze the attempt and extract actionable insights.
|
||||
"""
|
||||
|
||||
from __future__ import annotations
|
||||
|
||||
import logging
|
||||
from typing import Optional
|
||||
|
||||
from self_coding.modification_journal import ModificationAttempt, Outcome
|
||||
|
||||
logger = logging.getLogger(__name__)
|
||||
|
||||
|
||||
REFLECTION_SYSTEM_PROMPT = """You are a software engineering mentor analyzing a self-modification attempt.
|
||||
|
||||
Your goal is to provide constructive, specific feedback that helps improve future attempts.
|
||||
Focus on patterns and principles rather than one-off issues.
|
||||
|
||||
Be concise but insightful. Maximum 300 words."""
|
||||
|
||||
|
||||
REFLECTION_PROMPT_TEMPLATE = """A software agent just attempted to modify its own source code.
|
||||
|
||||
Task: {task_description}
|
||||
Approach: {approach}
|
||||
Files modified: {files_modified}
|
||||
Outcome: {outcome}
|
||||
Test results: {test_results}
|
||||
{failure_section}
|
||||
|
||||
Reflect on this attempt:
|
||||
1. What went well? (Be specific about techniques or strategies)
|
||||
2. What could be improved? (Focus on process, not just the code)
|
||||
3. What would you do differently next time?
|
||||
4. What general lesson can be extracted for future similar tasks?
|
||||
|
||||
Provide your reflection in a structured format:
|
||||
|
||||
**What went well:**
|
||||
[Your analysis]
|
||||
|
||||
**What could be improved:**
|
||||
[Your analysis]
|
||||
|
||||
**Next time:**
|
||||
[Specific actionable change]
|
||||
|
||||
**General lesson:**
|
||||
[Extracted principle for similar tasks]"""
|
||||
|
||||
|
||||
class ReflectionService:
|
||||
"""Generates reflections on self-modification attempts.
|
||||
|
||||
Uses an LLM to analyze attempts and extract lessons learned.
|
||||
Stores reflections in the Modification Journal for future reference.
|
||||
|
||||
Usage:
|
||||
from self_coding.reflection import ReflectionService
|
||||
from timmy.cascade_adapter import TimmyCascadeAdapter
|
||||
|
||||
adapter = TimmyCascadeAdapter()
|
||||
reflection_service = ReflectionService(llm_adapter=adapter)
|
||||
|
||||
# After a modification attempt
|
||||
reflection_text = await reflection_service.reflect_on_attempt(attempt)
|
||||
|
||||
# Store in journal
|
||||
await journal.update_reflection(attempt_id, reflection_text)
|
||||
"""
|
||||
|
||||
def __init__(
|
||||
self,
|
||||
llm_adapter: Optional[object] = None,
|
||||
model_preference: str = "fast", # "fast" or "quality"
|
||||
) -> None:
|
||||
"""Initialize ReflectionService.
|
||||
|
||||
Args:
|
||||
llm_adapter: LLM adapter (e.g., TimmyCascadeAdapter)
|
||||
model_preference: "fast" for quick reflections, "quality" for deeper analysis
|
||||
"""
|
||||
self.llm_adapter = llm_adapter
|
||||
self.model_preference = model_preference
|
||||
logger.info("ReflectionService initialized")
|
||||
|
||||
async def reflect_on_attempt(self, attempt: ModificationAttempt) -> str:
|
||||
"""Generate a reflection on a modification attempt.
|
||||
|
||||
Args:
|
||||
attempt: The modification attempt to reflect on
|
||||
|
||||
Returns:
|
||||
Reflection text (structured markdown)
|
||||
"""
|
||||
# Build the prompt
|
||||
failure_section = ""
|
||||
if attempt.outcome == Outcome.FAILURE and attempt.failure_analysis:
|
||||
failure_section = f"\nFailure analysis: {attempt.failure_analysis}"
|
||||
|
||||
prompt = REFLECTION_PROMPT_TEMPLATE.format(
|
||||
task_description=attempt.task_description,
|
||||
approach=attempt.approach or "(No approach documented)",
|
||||
files_modified=", ".join(attempt.files_modified) if attempt.files_modified else "(No files modified)",
|
||||
outcome=attempt.outcome.value.upper(),
|
||||
test_results=attempt.test_results[:500] if attempt.test_results else "(No test results)",
|
||||
failure_section=failure_section,
|
||||
)
|
||||
|
||||
# Call LLM if available
|
||||
if self.llm_adapter:
|
||||
try:
|
||||
response = await self.llm_adapter.chat(
|
||||
message=prompt,
|
||||
context=REFLECTION_SYSTEM_PROMPT,
|
||||
)
|
||||
reflection = response.content.strip()
|
||||
logger.info("Generated reflection for attempt (via %s)",
|
||||
response.provider_used)
|
||||
return reflection
|
||||
except Exception as e:
|
||||
logger.error("LLM reflection failed: %s", e)
|
||||
return self._generate_fallback_reflection(attempt)
|
||||
else:
|
||||
# No LLM available, use fallback
|
||||
return self._generate_fallback_reflection(attempt)
|
||||
|
||||
def _generate_fallback_reflection(self, attempt: ModificationAttempt) -> str:
|
||||
"""Generate a basic reflection without LLM.
|
||||
|
||||
Used when no LLM adapter is available or LLM call fails.
|
||||
|
||||
Args:
|
||||
attempt: The modification attempt
|
||||
|
||||
Returns:
|
||||
Basic reflection text
|
||||
"""
|
||||
if attempt.outcome == Outcome.SUCCESS:
|
||||
return f"""**What went well:**
|
||||
Successfully completed: {attempt.task_description}
|
||||
Files modified: {', '.join(attempt.files_modified) if attempt.files_modified else 'N/A'}
|
||||
|
||||
**What could be improved:**
|
||||
Document the approach taken for future reference.
|
||||
|
||||
**Next time:**
|
||||
Use the same pattern for similar tasks.
|
||||
|
||||
**General lesson:**
|
||||
Modifications to {', '.join(attempt.files_modified) if attempt.files_modified else 'these files'} should include proper test coverage."""
|
||||
|
||||
elif attempt.outcome == Outcome.FAILURE:
|
||||
return f"""**What went well:**
|
||||
Attempted: {attempt.task_description}
|
||||
|
||||
**What could be improved:**
|
||||
The modification failed after {attempt.retry_count} retries.
|
||||
{attempt.failure_analysis if attempt.failure_analysis else 'Failure reason not documented.'}
|
||||
|
||||
**Next time:**
|
||||
Consider breaking the task into smaller steps.
|
||||
Validate approach with simpler test case first.
|
||||
|
||||
**General lesson:**
|
||||
Changes affecting {', '.join(attempt.files_modified) if attempt.files_modified else 'multiple files'} require careful dependency analysis."""
|
||||
|
||||
else: # ROLLBACK
|
||||
return f"""**What went well:**
|
||||
Recognized failure and rolled back to maintain stability.
|
||||
|
||||
**What could be improved:**
|
||||
Early detection of issues before full implementation.
|
||||
|
||||
**Next time:**
|
||||
Run tests more frequently during development.
|
||||
Use smaller incremental commits.
|
||||
|
||||
**General lesson:**
|
||||
Rollback is preferable to shipping broken code."""
|
||||
|
||||
async def reflect_with_context(
|
||||
self,
|
||||
attempt: ModificationAttempt,
|
||||
similar_attempts: list[ModificationAttempt],
|
||||
) -> str:
|
||||
"""Generate reflection with context from similar past attempts.
|
||||
|
||||
Includes relevant past reflections to build cumulative learning.
|
||||
|
||||
Args:
|
||||
attempt: The current modification attempt
|
||||
similar_attempts: Similar past attempts (with reflections)
|
||||
|
||||
Returns:
|
||||
Reflection text incorporating past learnings
|
||||
"""
|
||||
# Build context from similar attempts
|
||||
context_parts = []
|
||||
for past in similar_attempts[:3]: # Top 3 similar
|
||||
if past.reflection:
|
||||
context_parts.append(
|
||||
f"Past similar task ({past.outcome.value}):\n"
|
||||
f"Task: {past.task_description}\n"
|
||||
f"Lesson: {past.reflection[:200]}..."
|
||||
)
|
||||
|
||||
context = "\n\n".join(context_parts)
|
||||
|
||||
# Build enhanced prompt
|
||||
failure_section = ""
|
||||
if attempt.outcome == Outcome.FAILURE and attempt.failure_analysis:
|
||||
failure_section = f"\nFailure analysis: {attempt.failure_analysis}"
|
||||
|
||||
enhanced_prompt = f"""A software agent just attempted to modify its own source code.
|
||||
|
||||
Task: {attempt.task_description}
|
||||
Approach: {attempt.approach or "(No approach documented)"}
|
||||
Files modified: {', '.join(attempt.files_modified) if attempt.files_modified else "(No files modified)"}
|
||||
Outcome: {attempt.outcome.value.upper()}
|
||||
Test results: {attempt.test_results[:500] if attempt.test_results else "(No test results)"}
|
||||
{failure_section}
|
||||
|
||||
---
|
||||
|
||||
Relevant past attempts:
|
||||
|
||||
{context if context else "(No similar past attempts)"}
|
||||
|
||||
---
|
||||
|
||||
Given this history, reflect on the current attempt:
|
||||
1. What went well?
|
||||
2. What could be improved?
|
||||
3. How does this compare to past similar attempts?
|
||||
4. What pattern or principle should guide future similar tasks?
|
||||
|
||||
Provide your reflection in a structured format:
|
||||
|
||||
**What went well:**
|
||||
**What could be improved:**
|
||||
**Comparison to past attempts:**
|
||||
**Guiding principle:**"""
|
||||
|
||||
if self.llm_adapter:
|
||||
try:
|
||||
response = await self.llm_adapter.chat(
|
||||
message=enhanced_prompt,
|
||||
context=REFLECTION_SYSTEM_PROMPT,
|
||||
)
|
||||
return response.content.strip()
|
||||
except Exception as e:
|
||||
logger.error("LLM reflection with context failed: %s", e)
|
||||
return await self.reflect_on_attempt(attempt)
|
||||
else:
|
||||
return await self.reflect_on_attempt(attempt)
|
||||
Reference in New Issue
Block a user