Edit Strategies & the Test-Driven Repair Loop

import pathlib, shutil

REPO = pathlib.Path("/sandbox/repo")

def apply_edit(rel: str, old: str, new: str) -> str:
    """Replace an exact block. Returns a clear error string on any mismatch."""
    target = (REPO / rel).resolve()
    if REPO not in target.parents:
        return "error: path escapes the repo"
    if not target.is_file():
        return f"error: {rel} does not exist"

    text = target.read_text()
    count = text.count(old)
    if count == 0:
        # The single most common failure: block drifted. Tell the model to re-read.
        return ("error: old block not found. Re-read the file and copy the "
                "exact current text (including indentation) before editing.")
    if count > 1:
        return (f"error: old block appears {count} times; include more "
                "surrounding context to make it unique.")

    backup = target.with_suffix(target.suffix + ".bak")
    shutil.copy(target, backup)              # cheap rollback point
    target.write_text(text.replace(old, new, 1))
    return f"applied: 1 replacement in {rel}"

import subprocess, pathlib

REPO = pathlib.Path("/sandbox/repo")

def run_tests(target: str = "") -> dict:
    """Run pytest in the sandbox; return pass/fail + trimmed output."""
    cmd = ["python", "-m", "pytest", "-q", "--no-header"]
    if target:
        cmd.append(target)
    proc = subprocess.run(
        cmd, cwd=REPO, capture_output=True, text=True, timeout=600,
    )
    output = proc.stdout + proc.stderr
    # Trim so a huge traceback dump doesn't blow the context budget.
    tail = "\n".join(output.splitlines()[-60:])
    return {
        "passed": proc.returncode == 0,
        "returncode": proc.returncode,
        "output_tail": tail,
    }

MAX_ATTEMPTS = 5

def repair(plan: dict, issue_text: str) -> dict:
    system = (
        "Fix the bug per the plan. FIRST write a test that reproduces the "
        "issue and fails (red). Then edit source with apply_edit until that "
        "test AND all existing tests pass (green). Read failing output before "
        "each edit. Tools: read_file, apply_edit, run_tests."
    )
    messages = [{"role": "user",
                 "content": f"Issue:\n{issue_text}\n\nPlan:\n{plan}"}]

    for attempt in range(1, MAX_ATTEMPTS + 1):
        resp = call_model(system, REPAIR_TOOLS, messages)
        messages.append({"role": "assistant", "content": resp.content})

        if resp.stop_reason != "tool_use":
            # Model thinks it's done — verify independently, never take its word.
            final = run_tests()
            if final["passed"]:
                return {"status": "success", "attempts": attempt}
            messages.append({"role": "user", "content":
                f"You stopped but tests still fail:\n{final['output_tail']}\n"
                "Keep fixing."})
            continue

        results = []
        for block in resp.content:
            if block.type == "tool_use":
                results.append(run_repair_tool(block))   # dispatches to the 3 tools
        messages.append({"role": "user", "content": results})

    # Bounded: give up cleanly rather than loop forever.
    return {"status": "exhausted", "attempts": MAX_ATTEMPTS,
            "last_tests": run_tests()}