homegraph 1.4.1 → 1.5.1

package/scripts/exp_boundary_eval/run_one.py

@@ -1,165 +1,165 @@
-#!/usr/bin/env python3
-"""Run a single one-shot experiment. Supports multiple agents via --agent."""
-
-import os
-import subprocess
-import sys
-from pathlib import Path
-from typing import Optional
-
-import setup as setup_mod
-from _utils import (OUTPUT_DIR, append_run_log, build_agent_cmd, current_time_ms,
-                    error, find_experiment, get_agent, info, header,
-                    parse_output, resolve_agent_model, run_monitored_subprocess,
-                    subprocess_no_window_kwargs, warn, write_json)
-from deveco_arm import analyze_homegraph_stream
-
-
-def run_experiment(exp_id: str, skip_permissions: bool = True,
-                   agent_name: str = "",
-                   results_dir: Optional[Path] = None,
-                   state_dir: Optional[Path] = None) -> bool:
-    exp = find_experiment(exp_id)
-    if exp is None:
-        error(f"Experiment {exp_id} not found in config")
-        return False
-    if exp.get("type") == "session":
-        error(f"Experiment {exp_id} is session type. Use run_session.py.")
-        return False
-
-    agent = get_agent(agent_name)
-    clone_path = os.environ.get("REPO_CLONE", "")
-    local_repo = os.environ.get("REPO_PATH", "")
-    if clone_path:
-        repo_root = Path(clone_path)
-    elif local_repo:
-        repo_root = Path(local_repo)
-    else:
-        error("No repo. Use run_all.py to clone, or set REPO_PATH / REPO_CLONE.")
-        return False
-
-    title = exp["title"]
-    prompt = exp["prompt"]
-    max_turns = exp.get("max_turns", 15)
-
-    header(f"Experiment {exp_id}: {title}")
-    info(f"Agent: {agent['name']} ({agent['binary']}) | Repo: {repo_root}")
-    info(f"Max turns: {max_turns} | Permissions: {'skip' if skip_permissions else 'ask'}")
-
-    session_id, _ = setup_mod.prepare_environment(exp_id, "full", agent_name,
-                                                    state_dir=state_dir)
-
-    result_dir = results_dir / exp_id if results_dir else OUTPUT_DIR / exp_id
-    result_dir.mkdir(parents=True, exist_ok=True)
-    (result_dir / "prompt.txt").write_text(prompt, encoding="utf-8")
-    write_json(result_dir / "experiment_config.json", {**exp, "agent": agent_name or "default"})
-
-    info("Starting agent...")
-    start_ms = current_time_ms()
-    os.chdir(repo_root)
-
-    cmd = build_agent_cmd(agent, session_id, max_turns, prompt,
-                          resume=False, skip_permissions=skip_permissions,
-                          repo_root=repo_root)
-
-    stream_path = result_dir / "stream_output.jsonl"
-    text_path = result_dir / "raw_output.txt"
-    track_hg = os.environ.get("EVAL_ARM", "") == "homegraph"
-
-    with open(stream_path, "w", encoding="utf-8") as stream_f:
-        result, mem = run_monitored_subprocess(
-            cmd, stdout=stream_f, stderr=subprocess.STDOUT, text=True,
-            track_homegraph=track_hg,
-        )
-    exit_code = result.returncode
-
-    end_ms = current_time_ms()
-    dur_ms = end_ms - start_ms
-    dur_s = f"{dur_ms / 1000:.1f}"
-    info(f"Completed in {dur_s}s (exit={exit_code})")
-
-    # Parse agent output
-    parser_name = agent.get("parser", "claude_stream_json")
-    stats = parse_output(stream_path, parser_name)
-
-    if exit_code != 0:
-        error(f"Agent 退出码 {exit_code}")
-        for err in stats.errors[:3]:
-            error(f"  {err}")
-        if "Unexpected server error" in " ".join(stats.errors):
-            warn("DevEco 服务端报错 — 通常为 API/网络/配额问题，与评测脚本无关。"
-                 "请检查 DevEco 登录状态、模型配置，稍后重试 `deveco run`。")
-        elif stats.tool_calls == 0 and not stats.errors:
-            warn("无工具调用且无错误事件 — 检查 stream_output.jsonl 是否为空。")
-
-    # Human-readable text
-    text_content = "\n".join(stats.text_lines)
-    text_path.write_text(text_content or "(no text output)", encoding="utf-8")
-
-    # Modified files
-    r2 = subprocess.run(
-        ["git", "diff", "--name-only"], capture_output=True, text=True, cwd=repo_root,
-        **subprocess_no_window_kwargs(),
-    )
-    mod_files = len([l for l in r2.stdout.strip().split("\n") if l])
-
-    hg_stream = analyze_homegraph_stream(stream_path)
-
-    results = {
-        "experiment_id": exp_id, "title": title, "agent": agent["name"],
-        "agent_binary": agent["binary"], "session_id": session_id,
-        "start_time_ms": start_ms, "end_time_ms": end_ms,
-        "duration_ms": dur_ms, "duration_s": dur_s,
-        "exit_code": exit_code, "max_turns": max_turns,
-        "skip_permissions": skip_permissions,
-        "tool_calls": stats.tool_calls, "tool_names": stats.tool_names,
-        "homegraph_tool_calls": stats.homegraph_tool_calls,
-        "used_homegraph": stats.used_homegraph,
-        **hg_stream,
-        "eval_arm": os.environ.get("EVAL_ARM", ""),
-        "files_read": list(set(stats.files_read)),
-        "files_edited": list(set(stats.files_edited)),
-        "model": resolve_agent_model(agent, stats),
-        "deveco_session_id": stats.deveco_session_id,
-        "input_tokens": stats.total_input_tokens,
-        "output_tokens": stats.total_output_tokens,
-        "assistant_messages": stats.assistant_messages,
-        "max_turns_hit": stats.max_turns_hit,
-        "errors": stats.errors[:5],
-        "modified_files": mod_files,
-        "memory": mem.to_dict(),
-    }
-    write_json(result_dir / "results.json", results)
-
-    if state_dir:
-        append_run_log(state_dir, f"Exp {exp_id} finished in {dur_s}s, tools={stats.tool_calls}, exit={exit_code}")
-    info(f"Results saved: {result_dir}/")
-    print(f"  Duration: {dur_s}s | Tool calls: {stats.tool_calls} | "
-          f"HG tools: {stats.homegraph_tool_calls} | "
-          f"Files read: {len(set(stats.files_read))} | Edited: {mod_files} | "
-          f"Peak mem: {mem.peak_combined_rss_mb:.0f} MB")
-    if stats.max_turns_hit:
-        print(f"  ⚠️  Max turns ({max_turns}) hit")
-    return exit_code == 0
-
-
-if __name__ == "__main__":
-    # Parse: python run_one.py <id> [--agent <name>] [--no-skip-permissions]
-    args = sys.argv[1:]
-    exp_id = args[0] if args else ""
-    agent_name = ""
-    skip = True
-
-    i = 1
-    while i < len(args):
-        if args[i] == "--agent" and i + 1 < len(args):
-            agent_name = args[i + 1]; i += 2
-        elif args[i] == "--no-skip-permissions":
-            skip = False; i += 1
-        else:
-            i += 1
-
-    if not exp_id:
-        error("Usage: python run_one.py <experiment_id> [--agent <name>]")
-        sys.exit(1)
-    sys.exit(0 if run_experiment(exp_id, skip, agent_name) else 1)
+#!/usr/bin/env python3
+"""Run a single one-shot experiment. Supports multiple agents via --agent."""
+
+import os
+import subprocess
+import sys
+from pathlib import Path
+from typing import Optional
+
+import setup as setup_mod
+from _utils import (OUTPUT_DIR, append_run_log, build_agent_cmd, current_time_ms,
+                    error, find_experiment, get_agent, info, header,
+                    parse_output, resolve_agent_model, run_monitored_subprocess,
+                    subprocess_no_window_kwargs, warn, write_json)
+from deveco_arm import analyze_homegraph_stream
+
+
+def run_experiment(exp_id: str, skip_permissions: bool = True,
+                   agent_name: str = "",
+                   results_dir: Optional[Path] = None,
+                   state_dir: Optional[Path] = None) -> bool:
+    exp = find_experiment(exp_id)
+    if exp is None:
+        error(f"Experiment {exp_id} not found in config")
+        return False
+    if exp.get("type") == "session":
+        error(f"Experiment {exp_id} is session type. Use run_session.py.")
+        return False
+
+    agent = get_agent(agent_name)
+    clone_path = os.environ.get("REPO_CLONE", "")
+    local_repo = os.environ.get("REPO_PATH", "")
+    if clone_path:
+        repo_root = Path(clone_path)
+    elif local_repo:
+        repo_root = Path(local_repo)
+    else:
+        error("No repo. Use run_all.py to clone, or set REPO_PATH / REPO_CLONE.")
+        return False
+
+    title = exp["title"]
+    prompt = exp["prompt"]
+    max_turns = exp.get("max_turns", 15)
+
+    header(f"Experiment {exp_id}: {title}")
+    info(f"Agent: {agent['name']} ({agent['binary']}) | Repo: {repo_root}")
+    info(f"Max turns: {max_turns} | Permissions: {'skip' if skip_permissions else 'ask'}")
+
+    session_id, _ = setup_mod.prepare_environment(exp_id, "full", agent_name,
+                                                    state_dir=state_dir)
+
+    result_dir = results_dir / exp_id if results_dir else OUTPUT_DIR / exp_id
+    result_dir.mkdir(parents=True, exist_ok=True)
+    (result_dir / "prompt.txt").write_text(prompt, encoding="utf-8")
+    write_json(result_dir / "experiment_config.json", {**exp, "agent": agent_name or "default"})
+
+    info("Starting agent...")
+    start_ms = current_time_ms()
+    os.chdir(repo_root)
+
+    cmd = build_agent_cmd(agent, session_id, max_turns, prompt,
+                          resume=False, skip_permissions=skip_permissions,
+                          repo_root=repo_root)
+
+    stream_path = result_dir / "stream_output.jsonl"
+    text_path = result_dir / "raw_output.txt"
+    track_hg = os.environ.get("EVAL_ARM", "") == "homegraph"
+
+    with open(stream_path, "w", encoding="utf-8") as stream_f:
+        result, mem = run_monitored_subprocess(
+            cmd, stdout=stream_f, stderr=subprocess.STDOUT, text=True,
+            track_homegraph=track_hg,
+        )
+    exit_code = result.returncode
+
+    end_ms = current_time_ms()
+    dur_ms = end_ms - start_ms
+    dur_s = f"{dur_ms / 1000:.1f}"
+    info(f"Completed in {dur_s}s (exit={exit_code})")
+
+    # Parse agent output
+    parser_name = agent.get("parser", "claude_stream_json")
+    stats = parse_output(stream_path, parser_name)
+
+    if exit_code != 0:
+        error(f"Agent 退出码 {exit_code}")
+        for err in stats.errors[:3]:
+            error(f"  {err}")
+        if "Unexpected server error" in " ".join(stats.errors):
+            warn("DevEco 服务端报错 — 通常为 API/网络/配额问题，与评测脚本无关。"
+                 "请检查 DevEco 登录状态、模型配置，稍后重试 `deveco run`。")
+        elif stats.tool_calls == 0 and not stats.errors:
+            warn("无工具调用且无错误事件 — 检查 stream_output.jsonl 是否为空。")
+
+    # Human-readable text
+    text_content = "\n".join(stats.text_lines)
+    text_path.write_text(text_content or "(no text output)", encoding="utf-8")
+
+    # Modified files
+    r2 = subprocess.run(
+        ["git", "diff", "--name-only"], capture_output=True, text=True, cwd=repo_root,
+        **subprocess_no_window_kwargs(),
+    )
+    mod_files = len([l for l in r2.stdout.strip().split("\n") if l])
+
+    hg_stream = analyze_homegraph_stream(stream_path)
+
+    results = {
+        "experiment_id": exp_id, "title": title, "agent": agent["name"],
+        "agent_binary": agent["binary"], "session_id": session_id,
+        "start_time_ms": start_ms, "end_time_ms": end_ms,
+        "duration_ms": dur_ms, "duration_s": dur_s,
+        "exit_code": exit_code, "max_turns": max_turns,
+        "skip_permissions": skip_permissions,
+        "tool_calls": stats.tool_calls, "tool_names": stats.tool_names,
+        "homegraph_tool_calls": stats.homegraph_tool_calls,
+        "used_homegraph": stats.used_homegraph,
+        **hg_stream,
+        "eval_arm": os.environ.get("EVAL_ARM", ""),
+        "files_read": list(set(stats.files_read)),
+        "files_edited": list(set(stats.files_edited)),
+        "model": resolve_agent_model(agent, stats),
+        "deveco_session_id": stats.deveco_session_id,
+        "input_tokens": stats.total_input_tokens,
+        "output_tokens": stats.total_output_tokens,
+        "assistant_messages": stats.assistant_messages,
+        "max_turns_hit": stats.max_turns_hit,
+        "errors": stats.errors[:5],
+        "modified_files": mod_files,
+        "memory": mem.to_dict(),
+    }
+    write_json(result_dir / "results.json", results)
+
+    if state_dir:
+        append_run_log(state_dir, f"Exp {exp_id} finished in {dur_s}s, tools={stats.tool_calls}, exit={exit_code}")
+    info(f"Results saved: {result_dir}/")
+    print(f"  Duration: {dur_s}s | Tool calls: {stats.tool_calls} | "
+          f"HG tools: {stats.homegraph_tool_calls} | "
+          f"Files read: {len(set(stats.files_read))} | Edited: {mod_files} | "
+          f"Peak mem: {mem.peak_combined_rss_mb:.0f} MB")
+    if stats.max_turns_hit:
+        print(f"  ⚠️  Max turns ({max_turns}) hit")
+    return exit_code == 0
+
+
+if __name__ == "__main__":
+    # Parse: python run_one.py <id> [--agent <name>] [--no-skip-permissions]
+    args = sys.argv[1:]
+    exp_id = args[0] if args else ""
+    agent_name = ""
+    skip = True
+
+    i = 1
+    while i < len(args):
+        if args[i] == "--agent" and i + 1 < len(args):
+            agent_name = args[i + 1]; i += 2
+        elif args[i] == "--no-skip-permissions":
+            skip = False; i += 1
+        else:
+            i += 1
+
+    if not exp_id:
+        error("Usage: python run_one.py <experiment_id> [--agent <name>]")
+        sys.exit(1)
+    sys.exit(0 if run_experiment(exp_id, skip, agent_name) else 1)

package/scripts/exp_boundary_eval/run_session.py

@@ -1,158 +1,158 @@
-#!/usr/bin/env python3
-"""Run multi-turn session experiment (5). Supports multiple agents via --agent."""
-
-import os
-import subprocess
-import sys
-import time
-from pathlib import Path
-from typing import Optional
-
-import setup as setup_mod
-from _utils import (OUTPUT_DIR, append_run_log, build_agent_cmd,
-                    current_time_ms, find_experiment, get_agent, info, header,
-                    log, parse_output, resolve_agent_model, run_monitored_subprocess,
-                    subprocess_no_window_kwargs, write_json, warn)
-
-
-def run_session(exp_id: str = "5", skip_permissions: bool = True,
-                agent_name: str = "",
-                results_dir: Optional[Path] = None,
-                state_dir: Optional[Path] = None) -> bool:
-    exp = find_experiment(exp_id)
-    if exp is None:
-        print(f"ERROR: Experiment {exp_id} not found")
-        return False
-
-    agent = get_agent(agent_name)
-    clone_path = os.environ.get("REPO_CLONE", "")
-    local_repo = os.environ.get("REPO_PATH", "")
-    if clone_path:
-        repo_root = Path(clone_path)
-    elif local_repo:
-        repo_root = Path(local_repo)
-    else:
-        print("ERROR: No repo. Use run_all.py to clone, or set REPO_PATH / REPO_CLONE.")
-        return False
-    max_turns = exp.get("max_turns_per_round", 25)
-
-    header("Experiment 5: Session Persistence & Context Decay")
-    info(f"Agent: {agent['name']} ({agent['binary']}) | Repo: {repo_root}")
-    info(f"Max turns/round: {max_turns} | Permissions: {'skip' if skip_permissions else 'ask'}")
-
-    session_id, _ = setup_mod.prepare_environment("5", "full", agent_name,
-                                                    state_dir=state_dir)
-    log(f"Shared Session ID: {session_id}")
-
-    result_dir = results_dir / "5" if results_dir else OUTPUT_DIR / "5"
-    result_dir.mkdir(parents=True, exist_ok=True)
-
-    rounds_data = []
-    session_model = ""
-    deveco_session_id = ""
-    for round_num in range(1, 6):
-        ri = next((r for r in exp["rounds"] if r["round"] == round_num), None)
-        if not ri: continue
-
-        header(f"Round {round_num} of 5")
-        info(f"Title: {ri['title']}")
-
-        round_dir = result_dir / f"round_{round_num}"
-        round_dir.mkdir(parents=True, exist_ok=True)
-        (round_dir / "prompt.txt").write_text(ri["prompt"], encoding="utf-8")
-
-        start_ms = current_time_ms()
-        is_resume = round_num > 1
-
-        cmd = build_agent_cmd(agent, session_id, max_turns, ri["prompt"],
-                              resume=is_resume, skip_permissions=skip_permissions,
-                              repo_root=repo_root)
-
-        stream_path = round_dir / "stream_output.jsonl"
-        text_path = round_dir / "raw_output.txt"
-        track_hg = os.environ.get("EVAL_ARM", "") == "homegraph"
-
-        with open(stream_path, "w", encoding="utf-8") as stream_f:
-            result, mem = run_monitored_subprocess(
-                cmd, stdout=stream_f, stderr=subprocess.STDOUT, text=True,
-                track_homegraph=track_hg,
-            )
-        exit_code = result.returncode
-
-        end_ms = current_time_ms()
-        dur_ms = end_ms - start_ms
-        dur_s = f"{dur_ms / 1000:.1f}"
-
-        parser_name = agent.get("parser", "claude_stream_json")
-        stats = parse_output(stream_path, parser_name)
-        if not session_model:
-            session_model = resolve_agent_model(agent, stats)
-        if not deveco_session_id and stats.deveco_session_id:
-            deveco_session_id = stats.deveco_session_id
-
-        text_content = "\n".join(stats.text_lines)
-        text_path.write_text(text_content or "(no text output)", encoding="utf-8")
-
-        r2 = subprocess.run(
-            ["git", "diff", "--name-only"], capture_output=True, text=True, cwd=repo_root,
-            **subprocess_no_window_kwargs(),
-        )
-        mod_files = len([l for l in r2.stdout.strip().split("\n") if l])
-
-        rr = {"round": round_num, "title": ri["title"], "duration_ms": dur_ms,
-              "duration_s": dur_s, "exit_code": exit_code,
-              "tool_calls": stats.tool_calls, "tool_names": stats.tool_names,
-              "files_read": list(set(stats.files_read)),
-              "files_edited": list(set(stats.files_edited)),
-              "assistant_messages": stats.assistant_messages,
-              "max_turns_hit": stats.max_turns_hit,
-              "modified_files": mod_files, "max_turns": max_turns,
-              "memory": mem.to_dict()}
-        write_json(round_dir / "round_results.json", rr)
-        rounds_data.append(rr)
-
-        info(f"Round {round_num} done: {dur_s}s | Tool calls: {stats.tool_calls} | "
-             f"Files read: {len(set(stats.files_read))} | Edited: {mod_files}")
-        if stats.max_turns_hit:
-            warn(f"  Max turns ({max_turns}) hit in round {round_num}")
-        if round_num < 5:
-            time.sleep(2)
-
-    summary = {
-        "experiment_id": "5", "session_id": session_id, "agent": agent["name"],
-        "model": session_model,
-        "deveco_session_id": deveco_session_id,
-        "total_rounds": len(rounds_data),
-        "total_duration_ms": sum(r["duration_ms"] for r in rounds_data),
-        "total_tool_calls": sum(int(r["tool_calls"]) for r in rounds_data),
-        "memory_retention_curve": [
-            {"round": r["round"], "tool_calls": int(r["tool_calls"]),
-             "files_read": r.get("files_read", [])} for r in rounds_data
-        ],
-    }
-    write_json(result_dir / "session_results.json", summary)
-
-    print("\nMemory Retention Curve:")
-    for r in rounds_data:
-        tc = int(r["tool_calls"])
-        fr = len(r.get("files_read", []))
-        print(f"  Round {r['round']}: {'#' * (tc // 2)} ({tc} tool calls, {fr} files read)")
-
-    log(f"Results: {result_dir}/session_results.json")
-    return True
-
-
-if __name__ == "__main__":
-    args = sys.argv[1:]
-    exp_id = args[0] if args else "5"
-    agent_name = ""
-    skip = True
-    i = 1
-    while i < len(args):
-        if args[i] == "--agent" and i + 1 < len(args):
-            agent_name = args[i + 1]; i += 2
-        elif args[i] == "--no-skip-permissions":
-            skip = False; i += 1
-        else:
-            i += 1
-    sys.exit(0 if run_session(exp_id, skip, agent_name) else 1)
+#!/usr/bin/env python3
+"""Run multi-turn session experiment (5). Supports multiple agents via --agent."""
+
+import os
+import subprocess
+import sys
+import time
+from pathlib import Path
+from typing import Optional
+
+import setup as setup_mod
+from _utils import (OUTPUT_DIR, append_run_log, build_agent_cmd,
+                    current_time_ms, find_experiment, get_agent, info, header,
+                    log, parse_output, resolve_agent_model, run_monitored_subprocess,
+                    subprocess_no_window_kwargs, write_json, warn)
+
+
+def run_session(exp_id: str = "5", skip_permissions: bool = True,
+                agent_name: str = "",
+                results_dir: Optional[Path] = None,
+                state_dir: Optional[Path] = None) -> bool:
+    exp = find_experiment(exp_id)
+    if exp is None:
+        print(f"ERROR: Experiment {exp_id} not found")
+        return False
+
+    agent = get_agent(agent_name)
+    clone_path = os.environ.get("REPO_CLONE", "")
+    local_repo = os.environ.get("REPO_PATH", "")
+    if clone_path:
+        repo_root = Path(clone_path)
+    elif local_repo:
+        repo_root = Path(local_repo)
+    else:
+        print("ERROR: No repo. Use run_all.py to clone, or set REPO_PATH / REPO_CLONE.")
+        return False
+    max_turns = exp.get("max_turns_per_round", 25)
+
+    header("Experiment 5: Session Persistence & Context Decay")
+    info(f"Agent: {agent['name']} ({agent['binary']}) | Repo: {repo_root}")
+    info(f"Max turns/round: {max_turns} | Permissions: {'skip' if skip_permissions else 'ask'}")
+
+    session_id, _ = setup_mod.prepare_environment("5", "full", agent_name,
+                                                    state_dir=state_dir)
+    log(f"Shared Session ID: {session_id}")
+
+    result_dir = results_dir / "5" if results_dir else OUTPUT_DIR / "5"
+    result_dir.mkdir(parents=True, exist_ok=True)
+
+    rounds_data = []
+    session_model = ""
+    deveco_session_id = ""
+    for round_num in range(1, 6):
+        ri = next((r for r in exp["rounds"] if r["round"] == round_num), None)
+        if not ri: continue
+
+        header(f"Round {round_num} of 5")
+        info(f"Title: {ri['title']}")
+
+        round_dir = result_dir / f"round_{round_num}"
+        round_dir.mkdir(parents=True, exist_ok=True)
+        (round_dir / "prompt.txt").write_text(ri["prompt"], encoding="utf-8")
+
+        start_ms = current_time_ms()
+        is_resume = round_num > 1
+
+        cmd = build_agent_cmd(agent, session_id, max_turns, ri["prompt"],
+                              resume=is_resume, skip_permissions=skip_permissions,
+                              repo_root=repo_root)
+
+        stream_path = round_dir / "stream_output.jsonl"
+        text_path = round_dir / "raw_output.txt"
+        track_hg = os.environ.get("EVAL_ARM", "") == "homegraph"
+
+        with open(stream_path, "w", encoding="utf-8") as stream_f:
+            result, mem = run_monitored_subprocess(
+                cmd, stdout=stream_f, stderr=subprocess.STDOUT, text=True,
+                track_homegraph=track_hg,
+            )
+        exit_code = result.returncode
+
+        end_ms = current_time_ms()
+        dur_ms = end_ms - start_ms
+        dur_s = f"{dur_ms / 1000:.1f}"
+
+        parser_name = agent.get("parser", "claude_stream_json")
+        stats = parse_output(stream_path, parser_name)
+        if not session_model:
+            session_model = resolve_agent_model(agent, stats)
+        if not deveco_session_id and stats.deveco_session_id:
+            deveco_session_id = stats.deveco_session_id
+
+        text_content = "\n".join(stats.text_lines)
+        text_path.write_text(text_content or "(no text output)", encoding="utf-8")
+
+        r2 = subprocess.run(
+            ["git", "diff", "--name-only"], capture_output=True, text=True, cwd=repo_root,
+            **subprocess_no_window_kwargs(),
+        )
+        mod_files = len([l for l in r2.stdout.strip().split("\n") if l])
+
+        rr = {"round": round_num, "title": ri["title"], "duration_ms": dur_ms,
+              "duration_s": dur_s, "exit_code": exit_code,
+              "tool_calls": stats.tool_calls, "tool_names": stats.tool_names,
+              "files_read": list(set(stats.files_read)),
+              "files_edited": list(set(stats.files_edited)),
+              "assistant_messages": stats.assistant_messages,
+              "max_turns_hit": stats.max_turns_hit,
+              "modified_files": mod_files, "max_turns": max_turns,
+              "memory": mem.to_dict()}
+        write_json(round_dir / "round_results.json", rr)
+        rounds_data.append(rr)
+
+        info(f"Round {round_num} done: {dur_s}s | Tool calls: {stats.tool_calls} | "
+             f"Files read: {len(set(stats.files_read))} | Edited: {mod_files}")
+        if stats.max_turns_hit:
+            warn(f"  Max turns ({max_turns}) hit in round {round_num}")
+        if round_num < 5:
+            time.sleep(2)
+
+    summary = {
+        "experiment_id": "5", "session_id": session_id, "agent": agent["name"],
+        "model": session_model,
+        "deveco_session_id": deveco_session_id,
+        "total_rounds": len(rounds_data),
+        "total_duration_ms": sum(r["duration_ms"] for r in rounds_data),
+        "total_tool_calls": sum(int(r["tool_calls"]) for r in rounds_data),
+        "memory_retention_curve": [
+            {"round": r["round"], "tool_calls": int(r["tool_calls"]),
+             "files_read": r.get("files_read", [])} for r in rounds_data
+        ],
+    }
+    write_json(result_dir / "session_results.json", summary)
+
+    print("\nMemory Retention Curve:")
+    for r in rounds_data:
+        tc = int(r["tool_calls"])
+        fr = len(r.get("files_read", []))
+        print(f"  Round {r['round']}: {'#' * (tc // 2)} ({tc} tool calls, {fr} files read)")
+
+    log(f"Results: {result_dir}/session_results.json")
+    return True
+
+
+if __name__ == "__main__":
+    args = sys.argv[1:]
+    exp_id = args[0] if args else "5"
+    agent_name = ""
+    skip = True
+    i = 1
+    while i < len(args):
+        if args[i] == "--agent" and i + 1 < len(args):
+            agent_name = args[i + 1]; i += 2
+        elif args[i] == "--no-skip-permissions":
+            skip = False; i += 1
+        else:
+            i += 1
+    sys.exit(0 if run_session(exp_id, skip, agent_name) else 1)