Page refresh triggers destructive simulation restart — UI should not auto-restart on load

[jbraseth]

Problem

Refreshing the browser during a running (or paused/failed) simulation triggers a full simulation restart. The frontend appears to detect an incomplete state and automatically fires POST /api/simulation/start, which:

1. Clears all existing action logs (actions.jsonl)
2. Deletes the SQLite simulation databases (twitter_simulation.db, reddit_simulation.db)
3. Launches a new simulation process from round 0

This is destructive and unexpected. A page refresh (GET) should never trigger side effects that destroy data.

How I hit this

1. Ran a 40-round simulation. Reddit completed 40/40, Twitter reached 34/40 before Zep free tier quota was exhausted.
2. Stopped the simulation via POST /api/simulation/stop (which correctly sets status to PAUSED).
3. Refreshed the browser to check state.
4. Frontend detected the simulation was not complete and automatically restarted it from scratch — wiping 815 agent actions from the first run.

Additional context

• The Zep graph memory was being updated during the first run, so the knowledge graph now contains residual data from a simulation whose action logs no longer exist. The second run writes to the same graph, creating contaminated state.
• The stop endpoint sets SimulationStatus.PAUSED but there is no corresponding resume endpoint (line 203 of simulation_runner.py mentions resume support in the docstring, but no implementation exists).
• The run_parallel_simulation.py script deletes the existing database on startup (lines 1152-1153), making resume impossible at the OASIS level.

Suggested improvements

1. Never auto-restart on page load. The frontend should display the current state (paused, failed, completed) and require an explicit user action to restart.
2. Implement actual pause/resume. The database and action logs persist state. OASIS databases use SQLite. Resume could skip os.remove(db_path), use CREATE TABLE IF NOT EXISTS, and start the simulation loop from the last completed round.
3. Confirm before destructive restart. If the user explicitly clicks restart, show a confirmation dialog warning that existing data will be lost.
4. Separate graph memory from simulation lifecycle. Clearing action logs while leaving Zep graph state intact creates inconsistency.

Love the project — this came up during a real research simulation and the data loss was painful. Happy to discuss further.

---

Filed from the CLI via gh

[jbraseth]

中文翻译 / Chinese Translation

问题

在运行中（或暂停/失败的）模拟过程中刷新浏览器会触发完全重新启动。前端似乎检测到未完成的状态，并自动发送 POST /api/simulation/start，这会：

1. 清除所有现有的操作日志（actions.jsonl）
2. 删除 SQLite 模拟数据库（twitter_simulation.db、reddit_simulation.db）
3. 从第 0 轮开始启动新的模拟进程

这是破坏性的且出乎意料的。页面刷新（GET）不应该触发销毁数据的副作用。

复现步骤

1. 运行了一个 40 轮的模拟。Reddit 完成了 40/40，Twitter 在 Zep 免费层配额用尽之前达到了 34/40。
2. 通过 POST /api/simulation/stop 停止了模拟（正确地将状态设置为 PAUSED）。
3. 刷新浏览器以检查状态。
4. 前端检测到模拟未完成，自动从头重新启动——清除了第一次运行的 815 个智能体操作记录。

补充说明

• 第一次运行期间 Zep 图谱记忆正在被更新，因此知识图谱现在包含来自操作日志已不存在的模拟的残留数据。第二次运行写入同一图谱，造成数据污染。
• stop 端点设置了 SimulationStatus.PAUSED，但没有对应的 resume 端点（simulation_runner.py 第 203 行的文档字符串提到了恢复支持，但没有实现）。
• run_parallel_simulation.py 脚本在启动时删除现有数据库（第 1152-1153 行），使得在 OASIS 层面无法恢复。

改进建议

1. 页面加载时绝不自动重启。 前端应显示当前状态（暂停、失败、已完成），并要求用户明确操作才能重启。
2. 实现真正的暂停/恢复功能。 数据库和操作日志保留状态。OASIS 数据库使用 SQLite。恢复可以跳过 os.remove(db_path)，使用 CREATE TABLE IF NOT EXISTS，并从最后完成的轮次开始模拟循环。
3. 破坏性重启前确认。 如果用户明确点击重启，应显示确认对话框，警告现有数据将丢失。
4. 将图谱记忆与模拟生命周期分离。 清除操作日志而保留 Zep 图谱状态会造成不一致。

喜欢这个项目——这个问题是在真实的研究模拟中遇到的，数据丢失非常痛苦。欢迎进一步讨论。