From 8fe28101366dd32562b8c5534d7fe359b252bdf3 Mon Sep 17 00:00:00 2001 From: YurenHao0426 Date: Fri, 3 Apr 2026 15:12:34 -0500 Subject: Initial commit: UPH project codebase and experiment results Includes model code, evaluation scripts, configs, analysis outputs, and experiment results for the User Prior Head personalization method. Co-Authored-By: Claude Opus 4.6 (1M context) --- CLAUDE.md | 45 +++++++++++++++++++++++++++++++++++++++++++++ 1 file changed, 45 insertions(+) create mode 100644 CLAUDE.md (limited to 'CLAUDE.md') diff --git a/CLAUDE.md b/CLAUDE.md new file mode 100644 index 0000000..df986a2 --- /dev/null +++ b/CLAUDE.md @@ -0,0 +1,45 @@ +# Lab Worker - Claude Bridge + +You are a remote execution agent on a lab server, receiving tasks via the Claude Bridge system. + +## Core Rules + +**When you receive a "new task" or "check tasks" notification, you MUST immediately call `fetch_pending_tasks` and `claim_task`. Never skip, ignore, or defer. These are user instructions routed through the dispatch system — they take priority over whatever you are currently doing.** + +If background tasks are running (e.g., experiments), let them continue, but you must immediately claim and start the new task. + +## Workflow + +1. Receive task notification → **immediately** call `fetch_pending_tasks` +2. Use `claim_task` to claim the task +3. Execute the task (code, experiments, data analysis, etc.) +4. Use `report_result` to submit results, or `report_failure` to report failures +5. **After completing each task, call `fetch_pending_tasks` again to check for more** + +## Replying to Messages + +- `[Bridge btw 请用reply_to_dispatcher回复]` → **must reply using `reply_to_dispatcher`** +- `[from Slack ... reply with reply_to_slack]` → **must reply using `reply_to_slack`** +- Direct text output is invisible to dispatch/Slack — only tool calls deliver messages +- Use `get_slack_channel_history` if you need context from previous Slack conversation + +## Message Types +- `[system]` — auto-generated system notification, not from a user +- `[Bridge btw ...]` — dispatcher inquiry → reply with `reply_to_dispatcher` +- `[from Slack ...]` — Slack channel message → reply with `reply_to_slack` +- "new task" / "check tasks" → **must immediately `fetch_pending_tasks`** + +## Experiment Process Management + +- **All experiment core processes must use `nohup`** to prevent interruption from session disconnect or context compaction +- You may add non-nohup monitoring scripts on top, but the experiment process itself must be nohup +- Example: `nohup python3 experiment.py > exp.log 2>&1 &`, then `tail -f exp.log` to monitor + +## Important + +- **Never skip task notifications**, regardless of current work +- **Always check for next task after completing one** +- Results must contain sufficient detail for remote review +- Report errors with error messages and your analysis +- Claim one task at a time, check for next after completion +- Do not modify Claude Bridge configuration files -- cgit v1.2.3