🔬 Research Tab¶

The Research tab is the 5th tab in the RLM Code TUI, accessible via Ctrl+5 / F6. It provides a dedicated space for experiment tracking, trajectory viewing, benchmarks, session replay, and live event streaming, all wired to real data from the RLM runner.

📍 How to Access¶

🗂️ Sub-Tabs¶

The Research tab organizes data across 5 sub-tabs, each shown as a button bar at the top of the pane. Click a sub-tab to switch views.

┌─────────────────────────────────────────────────────┐
│ 🔬 Research                                         │
│ [Dashboard] [Trajectory] [Benchmarks] [Replay] [Events] │
│                                                     │
│  ┌─ Content area (changes per sub-tab) ──────────┐  │
│  │                                                │  │
│  │                                                │  │
│  └────────────────────────────────────────────────┘  │
└─────────────────────────────────────────────────────┘

📊 Dashboard¶

The default sub-tab. Shows a high-level summary of the most recent RLM run.

Widgets¶

How It Populates¶

1. Run /rlm run "your task" or /rlm bench preset=dspy_quick in the RLM tab
2. When the run completes, the Dashboard auto-populates via build_run_visualization()
3. The MetricsPanel updates its reactive properties (run_id, status, reward, steps, etc.)
4. The SparklineChart fills with cumulative reward values from the trajectory

Data Source¶

from rlm_code.rlm.visualizer import build_run_visualization

viz = build_run_visualization(run_path=run_path, run_dir=run_path.parent)
# viz["run_id"], viz["status"], viz["total_reward"],
# viz["step_count"], viz["reward_curve"], viz["timeline"]

📈 Trajectory¶

Step-by-step timeline of the RLM run, showing what the agent did at each step.

Table Columns¶

How It Populates¶

After any /rlm run or /rlm bench command, the trajectory is extracted from build_run_visualization()["timeline"] and rendered as a Rich table.

🏆 Benchmarks¶

Displays the leaderboard table from benchmark runs.

What You See¶

A Rich table ranked by reward, showing:

How It Populates¶

Run /rlm bench preset=<name> then switch to Research → Benchmarks. The data comes from:

from rlm_code.rlm.leaderboard import Leaderboard

lb = Leaderboard(workdir=Path.cwd() / ".rlm_code", auto_load=True)
table = lb.format_rich_table(limit=15)

⏪ Replay¶

Step-through controls for time-travel debugging of any RLM run.

Controls¶

What You See¶

• Step position: Step 3/8 indicator
• Step detail: Action code with syntax highlighting, output, reward, cumulative reward
• Reward curve: SparklineChart showing the full reward trajectory with current position

How It Populates¶

Run /rlm status to get a run id, then /rlm replay <run_id>. The TUI automatically switches to Research → Replay and loads the session:

from rlm_code.rlm.session_replay import SessionReplayer

replayer = SessionReplayer.from_jsonl(run_path)
replayer.step_forward()    # advance one step
replayer.step_backward()   # go back one step
replayer.goto_step(n)      # jump to step n

Reward Color Coding¶

📡 Events¶

Live event stream from the RLM event bus, showing real-time progress during active runs.

What You See¶

A RichLog widget that streams formatted events with timestamps:

[14:23:01] 🟢 RUN_START - Starting run abc123 (pure_rlm)
[14:23:02] 🔵 ITERATION_START - Step 1/8
[14:23:04] 🟡 LLM_CALL - Calling claude-sonnet-4-20250514 (450 tokens)
[14:23:06] 🟢 ITERATION_END - Step 1 complete (reward: +0.15)
[14:23:08] 🟢 RUN_END - Run complete (total reward: 0.72)

Event Types¶

The event bus supports 27+ event types including:

Thread Safety¶

Events flow from the RLM runner (which runs in a worker thread) to the UI via call_from_thread() for thread-safe rendering:

def _on_raw_rlm_event(self, event):
    self.call_from_thread(self._on_rlm_event, event)

🔗 Integration with Slash Commands¶

The Research tab auto-updates when you run RLM commands in the RLM tab:

🎨 Visual Design¶

The Research tab inherits the TUI's purple-accented dark theme with additional styling for research-specific elements:

📊 Widgets Used¶