3 min read 631 words Updated Mar 24, 2026 Created Mar 24, 2026

Session Replay

Replay lets you step through past agent sessions like a debugger — see every tool call, every thought, every result. Useful for understanding what happened, debugging unexpected changes, and learning from agent behavior.

How It Works

Pick a session — :DwightReplay opens a Telescope picker of all past sessions.
Navigate — step forward and backward through events with j/k.
Jump — skip between tool calls with J/K to focus on actions, not reasoning.
Toggle view — switch between single-event and cumulative view with v.

:DwightReplay                " Telescope picker of all sessions
:DwightReplay latest         " Jump to the most recent session
:DwightReplay 20250314       " Partial filename match

The picker shows each session's timestamp, event count, and task name. The Telescope preview displays the full event timeline.

Key	Action
`j` / `l` / `Space`	Next step
`k` / `h`	Previous step
`J`	Jump to next tool call
`K`	Jump to previous tool call
`gg`	First step
`G`	Last step
`v`	Toggle single/cumulative view
`q` / `Esc`	Close

Event Types

Each event is rendered with an icon and contextual details:

Icon	Event	Shows
Launch	Session start	Backend, task prompt
Tool	Bash command	`$ command`
Read	File read	File path
Write	File write	File path
Edit	File edit	File path
Search	Search	Pattern
Result	Tool result	Output text
Error	Tool error	Error message
Thought	Agent thought	Reasoning text
Usage	Token usage	Input/output counts
Cost	Final cost	Total tokens + duration
Status	Finish	Exit code, duration, output size

Events show time offsets from session start (+12s, +45s).

View Modes

Single view (default): Shows only the current event with a marker and running stats (tools used, thoughts, errors so far).

Cumulative view (toggle with v): Shows all events from session start up to the current position. Useful for seeing the full context leading to a tool call.

Both views include a progress bar: Step 5 / 32 [===========-------]

Data Source

Replay reads JSONL log files from .dwight/agentic-logs/. These are written automatically during every agent session (:DwightAgent, :DwightAuto, etc.). Each line is a JSON object with a type and timestamp field.

No configuration needed — if you've run any agent tasks, the logs exist.

Tips

Replay after every Auto session. Even if the result looks correct, stepping through the session reveals the agent's decision-making process and potential issues.
Use cumulative view for debugging. When something went wrong, cumulative view shows the full context up to the failure point.
Jump between tool calls with J/K. Agent thoughts are informative but verbose — jumping between actions gives you the executive summary.
Share replays for team debugging. The JSONL files in .dwight/agentic-logs/ are portable. Share a log file and the teammate can replay it locally.

Commands

Command	Args	Description
`:DwightReplay`	`[session]`	Step through a session. Args: `latest`, filename, or blank for picker
`:DwightSessionLog`		View the persistent session log
`:DwightLog`		View the job log (all operations)