Inspecting trajectories
Inspecting trajectories
- Trajectories are the main output of SWE-agent. They are the best way to understand what SWE-agent does, especially when running on many problem instances.
- We provide two tools for visualizing the
.traj
files from thetrajectories
folder more easily. - Use
swe-agent inspect
(orsweagent i
) to open the command line inspector. - Use
swe-agent inspector
(orsweagent I
) to open the web inspector. - Please complete the hello world tutorial before proceeding.
You should see a folder called trajectories
in your working directory. Let's go to one of the experiment directories:
cd trajectories/$USER/<some directory> # (1)!
- Don't have a folder here? Make sure to run SWE-agent at least once.
Command line inspector
Run the inspector in the directory containing your .traj
files:
sweagent inspect
# or
sweagent i
You will be put into a pager that lets you navigate between trajectories. Here's how to navigate (this is similar to vim keybindings):
- Use
q
to quit - Switching between trajectories:
H
andL
go to the previous/next trajectoryt
brings up a list of all trajectories. Use type-ahead search to find a specific trajectory (once your search string results in a single match, the trajectory will be opened). Press<TAB>
to cycle through the list of matches.
- Use
h
andl
to navigate between the steps in the trjaectory - Use
j
andk
to scroll down/up - By default we only show reduced information. You can press
v
to toggle the view. - Press
o
to open the logs - Sometimes you see that you can press
e
to open a file in your$EDITOR
. For this to work, theEDITOR
environment variable must be set (e.g., tonano
orvim
).
Web-based inspector
Run the inspector in this directory (this is where your *.traj
files are):
sweagent inspector
# or
sweagent I
Additional flags
--directory
: Directory of trajectories to inspect (Defaults to current directory)--port
: Port to host web app (Defaults to8000
).
Benchmark results
If you are running SWE-agent on a benchmark (see batch mode), you will see evaluation results as ✅ or ❌. Otherwise, you will see ❓.
Tip
- If you do not see evaluation results, make sure that the SWE-bench output
is called
results.json
and is in the same directory as the trajectories. - To see gold patches, point
--data_path
to the SWE-bench dataset.
-
Something broken? Report bug
-
Something unclear? Ask question