agent loop repeats

[sick vector]

Still trying to figure this issue out, it was happening earlier as well and I tried changing the prompt etc but it didn't seem to make a difference

Here's a trace https://v3.dagger.cloud/dagger/traces/4e1cacdec81ef03c57e39b00373a1e68

The agent completes the job halfway through the trace but when asked to summarize, seems to start at the beginning and repeat all the steps again

[ruby dew]

Since the llm still has the full context, it seems like it's still taking that initial prompt as it's assignment. The last prompt ("summarize your changes") doesn't override that in any way. It's probably easiest to get the diff out of after and pass it to a new llm I think

[sick vector]

Ok thanks, I'm trying that

[sick vector]

👋 trying what you suggested here but not sure if I'm doing this correctly - can you help me debug why the agent is not returning a standard diff at the end of the first run? https://github.com/vikram-dagger/fastapi-sample-app/pull/52 and trace at https://v3.dagger.cloud/dagger/traces/3bf02fbffb8ccad9a49c7cf305308b7c

i tried making the prompt more explicit but no change, this is the next iteration which has the same problem: https://v3.dagger.cloud/dagger/traces/267722cbaeed354005bf4ccd67625e86

[ruby dew]

yeah so you're asking the agent to run the diff tool and return the output:

"""
...
- Once all the tests pass, you are done. Run the diff tool and return its output.
        """
        diff = await (
            dag.llm()
            .with_workspace(before)
            .with_prompt(prompt)
            .last_reply()
        )

but really what you want is probably:

# (delete the last line of the prompt)
work = (
            dag.llm()
            .with_workspace(before)
            .with_prompt(prompt)
        )
diff = await work.workspace().diff()

just use the diff tool 🙂

[sick vector]

yes, makes sense - thanks! trying now

[sick vector]

this works consistently now, ty!