Repeating prompts based on contextual-failure/success

[mcgdb]

I'm trying to set up a workflow wherein I'm trying to track down a memory-leak/overflow issue.

It's on an embedded system, and I have tools to 'build', 'flash' and 'monitor to a file' and 'run a tools that will trigger the bug'.

I am trying to get workflows to support this (any one step can fail, and would require fixing that step before trying to continue...if you can't build it, then you can't flash/monitor/trigger-the-bug).

I'm unclear on how to do this based on the examples, and I think it's because of my misunderstanding of how the repeat_until flag is treated .

If the command returns '0', like build and flash do, then it 'seems' like following what's in the examples works just fine too...However, when there is a crash, the monitor process doesn't have an error...it's outputting that there was a crash to the logs, and that's what it's supposed to do, so evaulating the return value of the logging proceess that would get called with cmd_runner doesn't work...The LLM would need to read the logs and see that there's a stack-exception of mediation error or similar...

How can I get that type of evaluation ?

In a chat, if I ask it to review the logs, the LLM is almost always able to discern that things 'didn't work', but I'm not sure how to get this 'soft - failure' fed back into the workflow mechanisms.

Thanks!
Could someone provide a snippet or skeleton or just a comment of what I would need to adjust to make this conceptual workflow work?

Thanks again for this great tool...It's crazy that we're bossing around the thing that will likely be the death of us all! But that's a problem for next year...This is due this week! :)

I don't have much of a config to share, since I've been trying lot's of different things, but certainly here's something I'm running now that isn't working, for completeness.:

            {
              {
                role = "system",
                content = SYSTEM_PROMPT,
                persistent = true,
                opts = {
                  visible = false,
                },
              },
              {
                name = "Build",
                role = "user",
                opts = { auto_submit = true, visible = true },
                content = function(chat)
                  -- Enable turbo mode!!!
                  vim.g.codecompanion_auto_tool_mode = true

                  return [[### Instructions
Use the @{full_stack_dev} and @{vectorcode_toolbox} tools to do the following:
read the contents of `build.log`...if there are no errors and the build completed successfully, then continue.  If there are errors, review them, patch them with the @{full_stack_dev} tools, and then use @{cmd_runner} to run `getIDF && idf.py build`.
]]
                end,
              },
            },
            {
              {
                role = "system",
                content = SYSTEM_PROMPT,
                persistent = true,
                opts = {
                  visible = false,
                },
              },
              {
                name = "Run",
                role = "user",
                opts = { auto_submit = true, visible = true },
                content = function(chat)
                  -- Enable turbo mode!!!
                  vim.g.codecompanion_auto_tool_mode = true

                  return [[### Instructions
Use the @{full_stack_dev} and @{vectorcode_toolbox} tools to do the following:
read the contents of `esp.log`...if there are no errors and the build completed successfully, then we are all done!  If there are errors, review them, patch them with the @{full_stack_dev} tools, and then run `getIDF && idf.py build flash`.
]]
                end,
              },
            },
          },
        },