freediag warm-up, week 1: tuning the pipeline on a smaller corpse
Week one of the freediag warm-up run: what the agent pipeline actually did against a real dormant OBD-II codebase, and what it got wrong.
What is Freediag?
Freediag is a free tool for grabbing diagnostic codes from cars over OBDII connections. Before the days of cheap diagnostic readers, it was a reliable option for many models of cars.
Now, it has largely been superseded by cheaper all-in-one readers, and other more modern software. But as a older C code base, it can provide a indicator of what Claude Code agent swarms are capable of when porting abandoned or semi-abandoned code to new languages or environments.
In the course of working on this codebase, several tweaks were made to my existing practices as gotchas were discovered.
Gotchas
Fable has proven to be a powerful PM ( Program Manager ), and in general because of Anthropic's alignment training, it tends to take a compliant but cautious bent when following user prompts:
- Do the smallest thing possible to satisfy the user's request. ( From the system prompt )
The one gotcha is spawning agents. If you set the model to Fable in Claude Code, and you are setting it up to act as a PM, then you must realize that satisfying #1, as a PM with a lot of open tickets, will result in Fable spawning multiple copies of itself to handle every ticket in parallel. The "smallest thing possible" does not apply to agent counts apparently.
And you can imagine how expensive this mistake can be. Fable is very capable, and very expensive. It excels as a director of work, and for tackling the hardest problems.
Of course this workflow issue was quickly noticed and corrected in the following ways:
And then ensuring these changes are saved in the claude file. These can all massively reduce costs and let Fable shine where it is strongest.
Findings
Additionally I instructed Claude Code to capture any findings discovered during the C porting and vetting process. Several issues were found, including silent corruption and truncation. A later blog post will go over those findings and analysis.