The coder LLM claimed 'build successful' without actually running bild. We added post-coder verification which catches this, but ideally the coder should:
1. Actually run bild as instructed in the prompt 2. Fix any compilation errors before finishing 3. Not declare success until verified
The prompt says 'You MUST run bild' but the LLM ignores it. Consider:
After review: The post-coder verification approach is more reliable than trying to get the LLM to reliably run bild. The current implementation already verifies compilation after the coder finishes and fails the phase if it doesn't compile. Prompt engineering to make LLMs follow instructions is unreliable - external verification is the right pattern.