Pragma Tests Tool-Calling Reliability Floor

Strong systems post. The useful insight here is that orchestration is a different problem from code generation, and the failure mode is not “can it think?” but “can it stay inside the tool contract?”

• –The post is grounded in a practical local stack: llama.cpp, open-source models, and a visible reasoning loop.
• –The core claim is credible and specific: smaller models often fail on argument discipline before they fail on reasoning.
• –The proposed mitigations are directionally right, especially exact signatures in-prompt and tighter loop controls.
• –The repo angle makes this more than a rant; it reads like an early design note for a local agent harness.
• –Best follow-up for the ecosystem would be stricter schemas/grammar-constrained decoding and evaluation by failure class, not just overall task success.