[ihavespoons]

Summary

Phase 1 of the inference optimizer. The design is in docs/design/optimizer.md. The optimizer searches llama.cpp settings by measurement on the host's own hardware. An optional upstream LLM can propose candidates, but only a measurement decides, and the owner applies the result.

This PR adds the plumbing. The search, the API and the UI come in later phases.

Changes

• Model settings batch, ubatch and split_mode (internal/models), with --batch-size, --ubatch-size and --split-mode in the llama-swap command. They are in the shared form, the TUI and the console. -b, --batch-size, -ub and --ubatch-size are now typed flags, and LoadSettings moves old raw args into the typed fields.
• The host agent's tune.<model> request (os/rootfs/usr/libexec/sibyl/agent). It runs one tool in a backend image with the same container options as the benchmark: llama-bench, llama-batched-bench, or llama-server until /health answers. Every argument must match a pattern for its flag. The model path, the pool, the GPUs and the backend get fixed checks.
• hostctl.Tune.
• llama-batched-bench in the hip, vulkan and cuda images.
• The agent's paths can be set by SIBYL_AGENT_* variables, for the tests. The systemd unit sets none.

Verification

• TestAgentTuneAllowlist runs the agent's allowlist on 7 good and 22 bad argument lists, such as -m /etc/shadow, -rpc, --host 0.0.0.0, a flag as a value, and an -ot pattern that starts with a dash.
• TestAgentTuneRuns runs the whole agent with a fake podman for each tool and for refused arguments. It fails when --cap-drop all is removed from the tune container options.
• Tests for the new settings, the raw-arg move, and the generator.
• go vet, go test -race ./..., golangci-lint (0 issues) and shellcheck pass.
• Not run on the box, because it runs a training job.

🤖 Generated with Claude Code