0.4.13

Added

• Added tiles run flags for llama.cpp tuning: --context-length, --gpu-layers, --offload-kqv, and --batch-size #160.
• Added a daemon /config endpoint so the Python inference backend can read Rust-owned Tiles config #160.

Changed

• Persist llama.cpp settings under [llama] in config.toml instead of using TILES_LLAMA_CPP_* environment variables #160.
• Reload the Linux llama.cpp runner when llama configuration changes, even if the selected model path stays the same #160.
• Reworked link command UX around tiles link create, tiles link add, tiles link list-peers, and tiles link revoke, with support for both offline link codes and UCAN tokens #163.
• Renamed inference system controls from tiles inference to tiles server with start, stop, and daemon subcommands #163.
• Switched Harmony handling to tiles-harmony across active server manifests #162.

Fixed

• Improved gpt-oss tool call handling by normalizing malformed tool names, passing tool metadata into Harmony conversation replay, and detecting tool calls emitted through the analysis channel #162.
• Fixed tool-call streaming state handling so final answers and function-call arguments are emitted more reliably #162.
• Improved dev Modelfile handling so cargo run -- run modelfiles/gpt-oss-gguf no longer depends on a pre-existing copied default Modelfile #160.