Commit 8385977

@@ -3,20 +3,40 @@
 Elelem::Providers.register(:gguf) do
   gpu = %w[vulkan cuda metal].include?(Elelem::Net::GGUF.backend)
 
+  # No baked-in model path: a fresh user has no reason to have the same file
+  # this project's author does. With exactly one .gguf in ~/.agents/models,
+  # that's unambiguously "the model"; with zero or several, only the user
+  # can say which -- so ask, rather than silently guessing.
+  default_model = -> {
+    dir = File.expand_path("~/.agents/models")
+    link = File.join(dir, "default.gguf")
+    return link if File.symlink?(link) || File.exist?(link)
+
+    models = Dir[File.join(dir, "*.gguf")].sort
+    case models.length
+    when 1 then models.first
+    when 0 then raise "no .gguf model found in #{dir}. Set ELELEM_GGUF_MODEL=/path/to/model.gguf"
+    else raise "multiple .gguf models found in #{dir}. Set ELELEM_GGUF_MODEL=/path/to/model.gguf " \
+               "or symlink #{link} to the one to use by default"
+    end
+  }
+
   Elelem::Net::GGUF.new(
-    model: ENV.fetch("GGUF_MODEL", File.expand_path("~/.agents/models/Qwen3.8-27B-UD-Q4_K_XL.gguf")),
+    model: ENV.fetch("ELELEM_GGUF_MODEL") { default_model.call },
     # n_ctx/n_threads/n_gpu_layers/max_tokens are calculated, not configured:
-    # the native shim asks llama.cpp for the model's trained max context
-    # (falling back to smaller sizes if it doesn't fit) and picks thread count
-    # from the host's physical cores; every layer is offloaded whenever a GPU
-    # backend is present, and a reply may use whatever context the prompt
-    # leaves free. Convention over configuration -- there's no better answer a
-    # user could supply than what the hardware and model already determine.
+    # the native shim asks llama.cpp for the model's trained max context, capped
+    # at a size that leaves GPU headroom for a large prompt's compute buffer
+    # (see EL_MAX_AUTO_CTX in elelem_llama.cpp), falling back to smaller sizes
+    # still if that doesn't fit; thread count comes from the host's physical
+    # cores; every layer is offloaded whenever a GPU backend is present; a
+    # reply may use whatever context the prompt leaves free. Convention over
+    # configuration -- there's no better *number* a user could supply than
+    # what the hardware and model already determine.
     n_ctx: 0,
     n_threads: 0,
     n_gpu_layers: gpu ? 999 : 0,
-    max_tokens: Integer(ENV.fetch("GGUF_MAX_TOKENS", "0")),
-    temp: Float(ENV.fetch("GGUF_TEMP", "0.7")),
-    seed: Integer(ENV.fetch("GGUF_SEED", "-1"))
+    max_tokens: Integer(ENV.fetch("ELELEM_GGUF_MAX_TOKENS", "0")),
+    temp: Float(ENV.fetch("ELELEM_GGUF_TEMP", "0.7")),
+    seed: Integer(ENV.fetch("ELELEM_GGUF_SEED", "-1"))
   )
 end