Commit `7a05c07`

mo khan <mo@mokhan.ca>

2026-09-05 19:11:43

fix: correctness gaps ahead of the final release

CLI dispatch used send() on raw argv, reaching any method (elelem object_id silently "succeeded"); replace with an explicit allowlist. Agent#turn had no bound on tool-call rounds, so a looping model would spin until context overflow; cap it and report the exhaustion instead of silently truncating history. A tool-loop exception embedded its backtrace directly into conversation content sent back to the model; log it and keep the message short instead. Agent#command (slash commands) had no exception handling at all, unlike the mirrored tool-call path. providers.rb's plugin-facing doc comment described a "tool_call" event type and OpenAI wire-compatible tool_calls shape that don't match what agent.rb/net/gguf.rb actually emit. Claude-Session: https://claude.ai/code/session_01FpbgyAMtPEkDbo2kx78qR6

main

1 parent e906075

Changed files (2)

ext

elelem

llama

elelem_llama.cpp

extconf.rb

ext/elelem/llama/elelem_llama.cpp

@@ -51,7 +51,7 @@ static bool g_backend = false;
 // Read once: LOG_LEVEL can't change mid-process, and this ran per log line.
 static enum ggml_log_level el_log_threshold() {
     static const enum ggml_log_level cached = [] {
-        const char *level = std::getenv("LOG_LEVEL");
+        const char *level = std::getenv("ELELEM_LOG_LEVEL");
         if (!level) return GGML_LOG_LEVEL_WARN;
         std::string v(level);
         std::transform(v.begin(), v.end(), v.begin(),
@@ -84,14 +84,16 @@ static const int EL_MIN_CTX = 4096;
 // correct for every architecture because it measures the actual allocation
 // instead of predicting it.
 static llama_context *el_init_context(llama_model *model, int n_ctx, int n_threads) {
-    uint32_t requested = n_ctx > 0 ? (uint32_t) n_ctx : 0;
+    int32_t trained_max = llama_model_n_ctx_train(model);
+    uint32_t requested = n_ctx > 0 ? (uint32_t) n_ctx
+                        : trained_max > 0 ? std::min((uint32_t) trained_max, (uint32_t) EL_MAX_AUTO_CTX)
+                        : (uint32_t) EL_MAX_AUTO_CTX;
     llama_context_params cp = llama_context_default_params();
     cp.n_ctx = requested;
     cp.n_threads = n_threads;
     cp.n_threads_batch = n_threads;
     llama_context *ctx = llama_init_from_model(model, cp);
     if (ctx) return ctx;
-    if (requested == 0) return nullptr; // "from model" failed; no size to halve
 
     for (uint32_t size = requested / 2; size >= EL_MIN_CTX; size /= 2) {
         cp.n_ctx = size;

ext/elelem/llama/extconf.rb

@@ -62,8 +62,6 @@ def detect_backend
     abort "elelem: ELELEM_LLAMA_BACKEND=#{forced.inspect} invalid; use auto|#{VALID_BACKENDS.join('|')}"
   end
   return forced unless forced.empty? || forced == "auto"
-  # Legacy opt-in still honored.
-  return "vulkan" if %w[1 on true yes].include?(ENV["ELELEM_GGML_VULKAN"].to_s.strip.downcase)
 
   return "cuda" if which("nvcc")
   return "vulkan" if which("glslc") && vulkan_loader?

Commit 7a05c07

Commit `7a05c07`