solid_loop 0.0.4 → 0.0.5

data/app/services/solid_loop/middlewares/message_building.rb

@@ -36,10 +36,28 @@ module SolidLoop
         all_messages = env.loop.messages.where(is_hidden: false)
           .order(Arel.sql("COALESCE(conversation_order, id)")).includes(:tool_calls)
 
-        # Build tool_call_id -> function_name lookup for tool response messages
-        tool_call_name_index = all_messages
+        # Build tool_call_id -> function_name lookup for tool response messages.
+        #
+        # Seeded from `tool_calls_raw` before the persisted rows, because a
+        # REJECTED call never becomes a `ToolCall` row: when the model sends
+        # arguments that are not valid JSON, the gem writes a "Tool call
+        # rejected" tool message asking it to resend, and that message's
+        # tool_call_id resolves to nothing. Resolving it to nil put `"name":
+        # null` on the wire — which OpenAI-compatible servers vary on. vLLM
+        # ignored it; llama.cpp rejects the whole request with
+        # `type must be string, but is null`, and since the offending message
+        # stays in history the loop then fails on every retry, permanently.
+        # The name is right there in the raw payload, so use it.
+        tool_call_name_index = all_messages.each_with_object({}) do |m, h|
+          Array(m.tool_calls_raw).each do |raw|
+            id   = raw["id"] || raw[:id]
+            name = raw.dig("function", "name") || raw.dig(:function, :name)
+            h[id] = name if id.present? && name.present?
+          end
+        end
+        all_messages
           .flat_map { |m| m.tool_calls.sort_by(&:id) }
-          .each_with_object({}) { |tc, h| h[tc.tool_call_id] = tc.function_name }
+          .each { |tc| tool_call_name_index[tc.tool_call_id] = tc.function_name }
 
         all_messages.each do |msg|
           m = { role: msg.role, content: msg.content }
@@ -59,7 +77,9 @@ module SolidLoop
 
           if msg.role == "tool"
             m[:tool_call_id] = msg.tool_call_id
-            m[:name] = tool_call_name_index[msg.tool_call_id]
+            # `name` is optional in the schema; omit it rather than send null.
+            name = tool_call_name_index[msg.tool_call_id]
+            m[:name] = name if name.present?
             env.triggering_message = msg
           elsif msg.role == "user"
             env.triggering_message = msg
@@ -86,6 +106,20 @@ module SolidLoop
           stream: !!env.agent.streaming?
         }
         env.payload[:tools] = tools if tools.any?
+
+        # Only when the agent opts in: an absent key lets the provider apply its
+        # own default. Set before the merge below so `llm_params` can override it.
+        max_tokens = env.agent.max_tokens
+        env.payload[:max_tokens] = max_tokens if max_tokens
+
+        effort = env.agent.reasoning_effort
+        env.payload[:reasoning_effort] = effort if effort.present?
+
+        # Host-supplied keys win: merged last so an agent can add what the gem
+        # does not model, or override what it does. Symbolized so a host writing
+        # string keys cannot silently produce a duplicate entry alongside ours.
+        extra = env.agent.llm_params
+        env.payload = env.payload.deep_merge(extra.symbolize_keys) if extra.present?
       end
     end
   end

data/app/services/solid_loop/middlewares/response_parsing.rb

@@ -73,8 +73,24 @@ module SolidLoop
           tokens_prompt: usage_data[:tokens_prompt],
           tokens_completion: usage_data[:tokens_completion],
           tokens_total: usage_data[:tokens_total],
-          tokens_prompt_cached: usage_data[:tokens_prompt_cached],
-          tps: usage_data[:tps],
+          # Cached tokens are a SUBSET of the prompt, never more. Clamped at the
+          # persistence boundary rather than in one parser because the invariant
+          # belongs to the row: a provider that reports the two on different
+          # bases (llama.cpp counts `prompt_n` fresh-only) would otherwise store
+          # a cache figure larger than the prompt it supposedly came from, and
+          # every derived hit-rate would exceed 100%.
+          tokens_prompt_cached: usage_data[:tokens_prompt_cached].to_i
+            .clamp(0, usage_data[:tokens_prompt].to_i),
+          # The provider's own rate wins when it reports one (llama.cpp sends
+          # `predicted_per_second`); otherwise derive it here, where the decode
+          # window is known. Without this the column stayed 0.0 for every
+          # OpenAI/Anthropic/Gemini-dialect turn, so any SQL aggregate over it
+          # silently read as "zero throughput" instead of "not measured".
+          tps: (usage_data[:tps].to_f.positive? ? usage_data[:tps] : SolidLoop::Message.derive_tps(
+            tokens_completion: usage_data[:tokens_completion],
+            duration_generation: full_duration,
+            ttft: ttft
+          )),
           ttft: ttft,
           duration_thinking: duration_thinking,
           duration_generation: full_duration,
@@ -276,7 +292,7 @@ module SolidLoop
         env.error_headers = safe_response_headers(response)
         env.error_body = body
 
-        error_msg = "HTTP #{status}: #{body}"
+        error_msg = SolidLoop::Redaction.redact_credentials("HTTP #{status}: #{body}")
 
         is_context_limit = (status == 400) && (
           body.include?("maximum context length") ||

data/app/views/layouts/solid_loop/admin.html.erb

@@ -19,7 +19,7 @@
       --sl-bg-color: #0e1621;
       --sl-sidebar-bg: #17212b;
       --sl-hover-bg: #202b36;
-      --sl-border-color: #101921;
+      --sl-border-color: #2c3d4d;
       --sl-text-color: #f5f5f5;
       --sl-muted-text: #7f91a4;
       --sl-msg-in: #182533;
@@ -86,6 +86,17 @@
       overflow: hidden;
     }
 
+    .sl-subject-link {
+      color: var(--sl-muted-text);
+      text-decoration: none;
+      border-bottom: 1px dotted var(--sl-muted-text);
+    }
+
+    .sl-subject-link:hover {
+      color: var(--sl-accent);
+      border-bottom-color: var(--sl-accent);
+    }
+
     .sl-content-area {
       flex: 1;
       overflow-y: auto;
@@ -96,8 +107,8 @@
     .sl-card {
       background-color: var(--sl-sidebar-bg);
       border-radius: var(--sl-radius);
-      padding: 20px;
-      margin-bottom: 20px;
+      padding: 12px 14px;
+      margin-bottom: 10px;
       border: 1px solid var(--sl-border-color);
     }
 
@@ -137,6 +148,62 @@
     .sl-badge--failed, .sl-badge--error { background-color: rgba(255, 107, 107, 0.2); color: var(--sl-error); }
     .sl-badge--paused, .sl-badge--frozen { background-color: rgba(133, 77, 14, 0.2); color: #eab308; }
 
+    /* Event families — every event badge used to be the same blue, so
+       llm_completion and mcp_initialize looked like the same kind of thing.
+       LLM traffic is violet, MCP tool invocations are teal, and the MCP
+       lifecycle chatter (initialize / list_tools / list_prompts) is a quiet
+       outline so it recedes behind the calls that actually do work. */
+    .sl-badge--llm { background-color: rgba(167, 139, 250, 0.18); color: #a78bfa; }
+    .sl-badge--mcp { background-color: rgba(45, 212, 191, 0.18); color: #2dd4bf; }
+    .sl-badge--mcp-meta {
+      background-color: transparent;
+      color: var(--sl-muted-text);
+      border: 1px solid var(--sl-border-color);
+      padding: 3px 7px;
+    }
+
+    /* HTTP / wire outcome */
+    .sl-badge--http-ok { background-color: rgba(49, 181, 69, 0.15); color: var(--sl-success); }
+    .sl-badge--http-warn { background-color: rgba(241, 196, 15, 0.2); color: var(--sl-pending); }
+    .sl-badge--http-error { background-color: rgba(255, 107, 107, 0.2); color: var(--sl-error); }
+
+    /* Metric cells — durations, tokens, counts. Tabular figures so a column of
+       numbers aligns and an outlier is spottable by shape alone; severity
+       colours are applied by the duration/idle thresholds in MetricsHelper. */
+    .sl-metric {
+      font-variant-numeric: tabular-nums;
+      font-feature-settings: "tnum";
+      white-space: nowrap;
+    }
+
+    .sl-metric--warn { color: var(--sl-pending); }
+    .sl-metric--slow { color: var(--sl-error); font-weight: 600; }
+    .sl-metric--blank { color: var(--sl-muted-text); }
+
+    .sl-metric__sub {
+      display: block;
+      margin-top: 2px;
+      font-size: 0.8em;
+      color: var(--sl-muted-text);
+      font-variant-numeric: tabular-nums;
+    }
+
+    /* Long opaque identifiers (MCP session ids) — truncated, with the full
+       value one click away instead of eating half the table width. */
+    .sl-copyable {
+      display: inline-flex;
+      align-items: center;
+      gap: 8px;
+    }
+
+    .sl-copyable__text {
+      font-family: ui-monospace, SFMono-Regular, Menlo, Consolas, monospace;
+      font-size: 0.85em;
+      color: var(--sl-text-color);
+    }
+
+    .sl-copyable__btn { flex-shrink: 0; }
+
     .sl-badge--live {
       background-color: rgba(255, 107, 107, 0.2);
       color: var(--sl-error);
@@ -207,10 +274,9 @@
     .sl-json {
       font-family: monospace;
       font-size: 0.85rem;
-      white-space: pre-wrap;
       word-break: break-word;
       background: var(--sl-bg-color);
-      padding: 10px;
+      padding: 8px 10px;
       border-radius: 6px;
       border: 1px solid var(--sl-border-color);
     }
@@ -295,44 +361,89 @@
     
     .sl-message__header {
       display: flex;
-      justify-content: space-between;
       align-items: center;
-      margin-bottom: 10px;
+      gap: 8px;
+      margin-bottom: 6px;
     }
-    
+
     .sl-message__role {
       text-transform: uppercase;
-      font-size: 0.85em;
+      font-size: 0.75em;
+      letter-spacing: 0.06em;
       color: var(--sl-muted-text);
     }
-    
+
+    .sl-message__model {
+      font-weight: normal;
+      font-size: 0.75em;
+    }
+
+    /* Pushes the timestamp to the right without a space-between that would
+       strand the model badge in the middle of the row. */
     .sl-message__time {
-      font-size: 0.8em;
+      margin-left: auto;
+      font-size: 0.75em;
       color: var(--sl-muted-text);
     }
-    
+
+    .sl-message__body {
+      margin: 0;
+      font-family: inherit;
+      font-size: inherit;
+    }
+
     .sl-message__tools {
-      margin-top: 10px;
-      color: var(--sl-pending);
+      display: flex;
+      flex-wrap: wrap;
+      align-items: center;
+      gap: 5px;
+      margin-top: 8px;
     }
-    
+
+    .sl-message__tools-label {
+      text-transform: uppercase;
+      font-size: 0.7em;
+      letter-spacing: 0.06em;
+      color: var(--sl-muted-text);
+    }
+
+    .sl-badge__link { color: inherit; text-decoration: none; }
+
+    /* Collapsed by default: reasoning is routinely thousands of characters and
+       used to push the metrics off-screen on every single turn. */
     .sl-message__reasoning {
-      margin-top: 10px;
-      padding-left: 10px;
-      border-left: 2px solid var(--sl-border-color);
+      margin-top: 8px;
+      padding-left: 8px;
+      border-left: 2px solid var(--sl-accent);
       color: var(--sl-muted-text);
     }
-    
+
+    .sl-message__reasoning-toggle {
+      cursor: pointer;
+      font-size: 0.8em;
+      text-transform: uppercase;
+      letter-spacing: 0.06em;
+      user-select: none;
+    }
+
+    .sl-message__reasoning-toggle:hover { color: var(--sl-text-color); }
+    .sl-message__reasoning pre { margin: 6px 0 0; }
+
+    .sl-message__cached { color: var(--sl-success); }
+
     .sl-message__footer {
-      margin-top: 10px;
-      padding-top: 8px;
+      margin-top: 8px;
+      padding-top: 6px;
       border-top: 1px solid var(--sl-border-color);
       display: flex;
       flex-wrap: wrap;
-      gap: 15px;
-      font-size: 0.8em;
+      align-items: center;
+      gap: 12px;
+      font-size: 0.75em;
       color: var(--sl-muted-text);
     }
+
+    .sl-message__footer-item--end { margin-left: auto; }
     
     .sl-message__footer-item {
       display: flex;