llm_meta_widget 0.15.0 → 0.15.2

checksums.yaml

@@ -1,7 +1,7 @@
 ---
 SHA256:
-  metadata.gz: 8797cb824bc5421990348a5de7fabd7eb95a0ec0c10f8e2d579d29d3483e7b29
-  data.tar.gz: 8151a9ab5e27e0e1bd1f46d62750e2a41137660d65c501713e9a10e94269858f
+  metadata.gz: 863b128096d9e0813388bb13a03d96a8f2b0f202f2aa5c33a27bdd0a48416c7b
+  data.tar.gz: e14113a514e1fa7ffb09e067cc44150e32e90a61fb5dd80946b9d7cfb34c6ad6
 SHA512:
-  metadata.gz: 12d33498cb17917ebc7f6a5950185b003267c52eab06f50e9632d36e5fd7e9c1bfe54d18134be7bc5e9082ff4c5e087769b0235983ff72aa6ec3dbd8461641af
-  data.tar.gz: 0f2975b8f85714bb11e64b43192755e26b61f90a9a80e6f20871895a76e292bfbb3f245c7e7f679e81331f1ab9be7d0362e91b8ce4e32b96d16ac10ce9eb8ded
+  metadata.gz: c1c3f21142b8de93b5a8216472aa700ac7fd72c589ca810f95eb8e3d368a60045974cec8f56a8e6fd64803da7b6a90a18be296312b7f3dc6f42c3c252ccf0f62
+  data.tar.gz: 5ee508999c3a2fe61e26f258b4362e4fac981097d7a3c5a375d6dcfe129e75f0d415d0a26d1404c3d2bc99e7f140347851b55609ca157f25bdb0b889b3fe1703

data/CHANGELOG.md

@@ -3,6 +3,57 @@
 This file starts at 0.8.0, the first release with a breaking change worth
 calling out. Earlier releases are in the git history (`git log --oneline v0.7.6`).
 
+## 0.15.2
+
+### A page that hides the model picker now learns the context window too
+
+Both requests for the hub's model list sat inside the branches that fill the
+model picker. A page that pins one model and hides the picker therefore never
+learned that model's context window — and hiding the picker is a reasonable
+choice when the host has already chosen the model.
+
+Without a window the resource budget falls back to its 8 KB floor, so a page's
+reference data is reported as skipped even when it would have fitted, and
+oversized tool results are neither trimmed nor evicted, because each of those
+steps returns its input unchanged when it has no window to spend. Release
+0.15.1 could not even warn about it: the warning there fires when a request
+comes back unusable, and here no request was made.
+
+The model list is now fetched whenever the hub is the provider, and only
+filling the picker remains conditional. A page configured to talk straight to
+Ollama is left alone, since there is no such endpoint to ask.
+
+The widget also says so once, in the browser console, if a turn starts with no
+known window — naming the model and what is lost. Three different causes have
+produced an unknown window so far, and each was reported only where it
+happened, if at all.
+
+## 0.15.1
+
+### A repeat visit no longer loses the model's context window or the hub tools
+
+The hub answers `/api/llms` and `/api/mcp_servers` with an ETag and
+`must-revalidate`, so a browser revalidated them on the second page load and
+got `304 Not Modified`. `Response.ok` is false for a 304, and both handlers
+turned that into an empty payload.
+
+Nothing announced it. Without the model's context window the resource budget
+fell to its 8 KB floor, so a page's reference data was reported as skipped even
+when it would have fitted; and tool-result truncation and oldest-result
+eviction both stood down, because each returns its input unchanged when it has
+no window to spend. An empty hub list meanwhile showed as "No hub-registered
+tools available". It all worked on a first load and stopped on every one after,
+which is why a hard reload always looked fine.
+
+Both requests now set `cache: "no-store"`, so no conditional request is sent
+and the answer always carries a body. A response that is not ok is reported in
+the browser console, naming what is lost, rather than being treated as an empty
+result.
+
+The hub stopped sending those ETags as well, so this release is not required to
+fix a page served by an up-to-date hub. It matters for any other server that
+does send them, and for making the next failure audible.
+
 ## 0.15.0
 
 ### The declared global is resolved when a handler is needed, not at startup

data/README.md

@@ -61,7 +61,7 @@ Option 3 if it does not.
 
 ```ruby
 # Gemfile
-gem "llm_meta_widget", "~> 0.15.0"
+gem "llm_meta_widget", "~> 0.15.2"
 ```
 
 ```bash

data/app/assets/javascripts/llm_meta_widget/element.js

@@ -502,9 +502,33 @@ function boot(cfg) {
   				});
   				if (flat.length > 1) modelPicker.style.display = "";
   			}));
-  		} else if (ENABLE_MODEL_PICKER && modelPicker) {
-  			tasks.push(fetch(LLM_BASE + "/api/llms", { headers: { "Accept": "application/json" } })
-  				.then(function(r) { return r.ok ? r.json() : { llms: [] }; })
+  		}
+  		// The hub answers with each model's context window, and every budget below
+  		// is derived from it, so ASK REGARDLESS OF THE PICKER. This used to sit
+  		// inside the picker branches: a page that pinned a model and hid the
+  		// picker — which is a sensible thing to do, and what the standalone E2E
+  		// fixture does — never learned the window at all. The resource budget then
+  		// fell to its 8 KB floor and both tool-result truncation and eviction stood
+  		// down, each returning its input unchanged without a window. No warning was
+  		// possible either, because no request was made to fail.
+  		//
+  		// Skipped only when talking straight to Ollama, which has no /api/llms.
+  		if (LLM_PROVIDER !== "ollama") {
+  			// cache: "no-store" so the browser never sends a conditional request. The
+  			// hub answers /api/llms with an ETag and must-revalidate, so a REPEAT visit
+  			// got 304 Not Modified — and Response.ok is false for a 304, which became
+  			// an empty payload here. The model's context window was then unknown, so
+  			// the resource budget fell to its 8 KB floor and BOTH tool-result
+  			// truncation and eviction stood down, each returning unchanged without a
+  			// window. It worked on a first load and quietly stopped on every one after.
+  			tasks.push(fetch(LLM_BASE + "/api/llms", { headers: { "Accept": "application/json" }, cache: "no-store" })
+  				.then(function(r) {
+  					if (r.ok) return r.json();
+  					console.warn("[llm-meta-widget] /api/llms answered " + r.status +
+  					             "; the model context window is unknown, so reference data is" +
+  					             " budgeted at the floor and oversized tool results are not trimmed");
+  					return { llms: [] };
+  				})
   				.then(function(payload) {
   					// /api/llms is heterogeneous per family:
   					//   - OpenAI/Anthropic/Google: `models: [{ name, display_name, ... }]`
@@ -515,6 +539,13 @@ function boot(cfg) {
   					// configured model is often hosted while the picker shows local ones.
   					llmsPayload = payload;
   					modelWindow = modelWindowFrom(payload, MODEL);
+
+  					// The window is the part every page needs; filling the picker is only for
+  					// pages that show one, and the code below would dereference a null picker.
+					// Without this the TypeError only skips the rest of boot, whose
+					// remaining job is the picker's change handler — so nothing a
+					// no-picker page needs. Kept for intent, not for a visible bug.
+  					if (!(ENABLE_MODEL_PICKER && modelPicker)) return;
   					var flat = [];
   					(payload.llms || []).forEach(function(llm) {
   						if (llm.family !== "ollama") return;
@@ -551,8 +582,16 @@ function boot(cfg) {
   		}
 
   		if (ENABLE_TOOL_PICKER && toolsListEl) {
-  			tasks.push(fetch(TOOL_HUB_BASE + "/api/mcp_servers", { headers: { "Accept": "application/json" } })
-  				.then(function(r) { return r.ok ? r.json() : { mcp_servers: [] }; })
+  			// Same reason as /api/llms above: a 304 on a repeat visit emptied this, and
+  			// the panel then said "No hub-registered tools available" although the hub
+  			// had them.
+  			tasks.push(fetch(TOOL_HUB_BASE + "/api/mcp_servers", { headers: { "Accept": "application/json" }, cache: "no-store" })
+  				.then(function(r) {
+  					if (r.ok) return r.json();
+  					console.warn("[llm-meta-widget] /api/mcp_servers answered " + r.status +
+  					             "; hub-registered tools will be listed as unavailable");
+  					return { mcp_servers: [] };
+  				})
   				.then(function(payload) {
   					hubMcpServers = (payload.mcp_servers || []).filter(function(s) {
   						return anyAllowedByAllowlist(s.name, HUB_TOOLS_ALLOWLIST);
@@ -702,6 +741,7 @@ function boot(cfg) {
   	// call answers, and null forever if the hub is older than the field — in
   	// which case the loop budgets nothing, exactly as it always did.
   	var modelWindow = null;
+  	var warnedNoWindow = false;
   	// The /api/llms answer, kept so switching model in the picker can resolve
   	// that model's window. Without it a visitor who switches from a 262k model
   	// to a 32k one keeps the big window — no eviction, and a resource budget
@@ -1351,6 +1391,18 @@ function boot(cfg) {
   		inputEl.value = "";
   		appendTurn("user", userText, pendingTurnLabel);
 
+  		// An unknown window is not a small window: every budget below then uses a
+  		// floor, and oversized tool results are neither trimmed nor evicted. Say so
+  		// once, at the moment it affects an answer. Three separate causes have led
+  		// here — a 304 read as empty, a page that hid the model picker, and a model
+  		// the hub does not list — and the first two were silent for days.
+  		if (!modelWindow && !warnedNoWindow) {
+  			warnedNoWindow = true;
+  			console.warn("[llm-meta-widget] no context window known for " + MODEL +
+  			             "; reference data is budgeted at the floor and oversized tool" +
+  			             " results are not trimmed");
+  		}
+
   		// Reference data that did not fit, reported once per budget, at the moment
   		// it affects an answer rather than at boot.
   		// Plan the page's resources first: the refusals are recorded while planning,

data/app/assets/javascripts/llm_meta_widget/llm-meta-widget.js

@@ -4269,12 +4269,16 @@ function boot(cfg) {
         });
         if (flat.length > 1) modelPicker.style.display = "";
       }));
-    } else if (ENABLE_MODEL_PICKER && modelPicker) {
-      tasks.push(fetch(LLM_BASE + "/api/llms", { headers: { "Accept": "application/json" } }).then(function(r) {
-        return r.ok ? r.json() : { llms: [] };
+    }
+    if (LLM_PROVIDER !== "ollama") {
+      tasks.push(fetch(LLM_BASE + "/api/llms", { headers: { "Accept": "application/json" }, cache: "no-store" }).then(function(r) {
+        if (r.ok) return r.json();
+        console.warn("[llm-meta-widget] /api/llms answered " + r.status + "; the model context window is unknown, so reference data is budgeted at the floor and oversized tool results are not trimmed");
+        return { llms: [] };
       }).then(function(payload) {
         llmsPayload = payload;
         modelWindow = modelWindowFrom(payload, MODEL);
+        if (!(ENABLE_MODEL_PICKER && modelPicker)) return;
         var flat = [];
         (payload.llms || []).forEach(function(llm) {
           if (llm.family !== "ollama") return;
@@ -4304,8 +4308,10 @@ function boot(cfg) {
       }));
     }
     if (ENABLE_TOOL_PICKER && toolsListEl) {
-      tasks.push(fetch(TOOL_HUB_BASE + "/api/mcp_servers", { headers: { "Accept": "application/json" } }).then(function(r) {
-        return r.ok ? r.json() : { mcp_servers: [] };
+      tasks.push(fetch(TOOL_HUB_BASE + "/api/mcp_servers", { headers: { "Accept": "application/json" }, cache: "no-store" }).then(function(r) {
+        if (r.ok) return r.json();
+        console.warn("[llm-meta-widget] /api/mcp_servers answered " + r.status + "; hub-registered tools will be listed as unavailable");
+        return { mcp_servers: [] };
       }).then(function(payload) {
         hubMcpServers = (payload.mcp_servers || []).filter(function(s) {
           return anyAllowedByAllowlist(s.name, HUB_TOOLS_ALLOWLIST);
@@ -4421,6 +4427,7 @@ function boot(cfg) {
     return Object.assign({}, p, { source: "page" });
   });
   var modelWindow = null;
+  var warnedNoWindow = false;
   var llmsPayload = null;
   var resourceContext = null;
   var resourcePlan = null;
@@ -4864,6 +4871,10 @@ function boot(cfg) {
     if (!userText) return;
     inputEl.value = "";
     appendTurn("user", userText, pendingTurnLabel);
+    if (!modelWindow && !warnedNoWindow) {
+      warnedNoWindow = true;
+      console.warn("[llm-meta-widget] no context window known for " + MODEL + "; reference data is budgeted at the floor and oversized tool results are not trimmed");
+    }
     pageResourcePlansNow();
     if (skippedResources.length && !announcedSkips) {
       announcedSkips = true;

data/lib/llm_meta_widget/version.rb

@@ -1,3 +1,3 @@
 module LlmMetaWidget
-  VERSION = "0.15.0"
+  VERSION = "0.15.2"
 end

metadata

@@ -1,7 +1,7 @@
 --- !ruby/object:Gem::Specification
 name: llm_meta_widget
 version: !ruby/object:Gem::Version
-  version: 0.15.0
+  version: 0.15.2
 platform: ruby
 authors:
 - jdkim