Merge pull request #2621 from esokullu/main

Fix runtime trace config attribution for unlimited and tabless runs

Author: esokullu

src/chrome/src/agent/agent.js

@@ -761,29 +761,38 @@ export class Agent extends LoopDetector {
     return this.conversationIds.get(tabId) || null;
   }
 
+  /**
+   * Snapshot the effective runtime settings for a run. Anything we cannot
+   * observe is omitted rather than guessed: an absent field reads as "unknown"
+   * in a dump, while a hard `false`/'ask' would assert a setting the run never
+   * actually had — exactly the misattribution this metadata exists to prevent.
+   */
   _runtimeTraceConfig(provider, { tabId = null, mode = null } = {}) {
     let extensionVersion = '';
-    let promptTier = 'full';
     try { extensionVersion = chrome.runtime.getManifest().version || ''; } catch {}
-    try { promptTier = provider?.promptTier || 'full'; } catch {}
-    const effectiveMode = mode
-      || (tabId != null ? this._effectiveRunMode(tabId) : 'ask');
+    const effectiveMode = mode || (tabId != null ? this._effectiveRunMode(tabId) : null);
     return normalizeRuntimeTraceConfig({
       extension_version: extensionVersion,
       browser_target: 'chrome',
-      mode: effectiveMode,
-      prompt_tier: promptTier,
+      ...(effectiveMode ? { mode: effectiveMode } : {}),
+      prompt_tier: this._resolvePromptTier(provider),
       screenshot_redaction: this.screenshotRedaction === true,
       strict_secret_mode: this.strictSecretMode === true,
       plan_before_act_mode: this._normalizePlanBeforeActMode(this.planBeforeActMode),
       auto_screenshot: this.autoScreenshot,
       use_site_adapters: this.useSiteAdapters === true,
       web_mcp_enabled: this.webMcpEnabled === true,
-      api_mutations_allowed: tabId != null && this.apiAllowedTabs.has(tabId),
       user_memory_enabled: this.userMemoryEnabled === true,
-      selection_grounded: tabId != null && this.selectionGroundingScopes.has(tabId),
+      // Per-tab authorizations only mean something in a tab's context.
+      ...(tabId != null ? {
+        api_mutations_allowed: this.apiAllowedTabs.has(tabId),
+        selection_grounded: this.selectionGroundingScopes.has(tabId),
+      } : {}),
       image_detail: this.imageDetail,
-      max_agent_steps: this.maxSteps,
+      // The steps slider stores 0 for "unlimited", which the agent hydrates as
+      // Infinity. Round-trip that back to 0 so an unlimited run records as
+      // unlimited instead of being dropped as a non-integer.
+      max_agent_steps: Number.isFinite(this.maxSteps) ? this.maxSteps : 0,
       max_image_dimension: this.maxImageDimension,
       max_screenshots_per_turn: this.maxScreenshotsPerTurn,
     });
@@ -9357,14 +9366,15 @@ Rules: no prose intro, no conclusion, no "this screenshot shows...", no layout d
   }
 
   /**
-   * Resolve the active provider's prompt tier ('compact' | 'mid' | 'full').
-   * The provider getter already forces 'full' for cloud providers and applies
-   * the per-category defaults (local → 'mid'); we just guard the case where
-   * no provider is ready yet (fall back to the full prompt).
+   * Resolve a provider's prompt tier ('compact' | 'mid' | 'full'), defaulting
+   * to the active provider. The provider getter already forces 'full' for
+   * cloud providers and applies the per-category defaults (local → 'mid'); we
+   * just guard the case where no provider is ready yet (fall back to the full
+   * prompt).
    */
-  _resolvePromptTier() {
+  _resolvePromptTier(provider = null) {
     try {
-      return this.providerManager.getActive().promptTier || 'full';
+      return (provider || this.providerManager.getActive()).promptTier || 'full';
     } catch { return 'full'; }
   }
 

src/chrome/src/trace/runtime-config.js

@@ -17,9 +17,13 @@ const BOOLEAN_FIELDS = Object.freeze([
   'selection_grounded',
 ]);
 
+// Bounds keep the payload sane, not to re-validate settings: each range is a
+// superset of what the agent's own normalizers can hold (steps ≤ 200 with 0 =
+// unlimited, dimension ≤ 2048, screenshots ≤ 5), so a legitimate setting is
+// never silently dropped for being out of range.
 const INTEGER_RANGES = Object.freeze({
   max_agent_steps: [0, 10_000],
-  max_image_dimension: [256, 16_384],
+  max_image_dimension: [1, 16_384],
   max_screenshots_per_turn: [0, 1_000],
 });
 

src/firefox/src/agent/agent.js

@@ -860,29 +860,38 @@ export class Agent extends LoopDetector {
     return this.conversationIds.get(tabId) || null;
   }
 
+  /**
+   * Snapshot the effective runtime settings for a run. Anything we cannot
+   * observe is omitted rather than guessed: an absent field reads as "unknown"
+   * in a dump, while a hard `false`/'ask' would assert a setting the run never
+   * actually had — exactly the misattribution this metadata exists to prevent.
+   */
   _runtimeTraceConfig(provider, { tabId = null, mode = null } = {}) {
     let extensionVersion = '';
-    let promptTier = 'full';
     try { extensionVersion = browser.runtime.getManifest().version || ''; } catch {}
-    try { promptTier = provider?.promptTier || 'full'; } catch {}
-    const effectiveMode = mode
-      || (tabId != null ? this._effectiveRunMode(tabId) : 'ask');
+    const effectiveMode = mode || (tabId != null ? this._effectiveRunMode(tabId) : null);
     return normalizeRuntimeTraceConfig({
       extension_version: extensionVersion,
       browser_target: 'firefox',
-      mode: effectiveMode,
-      prompt_tier: promptTier,
+      ...(effectiveMode ? { mode: effectiveMode } : {}),
+      prompt_tier: this._resolvePromptTier(provider),
       screenshot_redaction: this.screenshotRedaction === true,
       strict_secret_mode: this.strictSecretMode === true,
       plan_before_act_mode: this._normalizePlanBeforeActMode(this.planBeforeActMode),
       auto_screenshot: this.autoScreenshot,
       use_site_adapters: this.useSiteAdapters === true,
       web_mcp_enabled: this.webMcpEnabled === true,
-      api_mutations_allowed: tabId != null && this.apiAllowedTabs.has(tabId),
       user_memory_enabled: this.userMemoryEnabled === true,
-      selection_grounded: tabId != null && this.selectionGroundingScopes.has(tabId),
+      // Per-tab authorizations only mean something in a tab's context.
+      ...(tabId != null ? {
+        api_mutations_allowed: this.apiAllowedTabs.has(tabId),
+        selection_grounded: this.selectionGroundingScopes.has(tabId),
+      } : {}),
       image_detail: this.imageDetail,
-      max_agent_steps: this.maxSteps,
+      // The steps slider stores 0 for "unlimited", which the agent hydrates as
+      // Infinity. Round-trip that back to 0 so an unlimited run records as
+      // unlimited instead of being dropped as a non-integer.
+      max_agent_steps: Number.isFinite(this.maxSteps) ? this.maxSteps : 0,
       max_image_dimension: this.maxImageDimension,
       max_screenshots_per_turn: this.maxScreenshotsPerTurn,
     });
@@ -8152,14 +8161,15 @@ Rules: no prose intro, no conclusion, no "this screenshot shows...", no layout d
   }
 
   /**
-   * Resolve the active provider's prompt tier ('compact' | 'mid' | 'full').
-   * The provider getter already forces 'full' for cloud providers and applies
-   * the per-category defaults (local → 'mid'); we just guard the case where
-   * no provider is ready yet (fall back to the full prompt).
+   * Resolve a provider's prompt tier ('compact' | 'mid' | 'full'), defaulting
+   * to the active provider. The provider getter already forces 'full' for
+   * cloud providers and applies the per-category defaults (local → 'mid'); we
+   * just guard the case where no provider is ready yet (fall back to the full
+   * prompt).
    */
-  _resolvePromptTier() {
+  _resolvePromptTier(provider = null) {
     try {
-      return this.providerManager.getActive().promptTier || 'full';
+      return (provider || this.providerManager.getActive()).promptTier || 'full';
     } catch { return 'full'; }
   }
 

src/firefox/src/trace/runtime-config.js

@@ -17,9 +17,13 @@ const BOOLEAN_FIELDS = Object.freeze([
   'selection_grounded',
 ]);
 
+// Bounds keep the payload sane, not to re-validate settings: each range is a
+// superset of what the agent's own normalizers can hold (steps ≤ 200 with 0 =
+// unlimited, dimension ≤ 2048, screenshots ≤ 5), so a legitimate setting is
+// never silently dropped for being out of range.
 const INTEGER_RANGES = Object.freeze({
   max_agent_steps: [0, 10_000],
-  max_image_dimension: [256, 16_384],
+  max_image_dimension: [1, 16_384],
   max_screenshots_per_turn: [0, 1_000],
 });
 

test/run.js

@@ -4177,12 +4177,22 @@ test('trace record and JSON exports carry WebBrain version metadata', () => {
     assert.match(recorder, /runtimeConfig: normalizeRuntimeTraceConfig\(meta\.runtimeConfig\)/, `${label}: run record should retain only allowlisted runtime settings`);
     assert.match(agent, new RegExp(`webbrainVersion: ${runtimeName}\\.runtime\\.getManifest\\(\\)\\.version`), `${label}: trace start should read the runtime manifest`);
     assert.match(agent, /runtimeConfig: this\._runtimeTraceConfig\(provider, \{ tabId, mode \}\)/, `${label}: trace start should snapshot effective runtime settings`);
+    assert.match(
+      agent,
+      /runtimeConfig: this\._runtimeTraceConfig\(this\.providerManager\?\.getActive\?\.\(\), \{\s*tabId,\s*mode: 'act',\s*\}\)/,
+      `${label}: workflow runs should snapshot effective runtime settings too`,
+    );
     assert.match(traceUi, new RegExp(`exportedByWebBrainVersion: ${runtimeName}\\.runtime\\.getManifest\\(\\)\\.version`), `${label}: JSON export should identify the exporting build`);
     assert.match(traceUi, /schema: 'webbrain-trace\/1'/, `${label}: additive version metadata should retain the v1 schema`);
   }
 });
 
 test('runtime trace config is versioned, bounded, and secret-free in both browsers', () => {
+  assert.equal(
+    fs.readFileSync(path.join(ROOT, 'src/chrome/src/trace/runtime-config.js'), 'utf8'),
+    fs.readFileSync(path.join(ROOT, 'src/firefox/src/trace/runtime-config.js'), 'utf8'),
+    'chrome and firefox runtime trace config modules must remain byte-identical',
+  );
   const candidate = {
     schema_version: 99,
     extension_version: '24.7.0-beta.1',
@@ -4234,10 +4244,29 @@ test('runtime trace config is versioned, bounded, and secret-free in both browse
     browser_target: 'safari',
     mode: 'admin',
     max_agent_steps: Infinity,
-    max_image_dimension: 42,
+    max_image_dimension: 1568.5,
+    max_screenshots_per_turn: -1,
     screenshot_redaction: 'yes',
   });
   assert.deepEqual(rejected, { schema_version: 1 });
+
+  // Integer ranges exist to bound the payload, not to re-validate settings:
+  // every value the agent's own normalizers can produce must survive, or the
+  // dump silently loses the setting it was added to attribute.
+  for (const [field, value] of [
+    ['max_agent_steps', 0],
+    ['max_agent_steps', 200],
+    ['max_image_dimension', 1],
+    ['max_image_dimension', 2048],
+    ['max_screenshots_per_turn', 0],
+    ['max_screenshots_per_turn', 5],
+  ]) {
+    assert.equal(
+      RuntimeTraceConfigCh.normalizeRuntimeTraceConfig({ [field]: value })[field],
+      value,
+      `${field}=${value} is producible by the agent and must not be dropped`,
+    );
+  }
 });
 
 test('trace recorders normalize done only from explicit loop error evidence', () => {
@@ -32472,6 +32501,23 @@ test('WebBrain Cloud groups every generation in a stable conversation session wi
     assert.equal(main.webbrainRuntimeConfig?.strict_secret_mode, true, `${label}: strict secret setting missing`);
     assert.equal(main.webbrainRuntimeConfig?.api_mutations_allowed, true, `${label}: per-tab API authorization missing`);
     assert.ok(!JSON.stringify(main.webbrainRuntimeConfig).includes('apiKey'), `${label}: runtime metadata must remain an allowlist`);
+    assert.equal(main.webbrainRuntimeConfig?.max_agent_steps, agent.maxSteps, `${label}: step budget missing`);
+
+    // "Unlimited" steps hydrate as Infinity; record the stored 0 sentinel so an
+    // unlimited run is attributable instead of missing the field entirely.
+    const previousMaxSteps = agent.maxSteps;
+    agent.maxSteps = Infinity;
+    const unlimited = agent._cloudGenerationOptions(cloud, {}, { tabId, generationName: 'main' });
+    assert.equal(unlimited.webbrainRuntimeConfig?.max_agent_steps, 0, `${label}: unlimited step budget should record as 0`);
+    agent.maxSteps = previousMaxSteps;
+
+    // Without a tab there is no observable mode or per-tab authorization, so
+    // those fields must be absent rather than asserted as ask/false.
+    const untabbed = agent._runtimeTraceConfig(cloud);
+    assert.equal('mode' in untabbed, false, `${label}: unknown mode should be omitted, not guessed`);
+    assert.equal('api_mutations_allowed' in untabbed, false, `${label}: per-tab API authorization should be omitted without a tab`);
+    assert.equal('selection_grounded' in untabbed, false, `${label}: selection grounding should be omitted without a tab`);
+    assert.equal(untabbed.browser_target, label, `${label}: browser target should survive without a tab`);
 
     const byoOptions = agent._cloudGenerationOptions({ config: { providerName: 'openai' } }, { temperature: 0 }, { tabId, generationName: 'memory' });
     assert.deepEqual(byoOptions, { temperature: 0 }, `${label}: BYO provider received Cloud collection fields`);