deepline 0.2.47 → 0.2.48

package/dist/bundling-sources/sdk/src/release.ts

@@ -160,7 +160,7 @@ export const SDK_RELEASE = {
   // 0.2.0 makes Dataset Handles uniformly async-only after 0.1.320 briefly
   // exposed storage-dependent synchronous access. This deliberate minor
   // release keeps lazy paging semantics independent of row residency.
-  version: '0.2.47',
+  version: '0.2.48',
   contracts: {
     api: {
       name: 'sdk-http-api',

package/dist/bundling-sources/shared_libs/play-runtime/context.ts

@@ -1102,6 +1102,7 @@ function durableCtxKey(input: {
   id: string;
   semanticKey?: string | null;
   staleAfterSeconds?: number | null;
+  cacheEpochMs?: number;
 }): string {
   if (input.operation === 'tool') {
     throw new Error('Tool calls use tool receipt keys.');
@@ -1113,6 +1114,7 @@ function durableCtxKey(input: {
     id: input.id,
     semanticKey: input.semanticKey,
     staleAfterSeconds: input.staleAfterSeconds,
+    cacheEpochMs: input.cacheEpochMs,
   });
 }
 
@@ -1391,6 +1393,7 @@ export class PlayContextImpl implements ScalarPlayAuthoringRuntimeContext {
   private fixtureProviderPacingBypassLogged = false;
   private fixtureProviderPacingEnforcementLogged = false;
   private checkpoint: PlayCheckpoint;
+  private readonly durableCallCacheEpochMs: number;
   /**
    * Durable tool receipts are the replay/cache authority for the execution
    * paths the host supports. Keeping the same completed result in this
@@ -1664,6 +1667,14 @@ export class PlayContextImpl implements ScalarPlayAuthoringRuntimeContext {
     };
     this.#options = options;
     this.checkpoint = options.checkpoint ?? emptyCheckpoint();
+    const checkpointCacheEpochMs = this.checkpoint.durableCallCacheEpochMs;
+    this.durableCallCacheEpochMs =
+      typeof checkpointCacheEpochMs === 'number' &&
+      Number.isFinite(checkpointCacheEpochMs) &&
+      checkpointCacheEpochMs >= 0
+        ? checkpointCacheEpochMs
+        : Date.now();
+    this.checkpoint.durableCallCacheEpochMs = this.durableCallCacheEpochMs;
     this.durableMappedToolResultsBackedByReceipts = Boolean(
       (options.claimRuntimeStepReceipt || options.claimRuntimeStepReceipts) &&
       (options.getRuntimeStepReceipt || options.getRuntimeStepReceipts) &&
@@ -3104,6 +3115,7 @@ export class PlayContextImpl implements ScalarPlayAuthoringRuntimeContext {
       }),
       providerActionVersion,
       staleAfterSeconds: input.staleAfterSeconds,
+      cacheEpochMs: this.durableCallCacheEpochMs,
     });
   }
 
@@ -3286,6 +3298,7 @@ export class PlayContextImpl implements ScalarPlayAuthoringRuntimeContext {
         id,
         semanticKey: opts.semanticKey,
         staleAfterSeconds: stalePolicy.staleAfterSeconds,
+        cacheEpochMs: this.durableCallCacheEpochMs,
       });
     return await executeWithDurableRuntimeReceipt<T>({
       operation,

package/dist/bundling-sources/shared_libs/play-runtime/ctx-types.ts

@@ -779,6 +779,11 @@ export interface ContextOptions {
 }
 
 export interface PlayCheckpoint {
+  /**
+   * Clock captured when this logical run first creates its context. All
+   * staleAfterSeconds receipt buckets use this value across durable resumes.
+   */
+  durableCallCacheEpochMs?: number;
   /** Waterfall batches that have completed: key = `${toolName}:${provider}`, value = results array. */
   completedBatches: Record<string, BatchResult[]>;
   /** Tool call batches that have completed: key = toolId, value = sparse row-cache-key -> result map. */

package/dist/bundling-sources/shared_libs/play-runtime/durable-call-cache.ts

@@ -42,6 +42,8 @@ export function buildDurableToolCallCacheKey(input: {
   providerActionVersion?: string | null;
   cachePolicyVersion?: string | null;
   staleAfterSeconds?: number | null;
+  /** Run-stable clock used to choose the stale bucket. */
+  cacheEpochMs?: number;
   playLocalScope?: string | null;
 }): string {
   const orgId = input.orgId?.trim() || 'org';
@@ -70,6 +72,7 @@ export function buildDurableToolCallCacheKey(input: {
         input.cachePolicyVersion ?? DURABLE_CALL_CACHE_POLICY_VERSION,
       staleBucket: durableCacheStaleBucket({
         staleAfterSeconds: input.staleAfterSeconds,
+        nowMs: input.cacheEpochMs,
       }),
     }),
   );
@@ -84,6 +87,8 @@ export function buildDurableCtxCallCacheKey(input: {
   semanticKey?: string | null;
   cachePolicyVersion?: string | null;
   staleAfterSeconds?: number | null;
+  /** Run-stable clock used to choose the stale bucket. */
+  cacheEpochMs?: number;
 }): string {
   const orgId = input.orgId?.trim() || 'org';
   const playId = input.playId?.trim() || 'play';
@@ -104,6 +109,7 @@ export function buildDurableCtxCallCacheKey(input: {
         input.cachePolicyVersion ?? DURABLE_CALL_CACHE_POLICY_VERSION,
       staleBucket: durableCacheStaleBucket({
         staleAfterSeconds: input.staleAfterSeconds,
+        nowMs: input.cacheEpochMs,
       }),
     }),
   );

package/dist/bundling-sources/shared_libs/play-runtime/runner-backends/types.ts

@@ -32,6 +32,7 @@ export type PlayRunnerRuntimeResource = {
 
 export type RuntimeResourceTerminalReason =
   | 'completed'
+  | 'suspended'
   | 'runner_failed'
   | 'sandbox_missing'
   | 'sandbox_killed'

package/dist/cli/index.js

@@ -1044,7 +1044,7 @@ var SDK_RELEASE = {
   // 0.2.0 makes Dataset Handles uniformly async-only after 0.1.320 briefly
   // exposed storage-dependent synchronous access. This deliberate minor
   // release keeps lazy paging semantics independent of row residency.
-  version: "0.2.47",
+  version: "0.2.48",
   contracts: {
     api: {
       name: "sdk-http-api",

package/dist/cli/index.mjs

@@ -1030,7 +1030,7 @@ var SDK_RELEASE = {
   // 0.2.0 makes Dataset Handles uniformly async-only after 0.1.320 briefly
   // exposed storage-dependent synchronous access. This deliberate minor
   // release keeps lazy paging semantics independent of row residency.
-  version: "0.2.47",
+  version: "0.2.48",
   contracts: {
     api: {
       name: "sdk-http-api",

package/dist/index.js

@@ -763,7 +763,7 @@ var SDK_RELEASE = {
   // 0.2.0 makes Dataset Handles uniformly async-only after 0.1.320 briefly
   // exposed storage-dependent synchronous access. This deliberate minor
   // release keeps lazy paging semantics independent of row residency.
-  version: "0.2.47",
+  version: "0.2.48",
   contracts: {
     api: {
       name: "sdk-http-api",

package/dist/index.mjs

@@ -689,7 +689,7 @@ var SDK_RELEASE = {
   // 0.2.0 makes Dataset Handles uniformly async-only after 0.1.320 briefly
   // exposed storage-dependent synchronous access. This deliberate minor
   // release keeps lazy paging semantics independent of row residency.
-  version: "0.2.47",
+  version: "0.2.48",
   contracts: {
     api: {
       name: "sdk-http-api",

package/dist/viewer/viewer.css

@@ -881,6 +881,21 @@ body {
 .expectation-result.pass { color: var(--green); }
 .expectation-result.fail { color: var(--red); }
 .expectation-result.unknown { color: var(--yellow); }
+.infra-signals {
+    margin-top: 10px;
+    border: 1px solid color-mix(in srgb, var(--red) 45%, var(--viewer-border));
+    border-radius: var(--viewer-radius-md);
+    background: color-mix(in srgb, var(--red) 5%, var(--bg));
+    color: var(--text-dim);
+    font-size: 10px;
+}
+.infra-signals-heading { padding: 7px 9px; border-bottom: 1px solid var(--viewer-border); color: var(--red); font-weight: 700; letter-spacing: .04em; text-transform: uppercase; }
+.infra-signals ul { margin: 0; padding: 0; list-style: none; }
+.infra-signals li { padding: 7px 9px; }
+.infra-signals li + li { border-top: 1px solid var(--viewer-border); }
+.infra-signals strong { color: var(--text); font-family: var(--viewer-font-mono); }
+.infra-signals span { margin-left: 7px; color: var(--text-dimmer); font-family: var(--viewer-font-mono); }
+.infra-signals p { margin: 4px 0 0; overflow-wrap: anywhere; }
 .lane-prompt, .lane-result { border-bottom: 1px solid var(--viewer-border); }
 .lane-prompt summary, .lane-result summary { cursor: pointer; padding: 8px 13px; color: var(--text-dim); font-size: 11px; font-weight: 650; }
 .lane-prompt .prompt-text, .lane-result .md-rendered { padding: 0 13px 12px; font-size: 12px; }

package/dist/viewer/viewer.js

@@ -1354,16 +1354,38 @@ function renderExpectationMiniTable(session) {
     </div>`;
 }
 
+function normalizedInfraSignals(metric) {
+    if (!metric || !Array.isArray(metric.infra_signals)) return [];
+    return metric.infra_signals.filter(signal => signal && typeof signal === 'object').map(signal => ({
+        code: String(signal.code || 'unknown_infrastructure_error'),
+        source: String(signal.source || 'unknown'),
+        message: String(signal.message || 'No detail was recorded.'),
+    }));
+}
+
+function renderInfraSignals(metric) {
+    const signals = normalizedInfraSignals(metric);
+    if (!signals.length) return '';
+    const rows = signals.map(signal => `<li><strong>${esc(signal.code)}</strong><span>${esc(signal.source)}</span><p>${esc(signal.message)}</p></li>`).join('');
+    return `<div class="infra-signals"><div class="infra-signals-heading">Infrastructure signals</div><ul>${rows}</ul></div>`;
+}
+
 function renderLaneSession(session) {
     const i = session.sourceIndex;
     const d = session.dimensions;
     const prefix = 'lane-' + i;
     const errorRate = session.stats.tool_calls > 0 ? Math.round(session.stats.tool_errors / session.stats.tool_calls * 100) : 0;
+    const metric = session.evalMetric || {};
+    const infraAffected = metric.sample_validity === 'infra_affected';
+    const infraSignals = normalizedInfraSignals(metric);
+    const infraCodes = [...new Set(infraSignals.map(signal => signal.code))].join(', ');
+    const embeddedFailures = Number(metric.embedded_failure_count || 0);
     let html = `<article class="lane-session">
         <header class="lane-session-header">
             <div class="lane-title-row"><div><span class="run-kicker">Run ${esc(d.run)}</span><h3>${esc(d.agent)} / ${esc(d.model)}</h3></div><button class="download-btn" onclick="downloadJsonl(${i})">JSONL</button></div>
-            <div class="dimension-chips"><span>${esc(d.skill)}</span><span>${formatDuration(session.stats.duration_s)}</span><span>${session.stats.tool_calls} calls</span><span class="${session.stats.tool_errors ? 'chip-error' : ''}">${session.stats.tool_errors} errors${errorRate ? ' · ' + errorRate + '%' : ''}</span></div>
+            <div class="dimension-chips"><span>${esc(d.skill)}</span><span>${formatDuration(session.stats.duration_s)}</span><span>${session.stats.tool_calls} calls</span><span class="${session.stats.tool_errors ? 'chip-error' : ''}">${session.stats.tool_errors} errors${errorRate ? ' · ' + errorRate + '%' : ''}</span>${embeddedFailures ? `<span class="chip-error">${embeddedFailures} embedded failures</span>` : ''}${infraAffected ? `<span class="chip-error">infra affected${infraCodes ? ' · ' + esc(infraCodes) : ''}</span>` : ''}</div>
             ${renderExpectationMiniTable(session)}
+            ${renderInfraSignals(metric)}
         </header>`;
     if (!session.transcript_available) {
         html += '<div class="transcript-unavailable">No transcript was retained for this completed eval. Its scored expectations remain included in this comparison.</div>';
@@ -1431,6 +1453,7 @@ function downloadJsonl(i) {
 
 function renderSession(i) {
     const s = SESSIONS[i];
+    const metric = s.evalMetric || {};
     const main = document.getElementById('main-content');
     main.classList.add('full-width');
 
@@ -1458,11 +1481,22 @@ function renderSession(i) {
     html += item('Turns', s.stats.num_turns || '\u2014');
     html += item('Tool Calls', s.stats.tool_calls);
     html += item('Errors', s.stats.tool_errors + (s.stats.tool_calls > 0 ? ' (' + Math.round(s.stats.tool_errors / s.stats.tool_calls * 100) + '%)' : ''));
+    if (Number(metric.embedded_failure_count || 0) > 0) {
+        html += item('Embedded Failures', Number(metric.embedded_failure_count));
+    }
+    if (metric.sample_validity === 'infra_affected') {
+        html += item('Sample Validity', 'infra affected');
+    }
+    const infraSignals = normalizedInfraSignals(metric);
+    if (infraSignals.length) {
+        html += item('Infra Signal Codes', [...new Set(infraSignals.map(signal => signal.code))].join(', '));
+    }
     if (s.stats.cost_usd != null) html += item('Cost', '$' + s.stats.cost_usd.toFixed(2));
     if (s.stats.loop_groups > 0) html += item('Retry Loops', s.stats.loop_groups);
     if (s.stats.error_streak_groups > 0) html += item('Error Streaks', s.stats.error_streak_groups + ' (max ' + s.stats.max_error_streak + ')');
     html += '</div>';
     html += renderExpectationMiniTable(s);
+    html += renderInfraSignals(metric);
     if (!s.transcript_available) {
         html += '<div class="transcript-unavailable">No transcript was retained for this completed eval. Its scored expectations remain included in this comparison.</div>';
     }

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "deepline",
-  "version": "0.2.47",
+  "version": "0.2.48",
   "description": "Deepline SDK + CLI — B2B data enrichment powered by durable cloud execution",
   "license": "MIT",
   "repository": {