deepline 0.3.56 → 0.3.58

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

@@ -272,7 +272,7 @@ export class HttpClient {
       headers[RUNTIME_SCHEDULER_SCHEMA_OVERRIDE_HEADER] =
         runtimeSchedulerSchema.trim();
     }
-    // Automated test harnesses (e.g. tests/v2-plays) set DEEPLINE_SYNTHETIC_RUN
+    // Automated test harnesses (e.g. tests/gauntlet-e2e-tests) set DEEPLINE_SYNTHETIC_RUN
     // so their intentionally failing plays do not page the SDK CLI error
     // channel. Honored server-side only in non-prod (see plays/run route).
     const syntheticRun =

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

@@ -199,7 +199,7 @@ export const SDK_RELEASE = {
   // available at toolResponse.rawV2 while toolResponse.raw and all declared
   // getters keep their established compatibility behavior.
   // 0.3.1 deprecated the legacy `deepline enrich` command in favor of Plays.
-  version: '0.3.56',
+  version: '0.3.58',
   updateSummary:
     'Automatic CLI updates are now enabled by default. To opt out, run `deepline settings autoupdate off`; use `deepline settings autoupdate on` to re-enable updates or `deepline settings autoupdate pin <version>` to hold an exact release. This release also adds raw-v2 tool responses at toolResponse.rawV2 while preserving existing toolResponse.raw and declared getters.',
   packageCapabilities: {

package/dist/bundling-sources/shared_libs/play-runtime/coordinator-headers.ts

@@ -52,7 +52,7 @@ export const PREFERRED_RECEIPT_GATEWAY_MACHINE_ID_HEADER =
 /**
  * CLI→app marker (NOT a coordinator header — it lives here only because both
  * the SDK HTTP client and the run route already import this module). Set by
- * automated test harnesses (e.g. `tests/v2-plays`) so their intentionally
+ * automated test harnesses (e.g. `tests/gauntlet-e2e-tests`) so their intentionally
  * failing plays — depth-guard probes, error-path scenarios — do not page the
  * SDK CLI error channel as if a real customer run had failed. The run route
  * honors it ONLY in non-prod (see run/route.ts); a forged header from a real

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

@@ -311,11 +311,16 @@ async function createOneShotDaytonaSandbox(input: {
   });
 }
 
+type TimedOutDaytonaSandboxReconciliation =
+  | { kind: 'started'; sandbox: DaytonaSandbox }
+  | { kind: 'deleted' }
+  | { kind: 'not_found' };
+
 async function reconcileAndDeleteTimedOutDaytonaSandbox(input: {
   daytona: DaytonaClient;
   sandboxName: string;
-}): Promise<boolean> {
-  if (!input.daytona.get) return false;
+}): Promise<TimedOutDaytonaSandboxReconciliation> {
+  if (!input.daytona.get) return { kind: 'not_found' };
   let lastDeleteConflict: unknown = null;
   for (const delayMs of DAYTONA_TIMED_OUT_CREATE_RECONCILE_DELAYS_MS) {
     if (delayMs > 0) {
@@ -327,6 +332,14 @@ async function reconcileAndDeleteTimedOutDaytonaSandbox(input: {
     } catch {
       continue;
     }
+    // `create` waits only a bounded interval, but the named sandbox may finish
+    // the same create transition immediately afterwards. It is already the
+    // durable identity associated with this provider operation, so reusing it
+    // is safer and faster than attempting a conflicting delete or creating a
+    // second sandbox.
+    if (sandbox.state === 'started') {
+      return { kind: 'started', sandbox };
+    }
     try {
       await sandbox.delete(30);
     } catch (error) {
@@ -344,7 +357,7 @@ async function reconcileAndDeleteTimedOutDaytonaSandbox(input: {
         { cause: error },
       );
     }
-    return true;
+    return { kind: 'deleted' };
   }
   if (lastDeleteConflict) {
     throw new Error(
@@ -352,7 +365,7 @@ async function reconcileAndDeleteTimedOutDaytonaSandbox(input: {
       { cause: lastDeleteConflict },
     );
   }
-  return false;
+  return { kind: 'not_found' };
 }
 
 async function createRetriedOneShotDaytonaSandbox(input: {
@@ -402,16 +415,14 @@ async function createRetriedOneShotDaytonaSandbox(input: {
       // A normal rejected create has no durable provider resource. A timeout
       // can be ambiguous, so retain its lease until the crash TTL rather than
       // admitting another sandbox against an unknown provider outcome.
-      if (reservation && !isDaytonaSandboxStartTimeout(String(error))) {
-        await input.releaseSandboxCapacity?.(reservation.leaseId);
-      }
-      await input.observeCreateCall?.({
-        type: 'daytona_create_call_failed',
-        occurredAtMs: Date.now(),
-        providerAttempt: attempt,
-        errorClass: daytonaCreateErrorClass(error),
-      });
       const message = error instanceof Error ? error.message : String(error);
+      const recordCreateFailure = () =>
+        input.observeCreateCall?.({
+          type: 'daytona_create_call_failed',
+          occurredAtMs: Date.now(),
+          providerAttempt: attempt,
+          errorClass: daytonaCreateErrorClass(error),
+        });
       errors.push(message);
       input.emitStage('create:attempt_failed', {
         attempt,
@@ -427,18 +438,49 @@ async function createRetriedOneShotDaytonaSandbox(input: {
       });
       const immediateFallbackReason =
         resolveDaytonaSandboxAcquisitionUnavailableReason([message]);
-      if (immediateFallbackReason === 'daytona_total_cpu_limit_exceeded') {
-        throw new DaytonaSandboxAcquisitionUnavailableError(
-          immediateFallbackReason,
-          `Daytona sandbox create rejected by the organization CPU limit on provider attempt ${attempt}: ${message}`,
-        );
-      }
       if (isDaytonaSandboxStartTimeout(message)) {
-        const deleted = await reconcileAndDeleteTimedOutDaytonaSandbox({
-          daytona: input.daytona,
-          sandboxName,
-        });
-        if (deleted) {
+        let reconciliation: TimedOutDaytonaSandboxReconciliation;
+        try {
+          reconciliation = await reconcileAndDeleteTimedOutDaytonaSandbox({
+            daytona: input.daytona,
+            sandboxName,
+          });
+        } catch (reconciliationError) {
+          await recordCreateFailure();
+          throw reconciliationError;
+        }
+        if (reconciliation.kind === 'started') {
+          const acquiredAt = Date.now();
+          try {
+            await input.observeCreateCall?.({
+              type: 'daytona_create_call_succeeded',
+              occurredAtMs: acquiredAt,
+              providerAttempt: attempt,
+              sandboxId: reconciliation.sandbox.id,
+            });
+          } catch {
+            console.error(
+              '[play-runner.daytona.create_lifecycle_event_unrecorded]',
+              {
+                workflowId: input.context.workflowId ?? null,
+                runId: input.context.runId ?? null,
+                attempt,
+                sandboxId: reconciliation.sandbox.id,
+                eventType: 'daytona_create_call_succeeded',
+              },
+            );
+          }
+          return {
+            sandbox: reconciliation.sandbox,
+            attempt,
+            attemptElapsedMs: acquiredAt - attemptStartedAt,
+            ...(reservation
+              ? { sandboxCapacityLeaseId: reservation.leaseId }
+              : {}),
+          };
+        }
+        await recordCreateFailure();
+        if (reconciliation.kind === 'deleted') {
           // The provider-confirmed reconciliation means this reservation can
           // no longer represent a live sandbox. Release before returning the
           // acquisition-unavailable result so Modal fallback is not blocked
@@ -456,6 +498,16 @@ async function createRetriedOneShotDaytonaSandbox(input: {
           { cause: error },
         );
       }
+      if (reservation) {
+        await input.releaseSandboxCapacity?.(reservation.leaseId);
+      }
+      await recordCreateFailure();
+      if (immediateFallbackReason === 'daytona_total_cpu_limit_exceeded') {
+        throw new DaytonaSandboxAcquisitionUnavailableError(
+          immediateFallbackReason,
+          `Daytona sandbox create rejected by the organization CPU limit on provider attempt ${attempt}: ${message}`,
+        );
+      }
       continue;
     }
     // The outcome must be attributed immediately after Daytona acknowledges

package/dist/cli/index.js

@@ -1075,7 +1075,7 @@ var SDK_RELEASE = {
   // available at toolResponse.rawV2 while toolResponse.raw and all declared
   // getters keep their established compatibility behavior.
   // 0.3.1 deprecated the legacy `deepline enrich` command in favor of Plays.
-  version: "0.3.56",
+  version: "0.3.58",
   updateSummary: "Automatic CLI updates are now enabled by default. To opt out, run `deepline settings autoupdate off`; use `deepline settings autoupdate on` to re-enable updates or `deepline settings autoupdate pin <version>` to hold an exact release. This release also adds raw-v2 tool responses at toolResponse.rawV2 while preserving existing toolResponse.raw and declared getters.",
   packageCapabilities: {
     updatePreferences: 1
@@ -23759,10 +23759,16 @@ Concepts:
 Authoring:
   Use normal TypeScript interpolation for human-facing strings:
     const readyMessage = \`Put \${input.title} through to send-ready.\`;
-  An authored diagram is optional. Docflow parses only an explicit
+  An authored diagram is optional but highly recommended for every new or
+  substantially changed Play. Docflow parses only an explicit
   /** @mermaid ... */ block and its // @mermaid-node annotations; ordinary
   prose comments are never Docflow syntax. When Docflow is enabled, plays check
-  gives missing diagrams a non-blocking visualization warning.
+  gives missing diagrams a high-priority, non-blocking warning with the exact
+  Mermaid repair steps. Without a diagram, users see only an inferred execution
+  graph rather than a clear business-language map. It is easy to add: write a
+  short top-of-file flowchart, annotate its visible executable steps, then run
+  plays check again. Treat that warning as an authoring task before finishing
+  Play work unless the caller explicitly asks to leave the inferred graph.
 
 Common commands:
   deepline plays search email --json

package/dist/cli/index.mjs

@@ -1060,7 +1060,7 @@ var SDK_RELEASE = {
   // available at toolResponse.rawV2 while toolResponse.raw and all declared
   // getters keep their established compatibility behavior.
   // 0.3.1 deprecated the legacy `deepline enrich` command in favor of Plays.
-  version: "0.3.56",
+  version: "0.3.58",
   updateSummary: "Automatic CLI updates are now enabled by default. To opt out, run `deepline settings autoupdate off`; use `deepline settings autoupdate on` to re-enable updates or `deepline settings autoupdate pin <version>` to hold an exact release. This release also adds raw-v2 tool responses at toolResponse.rawV2 while preserving existing toolResponse.raw and declared getters.",
   packageCapabilities: {
     updatePreferences: 1
@@ -23821,10 +23821,16 @@ Concepts:
 Authoring:
   Use normal TypeScript interpolation for human-facing strings:
     const readyMessage = \`Put \${input.title} through to send-ready.\`;
-  An authored diagram is optional. Docflow parses only an explicit
+  An authored diagram is optional but highly recommended for every new or
+  substantially changed Play. Docflow parses only an explicit
   /** @mermaid ... */ block and its // @mermaid-node annotations; ordinary
   prose comments are never Docflow syntax. When Docflow is enabled, plays check
-  gives missing diagrams a non-blocking visualization warning.
+  gives missing diagrams a high-priority, non-blocking warning with the exact
+  Mermaid repair steps. Without a diagram, users see only an inferred execution
+  graph rather than a clear business-language map. It is easy to add: write a
+  short top-of-file flowchart, annotate its visible executable steps, then run
+  plays check again. Treat that warning as an authoring task before finishing
+  Play work unless the caller explicitly asks to leave the inferred graph.
 
 Common commands:
   deepline plays search email --json

package/dist/index.js

@@ -826,7 +826,7 @@ var SDK_RELEASE = {
   // available at toolResponse.rawV2 while toolResponse.raw and all declared
   // getters keep their established compatibility behavior.
   // 0.3.1 deprecated the legacy `deepline enrich` command in favor of Plays.
-  version: "0.3.56",
+  version: "0.3.58",
   updateSummary: "Automatic CLI updates are now enabled by default. To opt out, run `deepline settings autoupdate off`; use `deepline settings autoupdate on` to re-enable updates or `deepline settings autoupdate pin <version>` to hold an exact release. This release also adds raw-v2 tool responses at toolResponse.rawV2 while preserving existing toolResponse.raw and declared getters.",
   packageCapabilities: {
     updatePreferences: 1

package/dist/index.mjs

@@ -738,7 +738,7 @@ var SDK_RELEASE = {
   // available at toolResponse.rawV2 while toolResponse.raw and all declared
   // getters keep their established compatibility behavior.
   // 0.3.1 deprecated the legacy `deepline enrich` command in favor of Plays.
-  version: "0.3.56",
+  version: "0.3.58",
   updateSummary: "Automatic CLI updates are now enabled by default. To opt out, run `deepline settings autoupdate off`; use `deepline settings autoupdate on` to re-enable updates or `deepline settings autoupdate pin <version>` to hold an exact release. This release also adds raw-v2 tool responses at toolResponse.rawV2 while preserving existing toolResponse.raw and declared getters.",
   packageCapabilities: {
     updatePreferences: 1

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "deepline",
-  "version": "0.3.56",
+  "version": "0.3.58",
   "description": "GTM data CLI and TypeScript SDK for coding agents",
   "license": "MIT",
   "homepage": "https://code.deepline.com",