engineering-behavior-observatory 0.2.6 → 0.2.7

package/dist/src/cursor-sdk-runner.js

@@ -70,7 +70,7 @@ export async function runCursorSdkQueueEntry(options) {
             sandbox: { enabled: toolPolicy.sandbox.enabled },
             settingSources: [],
             autoReview: false,
-            enableAgentRetries: false,
+            enableAgentRetries: toolPolicy.enableAgentRetries,
         },
         maxNativeRecordBytes: limits.maxNativeRecordBytes,
     };
@@ -215,8 +215,8 @@ function validateConfigurationRecord(record, kind, reference) {
             if (!Array.isArray(record.settingSources) || record.settingSources.length !== 0) {
                 throw configurationError(reference, "settingSources must be [] so unrelated user/candidate settings are not loaded");
             }
-            if (record.autoReview !== false || record.enableAgentRetries !== false) {
-                throw configurationError(reference, "autoReview and enableAgentRetries must both be false for an observational attempt");
+            if (record.autoReview !== false || record.enableAgentRetries !== true) {
+                throw configurationError(reference, "autoReview must be false and enableAgentRetries must be true for native recovery");
             }
             return;
         case "capture-profile":

package/dist/src/cursor-sdk.d.ts

@@ -18,7 +18,7 @@ export type CursorSdkToolPolicy = {
     };
     settingSources: readonly [];
     autoReview: false;
-    enableAgentRetries: false;
+    enableAgentRetries: true;
 };
 export type CursorSdkCaptureConfiguration = {
     model: ModelSelection;

package/dist/src/cursor-sdk.js

@@ -2,7 +2,7 @@ import { createHash } from "node:crypto";
 import { existsSync } from "node:fs";
 import { lstat, mkdir, open, stat } from "node:fs/promises";
 import { dirname, join, parse, resolve } from "node:path";
-import { Agent, JSONL_LOCAL_AGENT_STORE_FILES, JsonlLocalAgentStore, } from "@cursor/sdk";
+import { Agent, Cursor, JSONL_LOCAL_AGENT_STORE_FILES, JsonlLocalAgentStore, } from "@cursor/sdk";
 import { assertNoDuplicateJsonKeys, digestMetadata } from "./artifacts.js";
 import { createAttemptIdentity, createRunIdentity, executeRunAttempt, } from "./lifecycle.js";
 import { createCapturedNativeEvidenceResolver, describeNormalizedDataset, validateNormalizedDataset, } from "./normalization-integrity.js";
@@ -56,6 +56,9 @@ export const CURSOR_SDK_CAPABILITIES = {
 export async function captureCursorSdkRun(options) {
     requireText(options.apiKey, "Cursor API key");
     requireText(options.prompt, "Cursor prompt");
+    if (options.configuration.toolPolicy.enableAgentRetries !== true) {
+        throw new Error("Cursor capture requires enableAgentRetries: true for native recovery.");
+    }
     if (options.definition.run.model.id !== options.configuration.model.id) {
         throw new Error("The declared model must match the Cursor SDK model configuration.");
     }
@@ -283,11 +286,13 @@ async function executeCursorSdk(options) {
     });
     try {
         await assertCursorWorkspaceIsolation(options.workspacePath);
+        Cursor.configure({ local: { useHttp1ForAgent: true } });
         await options.writer.record("configuration", {
             sdkVersion: CURSOR_SDK_VERSION,
             runtimeIdentity: { status: "not-exposed", detail: "The public SDK does not expose a separately versioned local agent runtime." },
             model: options.configuration.model,
             toolPolicy: options.configuration.toolPolicy,
+            transport: { useHttp1ForAgent: true },
             workingDirectory: options.workspacePath,
             workingDirectoryPolicy: "isolated-local.cwd",
             environmentKeys: environment.keys,
@@ -306,7 +311,7 @@ async function executeCursorSdk(options) {
                 settingSources: [],
                 sandboxOptions: { enabled: options.configuration.toolPolicy.sandbox.enabled },
                 autoReview: false,
-                enableAgentRetries: false,
+                enableAgentRetries: options.configuration.toolPolicy.enableAgentRetries,
             },
         });
         agent = await abortable(agentPromise, options.signal, "Cursor agent creation", async (lateAgent) => {

package/docs/harnesses/cursor-sdk.md

@@ -38,7 +38,21 @@ the following digest-pinned settings for every Cursor queue entry:
   MDM, or plugin settings are not silently loaded;
 - `autoReview: false`, so instrumentation does not add a behavior-changing
   classifier; and
-- `enableAgentRetries: false`, so a retained attempt is not silently replaced.
+- `enableAgentRetries: true`, enabling Cursor's native transport/stall recovery
+  within the owned run. EBO does not add a retry loop or replacement attempt.
+
+Before agent creation, EBO calls
+`Cursor.configure({ local: { useHttp1ForAgent: true } })` to select HTTP/1.1
+for local backend streams. This mitigates the observed `NGHTTP2_INTERNAL_ERROR`;
+it is not a proven root-cause fix. Both settings use the
+[public SDK API](https://cursor.com/docs/sdk/typescript).
+The native `configuration` record retains `transport.useHttp1ForAgent` and
+`toolPolicy.enableAgentRetries`. Terminal errors and partial evidence remain
+retained if native recovery fails; the attempt deadline still applies.
+
+Existing frozen policies with retries disabled are rejected at launch, not
+silently overridden. Prepare new configuration records and a new queue for
+future runs; preserve earlier frozen experiments and captures unchanged.
 
 SDK `1.0.31` treats an ancestor Git checkout as the local project root. EBO
 therefore rejects a materialized workspace nested under another `.git` path
@@ -79,7 +93,7 @@ provider ID containing dots.
 ```
 
 ```json
-{ "schemaVersion": "ebo.cursor-sdk-config/v1", "kind": "native-tool-policy", "tools": ["read", "edit", "grep", "glob", "ls"], "disallowedTools": ["shell", "task", "mcp", "webSearch", "webFetch"], "sandbox": { "enabled": false }, "settingSources": [], "autoReview": false, "enableAgentRetries": false }
+{ "schemaVersion": "ebo.cursor-sdk-config/v1", "kind": "native-tool-policy", "tools": ["read", "edit", "grep", "glob", "ls"], "disallowedTools": ["shell", "task", "mcp", "webSearch", "webFetch"], "sandbox": { "enabled": false }, "settingSources": [], "autoReview": false, "enableAgentRetries": true }
 ```
 
 ```json

package/examples/cursor-sdk/native-tool-policy.json

@@ -1 +1 @@
-{"schemaVersion":"ebo.cursor-sdk-config/v1","kind":"native-tool-policy","tools":["read","edit","grep","glob","ls"],"disallowedTools":["shell","task","mcp","webSearch","webFetch"],"sandbox":{"enabled":false},"settingSources":[],"autoReview":false,"enableAgentRetries":false}
+{"schemaVersion":"ebo.cursor-sdk-config/v1","kind":"native-tool-policy","tools":["read","edit","grep","glob","ls"],"disallowedTools":["shell","task","mcp","webSearch","webFetch"],"sandbox":{"enabled":false},"settingSources":[],"autoReview":false,"enableAgentRetries":true}

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "engineering-behavior-observatory",
-  "version": "0.2.6",
+  "version": "0.2.7",
   "description": "Capture engineering-agent trajectories, evaluate behavior, and inspect cited evidence across harnesses.",
   "license": "Apache-2.0",
   "repository": {
@@ -50,6 +50,7 @@
     "release/0.2.4/",
     "release/0.2.5/",
     "release/0.2.6/",
+    "release/0.2.7/",
     "schemas/",
     "scripts/atlas-grafana.sh"
   ],

package/release/0.2.7/KNOWN_LIMITATIONS.md

@@ -0,0 +1,8 @@
+# v0.2.7 known limitations
+
+The [v0.2.6 limitations](../0.2.6/KNOWN_LIMITATIONS.md) continue to apply.
+
+- HTTP/1.1 mitigates the observed HTTP/2 failure; it is not a proven root-cause fix.
+- Native recovery can still fail or reach the attempt deadline.
+- Deterministic tests verify configuration and evidence handling, not live
+  provider recovery. No new full trial was run for this patch.

package/release/0.2.7/README.md

@@ -0,0 +1,17 @@
+# v0.2.7: Cursor native recovery
+
+Cursor capture configures HTTP/1.1 before agent creation and enables the SDK's
+native transport/stall retries. The configuration evidence records both settings.
+EBO does not launch replacement attempts or add its own retry loop. Terminal
+errors and partial evidence remain available when native recovery fails.
+
+New queue policies must explicitly set `enableAgentRetries: true`. Earlier
+frozen policies are rejected at launch rather than silently overridden; prior
+experiments and retained captures remain unchanged.
+
+Regression tests cover configuration ordering, effective evidence, policy
+validation, and terminal-error retention with exactly one agent creation.
+
+See the [Cursor guide](../../docs/harnesses/cursor-sdk.md),
+[known limitations](KNOWN_LIMITATIONS.md), and
+[reproducibility manifest](reproducibility.json).

package/release/0.2.7/reproducibility.json

@@ -0,0 +1,74 @@
+{
+  "schemaVersion": "ebo.release-reproducibility/v1",
+  "release": {
+    "name": "engineering-behavior-observatory",
+    "version": "0.2.7"
+  },
+  "runtime": {
+    "node": "24.19.0",
+    "claudeAgentSdk": "0.3.258",
+    "openhandsAgentServer": "1.46.0",
+    "deepseekClient": "0.1.1-rc.2",
+    "deepseekProtocol": "0.1.1-rc.2",
+    "deepseekRuntime": "0.1.1-rc.2",
+    "codexAppServer": "0.153.4",
+    "grafana": "13.2.0",
+    "grafanaInfinity": "4.0.0",
+    "piSdk": "0.85.1",
+    "cursorSdk": "1.0.31"
+  },
+  "commands": [
+    "npm ci",
+    "npm run acceptance"
+  ],
+  "fixtureCoverage": {
+    "agent-sdk": "frozen queue entry through qualified capture, approved export, normalization, configurable judging, review, aggregation, and Atlas",
+    "openhands-agent-server": "pinned REST/WebSocket stream-final reconciliation, workspace/outcome evidence, partial capture, and retained evaluation",
+    "deepseek-harness": "official-client JSON-RPC composition, receipt-to-idle completion, stderr, interruption, shutdown, swaps, and retained evaluation",
+    "codex-app-server": "owned stdio lifecycle, full history, OTLP, interruption/failure, export, and retained evaluation; configurable workspace-write networking with explicit offline queue propagation",
+    "acceptance-cases": "seeded ordering, secret scanning, partial attempts, native references, abstention, disputes, denominators, and unsupported comparisons",
+    "pi-sdk": "frozen queue, passive hooks, durable session, partial cleanup, export, retained normalization and evaluation",
+    "cursor-sdk": "frozen queue, official store, callbacks/history, bounded cancellation, export, retained normalization and evaluation; HTTP/1.1 configuration ordering, native retry validation, terminal transport error retention without replacement",
+    "workspace-capture": "Contained absolute links relocated in derived captures; workspace packaging gaps preserve native completion and allow semantic analysis independently"
+  },
+  "determinism": {
+    "stable": [
+      "task, run, attempt, event, assertion, aggregate, and Atlas case identities",
+      "fixture, native-reference, normalized-dataset, source/cohort, archive, and package digests",
+      "seeded queue and review-sample ordering"
+    ],
+    "excluded": [
+      "native and EBO observation timestamps",
+      "lifecycle start/finish timestamps",
+      "provider timing and usage",
+      "temporary workspace and output paths"
+    ]
+  },
+  "fixtures": {
+    "tests/fixtures/task-packet.valid.v1.json": "ac2ef1043c0cc16bf4c21d05c1dc880ca1b4b8cc2f9ea7653328e91f0b0e2681",
+    "test/fixtures/agent-sdk-normalizer/complete.input.json": "7285dee088517e5de38ceda84fc7c58d9303949c7448ab3b5e10a5a2d2ebdfe3",
+    "test/fixtures/agent-sdk-normalizer/complete.expected.jsonl": "d32fa3a013bc179f1673f7e1a89b8a721bb8ddaf63eba5bc8a3f14f1cc148ed8",
+    "contracts/openhands-agent-server-v1.44.1.json": "e7a07977688a0703b15751a8a1ab17f29be45f9fdf438017e0dbcbc49cca37b0",
+    "contracts/openhands-agent-server-v1.46.0.json": "455cdddd2c206cb2f20245c2189b92e773657ec14d651236baa532b2dbadeb28",
+    "test/fixtures/openhands/v1.44.1/streamed-events.json": "d751083ba2ef94c9865117db481cc2f4933772bbbbb6b9fcf60571563715c1ea",
+    "test/fixtures/openhands/v1.44.1/final-events.json": "551db9904e860004d604c923692b5c5d3aaee966220492214c60003d2df415b6",
+    "test/fixtures/deepseek/golden-success.json": "f3867d38169bf0e28f64aba1dd60553f21cf07f41f81b4bad48261e49fa3ca88",
+    "test/fixtures/deepseek/golden-interrupted.json": "79e1f8404df1ee3b562d0bdbc0b2ee712ffc6d5774992b92970814ea815677b7",
+    "test/fixtures/deepseek/compositions/minimal/composition.json": "72fe4368e105f1bad82b7fdf8380e92c9fa689d8985f6ab91f534f0baf3eda02",
+    "contracts/codex-app-server-0.153.4/manifest.json": "e62281ff5e5d1cc71d07763b85e997ccc9ddbad9aa82a469da4181ee4b1890a8",
+    "test/fixtures/codex/legacy-0.150.1.dataset.json": "f71906b009e546afbf3b4e79396f223181382f3b3d8795e9cfcd032398a9d6bf",
+    "test/fixtures/behavior-assertions/abstained.json": "7f1962864f664c794925a67d93074e7b150af0f57c58041a81e062990d641197",
+    "test/fixtures/behavior-assertions/disputed.review.json": "450d3d9efe7feefaf7500b3ad20ab2803b412c7e53b91763125e15a83be87981",
+    "test/fixtures/comparison/exact.json": "cafa68e94b8d50dc308ddcfc2e73aa139c828b4ecd1e1c762bbc51bf1b44501e",
+    "test/fixtures/structural-observations/golden.json": "bad71a3a93524b76c0a2fdbb58e98e1e65be46cae14acf18ac89d9ba76e03e8a",
+    "test/atlas-fixture.ts": "9f03ce9cbcd0e5d3e3a5916866554ceb1c21f1c8712e58529499c1ea499f048d",
+    "test/pi.test.ts": "a00997c5617ca178e19b25b0d483cbdce8ce7c306d4c871557945b31a2e657cf",
+    "test/cursor-sdk.test.ts": "ceb8aeb719b9554d51bb3cfe78fd0bc796acdce8ddc5ef205a327fac4226f57e",
+    "examples/cursor-sdk/README.md": "a2cf481123e7f3deb1c5a682ba208a775c3f91c35e1a6c1a428c517c79692618",
+    "examples/cursor-sdk/capture-profile.json": "d1a5787875daf5523b3783f259b891c256a904cbb10279d5be214af846305a2d",
+    "examples/cursor-sdk/harness.json": "2f1da6f2b17258d0ee59431f687696901ce3692589175d182f3a4bbe8dd4d689",
+    "examples/cursor-sdk/model.json": "ad8f31ff5d2387c02d79226679022358afafc61ba5dda73040e3f754530a230a",
+    "examples/cursor-sdk/native-limits.json": "40fe451f56b0863e79b42f5c85defbce678e1dc90dd248d15c430db176d71511",
+    "examples/cursor-sdk/native-tool-policy.json": "8148b9ad57134d217a03e47cc11bcc85758e1115b64bb06b85564139aecd1cd9"
+  }
+}

package/release/README.md

@@ -5,6 +5,7 @@ For downloads and published artifacts, see
 
 | Version | Changes and verification | Support boundary |
 | :--- | :--- | :--- |
+| [0.2.7](0.2.7/README.md) | Cursor native transport/stall recovery and recorded HTTP/1.1 settings | [Known limitations](0.2.7/KNOWN_LIMITATIONS.md) |
 | [0.2.6](0.2.6/README.md) | Portable contained absolute links; isolate workspace packaging failures from native completion and semantic analysis | [Known limitations](0.2.6/KNOWN_LIMITATIONS.md) |
 | [0.2.5](0.2.5/README.md) | Preserve contained workspace symlinks and macOS AppleDouble files | [Known limitations](0.2.5/KNOWN_LIMITATIONS.md) |
 | [0.2.4](0.2.4/README.md) | Stream workspace snapshots and integrity checks; support archives up to 1 GiB | [Known limitations](0.2.4/KNOWN_LIMITATIONS.md) |