@stigmer/runner 3.15.0 → 3.15.2-dev.20260913223433

package/dist/activities/execute-deep-agent/__test-utils__/hermetic-deep-agent.js

@@ -0,0 +1,225 @@
+/**
+ * The native (deep-agent) slice of the hermetic activity driver: build the
+ * fixture record, bind the scripted model and the `Config`, run the REAL
+ * `ExecuteDeepAgent` activity under `MockActivityEnvironment` with the REAL
+ * deepagents graph, and hand back what it persisted.
+ *
+ * Layering (the `approval-contract/` + `gateway-substrate.ts` split): everything
+ * harness-agnostic lives in `src/__test-utils__/hermetic-activity.ts` (the
+ * activity environment, the execution record behind the client, the temp
+ * `~/.stigmer`, the ticking clock, the shutdown signal) and its fixtures
+ * (`execution-record-fixture.ts`, `model-registry-fixture.ts`,
+ * `git-workspace-fixture.ts`). This module adds only what is native's: the
+ * scripted model as the one LLM (`scripted-model.ts`, reached through the
+ * `shared/model-client.ts` double), the native `Config` slice with its
+ * checkpointer, the per-scenario reset of module caches and of the session's
+ * durable state, and the production wire shape of the activity input.
+ *
+ * What the two `vi.mock` calls a scenario file must carry look like — they are
+ * hoisted by vitest and so cannot live here:
+ *
+ * ```ts
+ * vi.mock("../../../../shared/model-client.js", async () =>
+ *   (await import("../../__test-utils__/scripted-model-module.js")).scriptedModelClientModule());
+ * vi.mock("../../../../client/stigmer-client.js", async () =>
+ *   (await import("../../../../__test-utils__/hermetic-activity.js")).hermeticStigmerClientModule());
+ * ```
+ *
+ * Two facts about this harness the driver encodes, both read from production:
+ *
+ *  - `thread_id` is `thread-{sessionId}` on EVERY invocation, the first
+ *    included (`invoke-agent-execution.ts` runs `EnsureThread` once and passes
+ *    the same id to `firstInvoke` and every `reinvoke`; `ensure-thread.ts`
+ *    derives it from the session). Cursor's driver sends `""` on a first turn
+ *    because Cursor's engine mints its own id; copying that here would key the
+ *    sqlite checkpoint on `""` in turn 1 and look it up under `thread-…` in
+ *    turn 2. A scenario overrides the default only to stage the empty-id edge.
+ *  - The `memory` checkpointer is recreated empty per invocation
+ *    (`checkpointer/factory.ts`), so a HITL reinvocation REPLAYS the graph; the
+ *    `sqlite` checkpointer keys one file per session under `~/.stigmer`, so a
+ *    reinvocation RESUMES from the interrupt — the durable posture with no
+ *    network. A scenario picks one; the HITL arms record both.
+ *
+ * The ONE knowledge of production this driver holds is `activityFactory`: how
+ * to obtain the `ExecuteDeepAgent` function from a `Config`. Since S3 M2a it
+ * is the harness registry's: the native adapter booted and bound to its
+ * byte-pinned activity name through `createHarnessActivities`, exactly the
+ * way the composition roots bind it since S3 M2b through `harness-adapters.ts`
+ * (and the way the Cursor driver binds its own, `hermetic-cursor.ts`). The
+ * orchestrator this adapter replaced (`index.ts`, `setup.ts`) was deleted at
+ * M2b; the goldens moved once, with the M2a flip, each hunk quoted with its
+ * ruling in its test's header (Q-M2a-1), and the switch of the roots moved
+ * none (Q-M2b-4).
+ *
+ * `boot` registers the deepagents harness profiles as production does
+ * (S3 M2a F-M2a-14: the orchestrator-driven runs never did, so deepagents
+ * auto-injected its ungated `general-purpose` sub-agent into every M0
+ * scenario; the scripted model never delegated to it).
+ *
+ * Determinism beyond S0's: the clock ticks once per model turn (the scripted
+ * model's `onTurn`), so status timestamps and the persist cadence land on the
+ * same instants every run; a scenario's interruption (pause, worker shutdown)
+ * is scheduled on a model turn through the same hook, never on a timer.
+ */
+import { rmSync } from "node:fs";
+import { dirname, join } from "node:path";
+import { getCheckpointDbPath } from "../../../shared/workspace/platform-dir.js";
+import { _resetRegistryCache } from "../../../shared/model-registry.js";
+import { _resetPricingCache } from "../../../shared/model-pricing-data.js";
+import { bindHermeticClient, runActivityHermetically, } from "../../../__test-utils__/hermetic-activity.js";
+import { executionRecordFixture } from "../../../__test-utils__/execution-record-fixture.js";
+import { FIXTURE_NATIVE_MODEL } from "../../../__test-utils__/model-registry-fixture.js";
+import { ScriptedModel } from "./scripted-model.js";
+import { bindScriptedModel } from "./scripted-model-module.js";
+// ---------------------------------------------------------------------------
+// Fixture ids — fixed so goldens are readable and byte-stable
+// ---------------------------------------------------------------------------
+const SESSION_ID = "ses_hermetic_0001";
+export const FIXTURE = {
+    org: "hermetic-org",
+    executionId: "aex_hermetic_0001",
+    sessionId: SESSION_ID,
+    agentInstanceId: "ain_hermetic_0001",
+    agentId: "agt_hermetic_0001",
+    agentName: "hermetic-agent",
+    /** The pinned model; in the registry document as the native entry. */
+    model: FIXTURE_NATIVE_MODEL,
+    /** The LangGraph thread id the control plane sends on every invocation (`ensure-thread.ts`). */
+    threadId: `thread-${SESSION_ID}`,
+};
+/** The four resources of one execution, wired by id into a chain, under the native fixture ids. */
+export function deepAgentExecutionRecord(options) {
+    return executionRecordFixture({
+        ...options,
+        modelName: options.modelName ?? FIXTURE.model,
+        ids: {
+            org: FIXTURE.org,
+            executionId: FIXTURE.executionId,
+            sessionId: FIXTURE.sessionId,
+            agentInstanceId: FIXTURE.agentInstanceId,
+            agentId: FIXTURE.agentId,
+            agentName: FIXTURE.agentName,
+        },
+    });
+}
+/** The OSS/local runner posture: no proxy, local mode, the scenario's checkpointer. */
+export function hermeticDeepAgentConfig(env, checkpointerType) {
+    return {
+        taskQueue: "hermetic-test-queue",
+        temporalAddress: "localhost:7233",
+        temporalNamespace: "default",
+        stigmerBackendEndpoint: "http://localhost:7234",
+        mcpBridgeEndpoint: null,
+        stigmerToken: null,
+        cursorApiKey: "",
+        workspaceRootDir: env.workspaceRootDir,
+        mode: "local",
+        proxyEndpoint: null,
+        maxConcurrentActivities: 1,
+        idleTimeoutSeconds: null,
+        cloudModeEnabled: false,
+        checkpointerType,
+        checkpointerProxyEndpoint: null,
+        artifactProxyEndpoint: null,
+        primaryModel: FIXTURE.model,
+        cursorStreamStallTimeoutMs: 180_000,
+        agentResolveTimeoutMs: 120_000,
+        workspaceLockTimeoutMs: 900_000,
+    };
+}
+/**
+ * The primary workspace directory a session with NO workspace entries runs in
+ * (`session-root.ts`: `<workspaceRootDir>/sessions/<sessionId>`). A scenario
+ * seeds the files its script reads here.
+ */
+export function sessionWorkspaceDir(env) {
+    return join(env.workspaceRootDir, "sessions", FIXTURE.sessionId);
+}
+/**
+ * The adapter through the registry: `createDeepAgentAdapter()` booted on the
+ * scenario's config, bound to `ExecuteDeepAgent` by `createHarnessActivities`
+ * over the turn runtime. Imported lazily so the scenario file's `vi.mock`
+ * declarations are in force before the engine slice loads the model client
+ * and before the registry loads the `StigmerClient` module.
+ */
+export const runtimeActivityFactory = async (config) => {
+    const { createDeepAgentAdapter } = await import("../adapter.js");
+    const { createHarnessActivities } = await import("../../../harness/registry.js");
+    const adapter = createDeepAgentAdapter();
+    await adapter.boot(config);
+    const activities = await createHarnessActivities([{ harness: "deep-agent", adapter }], config);
+    return activities.ExecuteDeepAgent;
+};
+/**
+ * Drop every module-level cache the harness keeps across invocations and
+ * every durable trace the session left behind (its sqlite checkpoint, its
+ * session workspace), so a scenario starts where a new session starts.
+ * Called at the START of a scenario, never between its invocations: a HITL
+ * scenario relies on the sqlite checkpoint surviving from turn 1 into turn 2
+ * exactly as it does in a live runner.
+ */
+export function resetDeepAgentScenarioState(env) {
+    _resetRegistryCache();
+    _resetPricingCache();
+    rmSync(dirname(getCheckpointDbPath(FIXTURE.sessionId)), { recursive: true, force: true });
+    rmSync(sessionWorkspaceDir(env), { recursive: true, force: true });
+}
+/** One scenario: the record, the one LLM, the runner it runs on, and the invocation in flight. */
+export class DeepAgentScenario {
+    env;
+    clock;
+    record;
+    config;
+    /** The controls of the invocation in flight; set by {@link runDeepAgentTurn} before the activity starts. */
+    controls;
+    model;
+    constructor(env, clock, record, 
+    /** The runner this scenario runs on: the OSS posture plus the scenario's knobs. */
+    config, script, onTurn) {
+        this.env = env;
+        this.clock = clock;
+        this.record = record;
+        this.config = config;
+        this.model = new ScriptedModel(script, [], {
+            onTurn: async (info) => {
+                this.clock.tick();
+                if (onTurn && this.controls)
+                    await onTurn(info, this.controls);
+            },
+        });
+    }
+}
+/** Bind the record and the model for a scenario; resets the harness's module and session state first. */
+export function beginDeepAgentScenario(options) {
+    resetDeepAgentScenarioState(options.env);
+    const scenario = new DeepAgentScenario(options.env, options.clock, options.record, { ...hermeticDeepAgentConfig(options.env, options.checkpointer ?? "memory"), ...options.config }, options.script, options.onTurn);
+    bindScriptedModel(scenario.model);
+    bindHermeticClient(options.record.client(options.clientOverrides));
+    return scenario;
+}
+/**
+ * Run ONE `ExecuteDeepAgent` invocation for the scenario. Built per invocation
+ * (the factory constructs its client, which is the one bound at
+ * `beginDeepAgentScenario`).
+ */
+export async function runDeepAgentTurn(scenario, options = {}) {
+    const activity = await (options.activityFactory ?? runtimeActivityFactory)(scenario.config);
+    const threadId = options.threadId ?? FIXTURE.threadId;
+    const turnSeq = options.turnSeq ?? 0;
+    const input = {
+        execution_id: scenario.record.executionId,
+        thread_id: threadId,
+        turn_seq: turnSeq,
+    };
+    const args = options.positional
+        ? [scenario.record.executionId, threadId]
+        : [input];
+    return runActivityHermetically(activity, args, {
+        taskQueue: scenario.config.taskQueue,
+        onControls: (controls) => {
+            scenario.controls = controls;
+            options.onControls?.(controls);
+        },
+    });
+}
+//# sourceMappingURL=hermetic-deep-agent.js.map

package/dist/activities/execute-deep-agent/__test-utils__/hermetic-deep-agent.js.map

@@ -0,0 +1 @@
+{"version":3,"file":"hermetic-deep-agent.js","sourceRoot":"","sources":["../../../../src/activities/execute-deep-agent/__test-utils__/hermetic-deep-agent.ts"],"names":[],"mappings":"AAAA;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;GA8DG;AAEH,OAAO,EAAE,MAAM,EAAE,MAAM,SAAS,CAAC;AACjC,OAAO,EAAE,OAAO,EAAE,IAAI,EAAE,MAAM,WAAW,CAAC;AAI1C,OAAO,EAAE,mBAAmB,EAAE,MAAM,2CAA2C,CAAC;AAChF,OAAO,EAAE,mBAAmB,EAAE,MAAM,mCAAmC,CAAC;AACxE,OAAO,EAAE,kBAAkB,EAAE,MAAM,uCAAuC,CAAC;AAC3E,OAAO,EAGL,kBAAkB,EAClB,uBAAuB,GAIxB,MAAM,8CAA8C,CAAC;AACtD,OAAO,EAAE,sBAAsB,EAA+B,MAAM,qDAAqD,CAAC;AAC1H,OAAO,EAAE,oBAAoB,EAAE,MAAM,mDAAmD,CAAC;AACzF,OAAO,EAAE,aAAa,EAA8C,MAAM,qBAAqB,CAAC;AAChG,OAAO,EAAE,iBAAiB,EAAE,MAAM,4BAA4B,CAAC;AAE/D,8EAA8E;AAC9E,8DAA8D;AAC9D,8EAA8E;AAE9E,MAAM,UAAU,GAAG,mBAAmB,CAAC;AAEvC,MAAM,CAAC,MAAM,OAAO,GAAG;IACrB,GAAG,EAAE,cAAc;IACnB,WAAW,EAAE,mBAAmB;IAChC,SAAS,EAAE,UAAU;IACrB,eAAe,EAAE,mBAAmB;IACpC,OAAO,EAAE,mBAAmB;IAC5B,SAAS,EAAE,gBAAgB;IAC3B,sEAAsE;IACtE,KAAK,EAAE,oBAAoB;IAC3B,gGAAgG;IAChG,QAAQ,EAAE,UAAU,UAAU,EAAE;CACxB,CAAC;AASX,mGAAmG;AACnG,MAAM,UAAU,wBAAwB,CAAC,OAA+B;IACtE,OAAO,sBAAsB,CAAC;QAC5B,GAAG,OAAO;QACV,SAAS,EAAE,OAAO,CAAC,SAAS,IAAI,OAAO,CAAC,KAAK;QAC7C,GAAG,EAAE;YACH,GAAG,EAAE,OAAO,CAAC,GAAG;YAChB,WAAW,EAAE,OAAO,CAAC,WAAW;YAChC,SAAS,EAAE,OAAO,CAAC,SAAS;YAC5B,eAAe,EAAE,OAAO,CAAC,eAAe;YACxC,OAAO,EAAE,OAAO,CAAC,OAAO;YACxB,SAAS,EAAE,OAAO,CAAC,SAAS;SAC7B;KACF,CAAC,CAAC;AACL,CAAC;AAQD,uFAAuF;AACvF,MAAM,UAAU,uBAAuB,CAAC,GAAwB,EAAE,gBAAsC;IACtG,OAAO;QACL,SAAS,EAAE,qBAAqB;QAChC,eAAe,EAAE,gBAAgB;QACjC,iBAAiB,EAAE,SAAS;QAC5B,sBAAsB,EAAE,uBAAuB;QAC/C,iBAAiB,EAAE,IAAI;QACvB,YAAY,EAAE,IAAI;QAClB,YAAY,EAAE,EAAE;QAChB,gBAAgB,EAAE,GAAG,CAAC,gBAAgB;QACtC,IAAI,EAAE,OAAO;QACb,aAAa,EAAE,IAAI;QACnB,uBAAuB,EAAE,CAAC;QAC1B,kBAAkB,EAAE,IAAI;QACxB,gBAAgB,EAAE,KAAK;QACvB,gBAAgB;QAChB,yBAAyB,EAAE,IAAI;QAC/B,qBAAqB,EAAE,IAAI;QAC3B,YAAY,EAAE,OAAO,CAAC,KAAK;QAC3B,0BAA0B,EAAE,OAAO;QACnC,qBAAqB,EAAE,OAAO;QAC9B,sBAAsB,EAAE,OAAO;KAChC,CAAC;AACJ,CAAC;AAED;;;;GAIG;AACH,MAAM,UAAU,mBAAmB,CAAC,GAAwB;IAC1D,OAAO,IAAI,CAAC,GAAG,CAAC,gBAAgB,EAAE,UAAU,EAAE,OAAO,CAAC,SAAS,CAAC,CAAC;AACnE,CAAC;AAYD;;;;;;GAMG;AACH,MAAM,CAAC,MAAM,sBAAsB,GAAoB,KAAK,EAAE,MAAM,EAAE,EAAE;IACtE,MAAM,EAAE,sBAAsB,EAAE,GAAG,MAAM,MAAM,CAAC,eAAe,CAAC,CAAC;IACjE,MAAM,EAAE,uBAAuB,EAAE,GAAG,MAAM,MAAM,CAAC,8BAA8B,CAAC,CAAC;IACjF,MAAM,OAAO,GAAG,sBAAsB,EAAE,CAAC;IACzC,MAAM,OAAO,CAAC,IAAI,CAAC,MAAM,CAAC,CAAC;IAC3B,MAAM,UAAU,GAAG,MAAM,uBAAuB,CAAC,CAAC,EAAE,OAAO,EAAE,YAAY,EAAE,OAAO,EAAE,CAAC,EAAE,MAAM,CAAC,CAAC;IAC/F,OAAO,UAAU,CAAC,gBAAiB,CAAC;AACtC,CAAC,CAAC;AAEF;;;;;;;GAOG;AACH,MAAM,UAAU,2BAA2B,CAAC,GAAwB;IAClE,mBAAmB,EAAE,CAAC;IACtB,kBAAkB,EAAE,CAAC;IACrB,MAAM,CAAC,OAAO,CAAC,mBAAmB,CAAC,OAAO,CAAC,SAAS,CAAC,CAAC,EAAE,EAAE,SAAS,EAAE,IAAI,EAAE,KAAK,EAAE,IAAI,EAAE,CAAC,CAAC;IAC1F,MAAM,CAAC,mBAAmB,CAAC,GAAG,CAAC,EAAE,EAAE,SAAS,EAAE,IAAI,EAAE,KAAK,EAAE,IAAI,EAAE,CAAC,CAAC;AACrE,CAAC;AAED,kGAAkG;AAClG,MAAM,OAAO,iBAAiB;IAMjB;IACA;IACA;IAEA;IATX,4GAA4G;IAC5G,QAAQ,CAAiC;IAChC,KAAK,CAAgB;IAE9B,YACW,GAAwB,EACxB,KAAoB,EACpB,MAAuB;IAChC,mFAAmF;IAC1E,MAAc,EACvB,MAAsB,EACtB,MAA0C;QANjC,QAAG,GAAH,GAAG,CAAqB;QACxB,UAAK,GAAL,KAAK,CAAe;QACpB,WAAM,GAAN,MAAM,CAAiB;QAEvB,WAAM,GAAN,MAAM,CAAQ;QAIvB,IAAI,CAAC,KAAK,GAAG,IAAI,aAAa,CAAC,MAAM,EAAE,EAAE,EAAE;YACzC,MAAM,EAAE,KAAK,EAAE,IAAI,EAAE,EAAE;gBACrB,IAAI,CAAC,KAAK,CAAC,IAAI,EAAE,CAAC;gBAClB,IAAI,MAAM,IAAI,IAAI,CAAC,QAAQ;oBAAE,MAAM,MAAM,CAAC,IAAI,EAAE,IAAI,CAAC,QAAQ,CAAC,CAAC;YACjE,CAAC;SACF,CAAC,CAAC;IACL,CAAC;CACF;AA0BD,yGAAyG;AACzG,MAAM,UAAU,sBAAsB,CAAC,OAAiC;IACtE,2BAA2B,CAAC,OAAO,CAAC,GAAG,CAAC,CAAC;IACzC,MAAM,QAAQ,GAAG,IAAI,iBAAiB,CACpC,OAAO,CAAC,GAAG,EACX,OAAO,CAAC,KAAK,EACb,OAAO,CAAC,MAAM,EACd,EAAE,GAAG,uBAAuB,CAAC,OAAO,CAAC,GAAG,EAAE,OAAO,CAAC,YAAY,IAAI,QAAQ,CAAC,EAAE,GAAG,OAAO,CAAC,MAAM,EAAE,EAChG,OAAO,CAAC,MAAM,EACd,OAAO,CAAC,MAAM,CACf,CAAC;IACF,iBAAiB,CAAC,QAAQ,CAAC,KAAK,CAAC,CAAC;IAClC,kBAAkB,CAAC,OAAO,CAAC,MAAM,CAAC,MAAM,CAAC,OAAO,CAAC,eAAe,CAAC,CAAC,CAAC;IACnE,OAAO,QAAQ,CAAC;AAClB,CAAC;AAaD;;;;GAIG;AACH,MAAM,CAAC,KAAK,UAAU,gBAAgB,CACpC,QAA2B,EAC3B,UAAgC,EAAE;IAElC,MAAM,QAAQ,GAAG,MAAM,CAAC,OAAO,CAAC,eAAe,IAAI,sBAAsB,CAAC,CAAC,QAAQ,CAAC,MAAM,CAAC,CAAC;IAC5F,MAAM,QAAQ,GAAG,OAAO,CAAC,QAAQ,IAAI,OAAO,CAAC,QAAQ,CAAC;IACtD,MAAM,OAAO,GAAG,OAAO,CAAC,OAAO,IAAI,CAAC,CAAC;IACrC,MAAM,KAAK,GAAyB;QAClC,YAAY,EAAE,QAAQ,CAAC,MAAM,CAAC,WAAW;QACzC,SAAS,EAAE,QAAQ;QACnB,QAAQ,EAAE,OAAO;KAClB,CAAC;IACF,MAAM,IAAI,GAA6C,OAAO,CAAC,UAAU;QACvE,CAAC,CAAC,CAAC,QAAQ,CAAC,MAAM,CAAC,WAAW,EAAE,QAAQ,CAAC;QACzC,CAAC,CAAC,CAAC,KAAK,CAAC,CAAC;IACZ,OAAO,uBAAuB,CAAC,QAAQ,EAAE,IAAI,EAAE;QAC7C,SAAS,EAAE,QAAQ,CAAC,MAAM,CAAC,SAAS;QACpC,UAAU,EAAE,CAAC,QAAQ,EAAE,EAAE;YACvB,QAAQ,CAAC,QAAQ,GAAG,QAAQ,CAAC;YAC7B,OAAO,CAAC,UAAU,EAAE,CAAC,QAAQ,CAAC,CAAC;QACjC,CAAC;KACF,CAAC,CAAC;AACL,CAAC"}

package/dist/activities/execute-deep-agent/__test-utils__/hitl-script.d.ts

@@ -0,0 +1,24 @@
+/**
+ * The one gated opening every HITL arm shares, so their first-invocation
+ * statuses are byte-identical and can match ONE golden file
+ * (`__tests__/hermetic/goldens/gate.turn1.status.json`). Matching the same
+ * file from the approve, skip, reject and sequential-gates arms is itself an
+ * assertion: the pending decision is invisible to the first invocation.
+ *
+ * Why `execute`: under capture mode (on whenever a workspace dir and artifact
+ * storage exist — every hermetic run) file writes FLOW into the CAS ledger
+ * and are reviewed post hoc; shell is the built-in that still interrupts
+ * (`middleware/approval-gate.ts`, `shared/tool-kind.ts` SHELL). `echo` is a
+ * real subprocess with byte-stable output, the same tool
+ * `__tests__/shell-execute-gate.test.ts` drives through the real gate.
+ */
+import type { ScriptedToolCall, ScriptedTurn } from "./scripted-model.js";
+export declare const EXECUTE_CALL_A: ScriptedToolCall;
+export declare const EXECUTE_CALL_B: ScriptedToolCall;
+/** The gate's copy for a SHELL built-in (`CATEGORY_APPROVAL_MESSAGE.shell` over the args). */
+export declare function executeApprovalMessage(call: ScriptedToolCall): string;
+/** Round 0 of every gated arm: one text block and the gated `execute` A. */
+export declare const GATED_OPENING_TURN: ScriptedTurn;
+/** The closing text after the decision has been carried out (or declined). */
+export declare const CLOSING_TURN: ScriptedTurn;
+export declare const GATE_TURN1_GOLDEN = "./goldens/gate.turn1.status.json";

package/dist/activities/execute-deep-agent/__test-utils__/hitl-script.js

@@ -0,0 +1,41 @@
+/**
+ * The one gated opening every HITL arm shares, so their first-invocation
+ * statuses are byte-identical and can match ONE golden file
+ * (`__tests__/hermetic/goldens/gate.turn1.status.json`). Matching the same
+ * file from the approve, skip, reject and sequential-gates arms is itself an
+ * assertion: the pending decision is invisible to the first invocation.
+ *
+ * Why `execute`: under capture mode (on whenever a workspace dir and artifact
+ * storage exist — every hermetic run) file writes FLOW into the CAS ledger
+ * and are reviewed post hoc; shell is the built-in that still interrupts
+ * (`middleware/approval-gate.ts`, `shared/tool-kind.ts` SHELL). `echo` is a
+ * real subprocess with byte-stable output, the same tool
+ * `__tests__/shell-execute-gate.test.ts` drives through the real gate.
+ */
+export const EXECUTE_CALL_A = {
+    id: "call-hermetic-exec-0001",
+    name: "execute",
+    args: { command: "echo hermetic-a" },
+};
+export const EXECUTE_CALL_B = {
+    id: "call-hermetic-exec-0002",
+    name: "execute",
+    args: { command: "echo hermetic-b" },
+};
+/** The gate's copy for a SHELL built-in (`CATEGORY_APPROVAL_MESSAGE.shell` over the args). */
+export function executeApprovalMessage(call) {
+    return `Execute command: ${String(call.args.command)}`;
+}
+/** Round 0 of every gated arm: one text block and the gated `execute` A. */
+export const GATED_OPENING_TURN = {
+    text: "I will run the command.",
+    toolCalls: [EXECUTE_CALL_A],
+    usage: { inputTokens: 1_400, outputTokens: 50 },
+};
+/** The closing text after the decision has been carried out (or declined). */
+export const CLOSING_TURN = {
+    text: "Done with the command.",
+    usage: { inputTokens: 1_600, outputTokens: 20 },
+};
+export const GATE_TURN1_GOLDEN = "./goldens/gate.turn1.status.json";
+//# sourceMappingURL=hitl-script.js.map

package/dist/activities/execute-deep-agent/__test-utils__/hitl-script.js.map

@@ -0,0 +1 @@
+{"version":3,"file":"hitl-script.js","sourceRoot":"","sources":["../../../../src/activities/execute-deep-agent/__test-utils__/hitl-script.ts"],"names":[],"mappings":"AAAA;;;;;;;;;;;;;GAaG;AAIH,MAAM,CAAC,MAAM,cAAc,GAAqB;IAC9C,EAAE,EAAE,yBAAyB;IAC7B,IAAI,EAAE,SAAS;IACf,IAAI,EAAE,EAAE,OAAO,EAAE,iBAAiB,EAAE;CACrC,CAAC;AAEF,MAAM,CAAC,MAAM,cAAc,GAAqB;IAC9C,EAAE,EAAE,yBAAyB;IAC7B,IAAI,EAAE,SAAS;IACf,IAAI,EAAE,EAAE,OAAO,EAAE,iBAAiB,EAAE;CACrC,CAAC;AAEF,8FAA8F;AAC9F,MAAM,UAAU,sBAAsB,CAAC,IAAsB;IAC3D,OAAO,oBAAoB,MAAM,CAAC,IAAI,CAAC,IAAI,CAAC,OAAO,CAAC,EAAE,CAAC;AACzD,CAAC;AAED,4EAA4E;AAC5E,MAAM,CAAC,MAAM,kBAAkB,GAAiB;IAC9C,IAAI,EAAE,yBAAyB;IAC/B,SAAS,EAAE,CAAC,cAAc,CAAC;IAC3B,KAAK,EAAE,EAAE,WAAW,EAAE,KAAK,EAAE,YAAY,EAAE,EAAE,EAAE;CAChD,CAAC;AAEF,8EAA8E;AAC9E,MAAM,CAAC,MAAM,YAAY,GAAiB;IACxC,IAAI,EAAE,wBAAwB;IAC9B,KAAK,EAAE,EAAE,WAAW,EAAE,KAAK,EAAE,YAAY,EAAE,EAAE,EAAE;CAChD,CAAC;AAEF,MAAM,CAAC,MAAM,iBAAiB,GAAG,kCAAkC,CAAC"}

package/dist/activities/execute-deep-agent/__test-utils__/scripted-model-module.d.ts

@@ -0,0 +1,54 @@
+/**
+ * The module-boundary double for `shared/model-client.ts` — how a
+ * `ScriptedModel` reaches the REAL native engine build.
+ *
+ * `buildChatModel` is the runner's single construction point for LangChain
+ * chat models, and a native turn reaches it three ways: the primary model
+ * (`turn-setup.ts` `buildEngine`), every sub-agent's `modelFactory` (the same
+ * `buildModelFor`), and the runtime's tier-2 structured-output extractor
+ * (`shared/extract-structured-output.ts`). None of them takes a model as an
+ * argument, so the one seam a hermetic run has is this module — the native analog of the Cursor
+ * driver's `vi.mock("@cursor/sdk")`: the vendor engine doubled where the
+ * runner imports it, and nothing else (S3 M0 ruling Q-M0-1, amending
+ * Q-S3-11).
+ *
+ * `vi.mock` is hoisted and must live in the test file; the factory can
+ * `await import()` this module:
+ *
+ * ```ts
+ * vi.mock("../../../../shared/model-client.js", async () =>
+ *   (await import("../../__test-utils__/scripted-model-module.js")).scriptedModelClientModule(),
+ * );
+ * ```
+ *
+ * Every build the activity requests is recorded (`builds`) so a scenario can
+ * assert WHAT production asked for — the registry id, the service tier, the
+ * proxy posture — while the model handed back is the scenario's script. The
+ * provider is inferred from the requested name exactly as the real module
+ * does, so a golden's provider-labelled copy reads as in production.
+ *
+ * A scenario binds ONE model (every build of every execution gets it) or a
+ * RESOLVER over the build's options: production scopes every build to its
+ * execution through `headerScope.executionId` (the proxy's attribution
+ * header), so a test that runs several executions at once — the harness
+ * contract kit's concurrency invariant — routes each build to that
+ * execution's own engine through the id production already sends, and
+ * through nothing added for the test (S3 M3).
+ */
+import type { BuildChatModelOptions, BuiltChatModel } from "../../../shared/model-client.js";
+import type { ScriptedModel } from "./scripted-model.js";
+/** Which `ScriptedModel` answers one `buildChatModel` call. */
+export type ScriptedModelResolver = (opts: BuildChatModelOptions) => ScriptedModel;
+/**
+ * Bind the model every `buildChatModel` call hands back for the current
+ * scenario — or a resolver that picks one per call — and forget the previous
+ * scenario's build log. Called by the driver at `beginDeepAgentScenario`,
+ * before the activity factory runs.
+ */
+export declare function bindScriptedModel(model: ScriptedModel | ScriptedModelResolver): void;
+/** The `buildChatModel` requests the activity made in this scenario, in order. */
+export declare function recordedModelBuilds(): readonly BuildChatModelOptions[];
+/** The factory a test passes to `vi.mock(".../shared/model-client.js", ...)`. */
+export declare function scriptedModelClientModule(): {
+    buildChatModel: (opts: BuildChatModelOptions) => Promise<BuiltChatModel>;
+};

package/dist/activities/execute-deep-agent/__test-utils__/scripted-model-module.js

@@ -0,0 +1,72 @@
+/**
+ * The module-boundary double for `shared/model-client.ts` — how a
+ * `ScriptedModel` reaches the REAL native engine build.
+ *
+ * `buildChatModel` is the runner's single construction point for LangChain
+ * chat models, and a native turn reaches it three ways: the primary model
+ * (`turn-setup.ts` `buildEngine`), every sub-agent's `modelFactory` (the same
+ * `buildModelFor`), and the runtime's tier-2 structured-output extractor
+ * (`shared/extract-structured-output.ts`). None of them takes a model as an
+ * argument, so the one seam a hermetic run has is this module — the native analog of the Cursor
+ * driver's `vi.mock("@cursor/sdk")`: the vendor engine doubled where the
+ * runner imports it, and nothing else (S3 M0 ruling Q-M0-1, amending
+ * Q-S3-11).
+ *
+ * `vi.mock` is hoisted and must live in the test file; the factory can
+ * `await import()` this module:
+ *
+ * ```ts
+ * vi.mock("../../../../shared/model-client.js", async () =>
+ *   (await import("../../__test-utils__/scripted-model-module.js")).scriptedModelClientModule(),
+ * );
+ * ```
+ *
+ * Every build the activity requests is recorded (`builds`) so a scenario can
+ * assert WHAT production asked for — the registry id, the service tier, the
+ * proxy posture — while the model handed back is the scenario's script. The
+ * provider is inferred from the requested name exactly as the real module
+ * does, so a golden's provider-labelled copy reads as in production.
+ *
+ * A scenario binds ONE model (every build of every execution gets it) or a
+ * RESOLVER over the build's options: production scopes every build to its
+ * execution through `headerScope.executionId` (the proxy's attribution
+ * header), so a test that runs several executions at once — the harness
+ * contract kit's concurrency invariant — routes each build to that
+ * execution's own engine through the id production already sends, and
+ * through nothing added for the test (S3 M3).
+ */
+import { stripProviderPrefix, tryInferProvider } from "../../../shared/llm-proxy.js";
+let resolve;
+const builds = [];
+/**
+ * Bind the model every `buildChatModel` call hands back for the current
+ * scenario — or a resolver that picks one per call — and forget the previous
+ * scenario's build log. Called by the driver at `beginDeepAgentScenario`,
+ * before the activity factory runs.
+ */
+export function bindScriptedModel(model) {
+    resolve = typeof model === "function" ? model : () => model;
+    builds.length = 0;
+}
+/** The `buildChatModel` requests the activity made in this scenario, in order. */
+export function recordedModelBuilds() {
+    return builds;
+}
+/** The factory a test passes to `vi.mock(".../shared/model-client.js", ...)`. */
+export function scriptedModelClientModule() {
+    return {
+        buildChatModel: async (opts) => {
+            if (!resolve) {
+                throw new Error("scripted-model-module: no ScriptedModel bound — call bindScriptedModel(model) " +
+                    "before constructing the activities");
+            }
+            builds.push(opts);
+            return {
+                model: resolve(opts),
+                provider: tryInferProvider(opts.modelName) ?? "anthropic",
+                apiModelId: stripProviderPrefix(opts.modelName),
+            };
+        },
+    };
+}
+//# sourceMappingURL=scripted-model-module.js.map

package/dist/activities/execute-deep-agent/__test-utils__/scripted-model-module.js.map

@@ -0,0 +1 @@
+{"version":3,"file":"scripted-model-module.js","sourceRoot":"","sources":["../../../../src/activities/execute-deep-agent/__test-utils__/scripted-model-module.ts"],"names":[],"mappings":"AAAA;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;GAoCG;AAGH,OAAO,EAAE,mBAAmB,EAAE,gBAAgB,EAAE,MAAM,8BAA8B,CAAC;AAMrF,IAAI,OAA0C,CAAC;AAC/C,MAAM,MAAM,GAA4B,EAAE,CAAC;AAE3C;;;;;GAKG;AACH,MAAM,UAAU,iBAAiB,CAAC,KAA4C;IAC5E,OAAO,GAAG,OAAO,KAAK,KAAK,UAAU,CAAC,CAAC,CAAC,KAAK,CAAC,CAAC,CAAC,GAAG,EAAE,CAAC,KAAK,CAAC;IAC5D,MAAM,CAAC,MAAM,GAAG,CAAC,CAAC;AACpB,CAAC;AAED,kFAAkF;AAClF,MAAM,UAAU,mBAAmB;IACjC,OAAO,MAAM,CAAC;AAChB,CAAC;AAED,iFAAiF;AACjF,MAAM,UAAU,yBAAyB;IAGvC,OAAO;QACL,cAAc,EAAE,KAAK,EAAE,IAAI,EAAE,EAAE;YAC7B,IAAI,CAAC,OAAO,EAAE,CAAC;gBACb,MAAM,IAAI,KAAK,CACb,gFAAgF;oBAC9E,oCAAoC,CACvC,CAAC;YACJ,CAAC;YACD,MAAM,CAAC,IAAI,CAAC,IAAI,CAAC,CAAC;YAClB,OAAO;gBACL,KAAK,EAAE,OAAO,CAAC,IAAI,CAAC;gBACpB,QAAQ,EAAE,gBAAgB,CAAC,IAAI,CAAC,SAAS,CAAC,IAAI,WAAW;gBACzD,UAAU,EAAE,mBAAmB,CAAC,IAAI,CAAC,SAAS,CAAC;aAChD,CAAC;QACJ,CAAC;KACF,CAAC;AACJ,CAAC"}