@moxxy/sdk 0.15.0 → 0.15.1

package/src/tool-dispatch.test.ts

@@ -0,0 +1,215 @@
+import { describe, expect, it } from 'vitest';
+import { asSessionId, asTurnId } from './ids.js';
+import type { EmittedEvent, MoxxyEvent } from './events.js';
+import type { ModeContext } from './mode.js';
+import type { CollectedToolUse, StuckLoopDetector, StuckSignal } from './mode-helpers.js';
+import type { ToolCallVerdict } from './hooks.js';
+import type { PermissionDecision } from './permission.js';
+import {
+  dispatchToolCall,
+  executeToolUses,
+  emitRequestsAndDetectStuck,
+  type StuckLoopReport,
+} from './tool-dispatch.js';
+
+/**
+ * The orphan-prevention guarantees in tool-dispatch are load-bearing: the event
+ * log must never end on a `tool_call_requested` without a paired `tool_result`,
+ * across hook throws, resolver throws, emit throws, mid-batch abort, and a
+ * stuck-loop trip. These tests drive each defensive path directly (the mode
+ * tests only exercise the happy path).
+ */
+
+interface StubOpts {
+  readonly verdict?: ToolCallVerdict | (() => ToolCallVerdict | Promise<ToolCallVerdict>);
+  readonly decision?: PermissionDecision;
+  readonly execute?: (name: string, input: unknown) => unknown;
+  /** Throw from `emit` for these 0-based emit indices. */
+  readonly throwOnEmitIndices?: ReadonlyArray<number>;
+  readonly aborted?: boolean;
+}
+
+function makeCtx(opts: StubOpts = {}): { ctx: ModeContext; events: MoxxyEvent[] } {
+  const events: MoxxyEvent[] = [];
+  let emitCount = 0;
+  const controller = new AbortController();
+  if (opts.aborted) controller.abort();
+
+  const emit = async (event: EmittedEvent): Promise<MoxxyEvent> => {
+    const idx = emitCount++;
+    if (opts.throwOnEmitIndices?.includes(idx)) {
+      throw new Error('emit boom');
+    }
+    const full = { ...event, id: `e${idx}`, ts: idx } as unknown as MoxxyEvent;
+    events.push(full);
+    return full;
+  };
+
+  const ctx = {
+    sessionId: asSessionId('s1'),
+    turnId: asTurnId('t1'),
+    cwd: '/tmp',
+    env: {},
+    log: { events: () => [], subscribe: () => () => {} } as unknown,
+    signal: controller.signal,
+    hooks: {
+      dispatchToolCall: async (): Promise<ToolCallVerdict> => {
+        const v = opts.verdict ?? { action: 'allow' };
+        return typeof v === 'function' ? v() : v;
+      },
+    } as unknown,
+    permissions: {
+      check: async (): Promise<PermissionDecision> => opts.decision ?? { mode: 'allow' },
+    } as unknown,
+    tools: {
+      get: () => undefined,
+      execute: async (name: string, input: unknown) =>
+        opts.execute ? opts.execute(name, input) : 'ok',
+    } as unknown,
+    emit,
+  } as unknown as ModeContext;
+
+  return { ctx, events };
+}
+
+async function drain(
+  gen: AsyncGenerator<MoxxyEvent, unknown, unknown>,
+): Promise<unknown> {
+  let res = await gen.next();
+  while (!res.done) res = await gen.next();
+  return res.value;
+}
+
+const tool: CollectedToolUse = { id: 'c1', name: 'Read', input: { x: 1 } };
+
+describe('dispatchToolCall — orphan prevention', () => {
+  it('happy path ends with a successful tool_result', async () => {
+    const { ctx, events } = makeCtx({ execute: () => 'output' });
+    await drain(dispatchToolCall(ctx, tool, 0));
+    const result = events.find((e) => e.type === 'tool_result') as { ok: boolean } | undefined;
+    expect(result?.ok).toBe(true);
+  });
+
+  it('synthesizes a failed result when the hook throws (no orphan)', async () => {
+    const { ctx, events } = makeCtx({
+      verdict: () => {
+        throw new Error('hook exploded');
+      },
+    });
+    await drain(dispatchToolCall(ctx, tool, 0));
+    const result = events.find((e) => e.type === 'tool_result') as
+      | { ok: boolean; error: { message: string } }
+      | undefined;
+    expect(result?.ok).toBe(false);
+    expect(result?.error.message).toContain('pre-execute failure');
+    expect(result?.error.message).toContain('hook exploded');
+  });
+
+  it('synthesizes a failed result when the resolver throws', async () => {
+    const { ctx, events } = makeCtx({
+      decision: undefined,
+    });
+    // Make permissions.check itself throw via a hostile override.
+    (ctx.permissions as unknown as { check: () => Promise<never> }).check = async () => {
+      throw new Error('resolver down');
+    };
+    await drain(dispatchToolCall(ctx, tool, 0));
+    const result = events.find((e) => e.type === 'tool_result') as
+      | { ok: boolean; error: { message: string } }
+      | undefined;
+    expect(result?.ok).toBe(false);
+    expect(result?.error.message).toContain('pre-execute failure');
+  });
+
+  it('labels a post-run emit failure as "tool ran but result emit failed"', async () => {
+    // Emit order: #0 = tool_call_approved, #1 = success tool_result (throws →
+    // inner catch), #2 = inner-catch failed tool_result (throws → outer catch).
+    // The outer catch must NOT mislabel a tool that already ran as a
+    // pre-execute failure.
+    const { ctx, events } = makeCtx({
+      execute: () => 'output',
+      throwOnEmitIndices: [1, 2],
+    });
+    await drain(dispatchToolCall(ctx, tool, 0));
+    const result = events.find((e) => e.type === 'tool_result') as
+      | { ok: boolean; error: { message: string } }
+      | undefined;
+    expect(result?.ok).toBe(false);
+    expect(result?.error.message).toContain('tool ran but result emit failed');
+    expect(result?.error.message).not.toContain('pre-execute failure');
+  });
+
+  it('emits a denied result + tool_result when the hook denies', async () => {
+    const { ctx, events } = makeCtx({ verdict: { action: 'deny', reason: 'nope' } });
+    await drain(dispatchToolCall(ctx, tool, 0));
+    expect(events.some((e) => e.type === 'tool_call_denied')).toBe(true);
+    const result = events.find((e) => e.type === 'tool_result') as { ok: boolean } | undefined;
+    expect(result?.ok).toBe(false);
+  });
+});
+
+describe('executeToolUses — mid-batch abort', () => {
+  it('synthesizes a tool_result for every un-run call and an abort when aborted up front', async () => {
+    const { ctx, events } = makeCtx({ aborted: true });
+    const calls: CollectedToolUse[] = [
+      { id: 'a', name: 'Read', input: {} },
+      { id: 'b', name: 'Read', input: {} },
+    ];
+    const stop = await drain(executeToolUses(ctx, calls, 0));
+    expect(stop).toBe(true);
+    const results = events.filter((e) => e.type === 'tool_result') as Array<{
+      callId: string;
+      error: { kind: string };
+    }>;
+    expect(new Set(results.map((r) => r.callId))).toEqual(new Set(['a', 'b']));
+    expect(results.every((r) => r.error.kind === 'aborted')).toBe(true);
+    expect(events.some((e) => e.type === 'abort')).toBe(true);
+  });
+});
+
+describe('emitRequestsAndDetectStuck — stuck trip', () => {
+  const report: StuckLoopReport = {
+    abortedResultMessage: 'stuck — not run',
+    nearHint: 'with nearly identical input',
+    fatalMessage: ({ toolName }) => `stuck on ${toolName}`,
+  };
+
+  function detector(trip: StuckSignal): StuckLoopDetector {
+    return { record: () => trip } as unknown as StuckLoopDetector;
+  }
+
+  it('synthesizes a failed tool_result for the emitted request before the fatal error', async () => {
+    const { ctx, events } = makeCtx();
+    const stop = await drain(
+      emitRequestsAndDetectStuck(
+        ctx,
+        [tool],
+        detector({ stuck: true, kind: 'exact', count: 3 }),
+        report,
+      ),
+    );
+    expect(stop).toBe(true);
+    // The request must be paired with a synthesized aborted result — no orphan.
+    expect(events.some((e) => e.type === 'tool_call_requested')).toBe(true);
+    const result = events.find((e) => e.type === 'tool_result') as
+      | { ok: boolean; error: { kind: string } }
+      | undefined;
+    expect(result?.ok).toBe(false);
+    expect(result?.error.kind).toBe('aborted');
+    expect(events.some((e) => e.type === 'error')).toBe(true);
+  });
+
+  it('returns false (no synthesis) when the detector does not trip', async () => {
+    const { ctx, events } = makeCtx();
+    const stop = await drain(
+      emitRequestsAndDetectStuck(
+        ctx,
+        [tool],
+        detector({ stuck: false, kind: 'exact', count: 0 }),
+        report,
+      ),
+    );
+    expect(stop).toBe(false);
+    expect(events.filter((e) => e.type === 'tool_result')).toHaveLength(0);
+  });
+});

package/src/tool-dispatch.ts

@@ -23,6 +23,11 @@ export async function* dispatchToolCall(
   t: CollectedToolUse,
   iteration: number,
 ): AsyncGenerator<MoxxyEvent, void, unknown> {
+  // Set once the tool body resolves, so the outer catch can tell a genuine
+  // pre-execute failure (hook/permission/emit-before-run threw) apart from a
+  // post-run failure (the tool ran — possibly with side effects — but emitting
+  // its success result threw). The two warrant different log wording.
+  let executed = false;
   try {
     const verdict = await ctx.hooks.dispatchToolCall({
       sessionId: ctx.sessionId,
@@ -70,6 +75,7 @@ export async function* dispatchToolCall(
         log: ctx.log,
         ...(ctx.subagents ? { subagents: ctx.subagents } : {}),
       });
+      executed = true;
       yield await ctx.emit({
         type: 'tool_result',
         sessionId: ctx.sessionId,
@@ -94,9 +100,13 @@ export async function* dispatchToolCall(
     }
   } catch (err) {
     // Defensive: a hook handler, permission resolver, or the emit itself threw
-    // before we could produce a tool_result. Synthesize a failed result so the
-    // event log stays well-formed (no orphan tool_call_requested).
+    // before (or after) we could produce a tool_result. Synthesize a failed
+    // result so the event log stays well-formed (no orphan tool_call_requested).
+    // When `executed` is set the tool actually ran (possibly with side effects)
+    // and only the success-result emit failed — don't mislabel that as a
+    // pre-execute failure.
     const message = err instanceof Error ? err.message : String(err);
+    const prefix = executed ? 'tool ran but result emit failed' : 'pre-execute failure';
     yield await ctx.emit({
       type: 'tool_result',
       sessionId: ctx.sessionId,
@@ -104,7 +114,7 @@ export async function* dispatchToolCall(
       source: 'tool',
       callId: asToolCallId(t.id),
       ok: false,
-      error: { kind: 'threw', message: `pre-execute failure: ${message}` },
+      error: { kind: 'threw', message: `${prefix}: ${message}` },
     });
   }
 }

package/src/tunnel.ts

@@ -82,8 +82,18 @@ function ensureExitHook(): void {
     liveChildren.clear();
   };
   process.once('exit', killAll);
-  process.once('SIGINT', killAll);
-  process.once('SIGTERM', killAll);
+  // Registering a SIGINT/SIGTERM listener suppresses Node's default
+  // terminate-on-signal behavior, so an entrypoint that spawns a tunnel but
+  // installs no exit handler of its own would swallow the first Ctrl-C and hang
+  // with the process still alive. Re-raise the signal after cleanup so the
+  // default termination still fires. `process.once` removes the listener before
+  // it runs, so re-raising hits Node's default disposition (terminate).
+  const onSignal = (sig: 'SIGINT' | 'SIGTERM') => (): void => {
+    killAll();
+    process.kill(process.pid, sig);
+  };
+  process.once('SIGINT', onSignal('SIGINT'));
+  process.once('SIGTERM', onSignal('SIGTERM'));
 }
 
 /** Track a spawned child; returns an `untrack()` that also kills it cleanly. */