From 1005f02db26dd6c80be47fbf48065cc6128e0ac9 Mon Sep 17 00:00:00 2001
From: Tom Boucher <trekkie@nomorestars.com>
Date: Wed, 15 Apr 2026 14:59:27 -0400
Subject: [PATCH] fix(sdk): stop duplicating prompt in runPhaseStepSession user
 message (#2223)

runPhaseStepSession was passing the full prompt string as both the
user-visible prompt: argument and as systemPrompt.append, sending
the same (potentially large) text twice per invocation and doubling
the token cost for every phase step session.

runPlanSession correctly uses a short directive as the user message
and reserves the full content for systemPrompt.append only. Apply
the same pattern to runPhaseStepSession: use a brief
"Execute this phase step: <step>" directive as the user message.

Closes #2194

Co-authored-by: Claude Sonnet 4.6 <noreply@anthropic.com>
---
 sdk/src/session-runner.test.ts | 98 ++++++++++++++++++++++++++++++++++
 sdk/src/session-runner.ts      |  2 +-
 2 files changed, 99 insertions(+), 1 deletion(-)
 create mode 100644 sdk/src/session-runner.test.ts

diff --git a/sdk/src/session-runner.test.ts b/sdk/src/session-runner.test.ts
new file mode 100644
index 000000000..56ef7a90a
--- /dev/null
+++ b/sdk/src/session-runner.test.ts
@@ -0,0 +1,98 @@
+/**
+ * Unit tests for session-runner.ts
+ *
+ * Regression test for #2194: runPhaseStepSession was passing the full prompt
+ * string as both the user-visible prompt: message and systemPrompt.append,
+ * doubling the token cost on every phase step invocation.
+ */
+
+import { describe, it, expect, vi, beforeEach } from 'vitest';
+import { PhaseStepType } from './types.js';
+import { CONFIG_DEFAULTS } from './config.js';
+import type { GSDConfig } from './config.js';
+
+// ─── Mock the Agent SDK ───────────────────────────────────────────────────────
+
+// Capture the query call options so we can assert on them without making real API calls.
+const mockQueryCalls: Array<{ prompt: string; options: Record<string, unknown> }> = [];
+
+vi.mock('@anthropic-ai/claude-agent-sdk', () => {
+  async function* fakeQueryStream() {
+    // Yield a minimal success result message so processQueryStream completes.
+    yield {
+      type: 'result',
+      subtype: 'success',
+      session_id: 'test-session',
+      total_cost_usd: 0,
+      duration_ms: 1,
+      num_turns: 1,
+      usage: {
+        input_tokens: 10,
+        output_tokens: 5,
+        cache_read_input_tokens: 0,
+        cache_creation_input_tokens: 0,
+      },
+    };
+  }
+
+  return {
+    query: vi.fn((args: { prompt: string; options: Record<string, unknown> }) => {
+      mockQueryCalls.push({ prompt: args.prompt, options: args.options });
+      return fakeQueryStream();
+    }),
+  };
+});
+
+// ─── Import SUT after mock is hoisted ────────────────────────────────────────
+
+import { runPhaseStepSession } from './session-runner.js';
+
+// ─── Helpers ─────────────────────────────────────────────────────────────────
+
+function makeConfig(overrides: Partial<GSDConfig> = {}): GSDConfig {
+  return { ...CONFIG_DEFAULTS, ...overrides };
+}
+
+// ─── Tests ───────────────────────────────────────────────────────────────────
+
+describe('runPhaseStepSession', () => {
+  beforeEach(() => {
+    mockQueryCalls.length = 0;
+  });
+
+  it('does not duplicate the prompt in both the user message and systemPrompt.append', async () => {
+    const fullPrompt = 'You are a researcher. Investigate the topic thoroughly.\n\n## Context\nDetailed instructions here...';
+
+    await runPhaseStepSession(fullPrompt, PhaseStepType.Research, makeConfig());
+
+    expect(mockQueryCalls).toHaveLength(1);
+
+    const call = mockQueryCalls[0];
+    const appendValue = (call.options.systemPrompt as { append?: string })?.append;
+
+    // The full prompt must appear in systemPrompt.append (that is its correct location).
+    expect(appendValue).toBe(fullPrompt);
+
+    // The user-visible prompt: must NOT be the full prompt — it should be a short directive.
+    expect(call.prompt).not.toBe(fullPrompt);
+    expect(call.prompt.length).toBeLessThan(fullPrompt.length);
+  });
+
+  it('passes the full prompt in systemPrompt.append', async () => {
+    const fullPrompt = 'Complex multi-line\nprompt with $VARIABLES and $(command) patterns.';
+
+    await runPhaseStepSession(fullPrompt, PhaseStepType.Execute, makeConfig());
+
+    const call = mockQueryCalls[0];
+    const appendValue = (call.options.systemPrompt as { append?: string })?.append;
+
+    expect(appendValue).toBe(fullPrompt);
+  });
+
+  it('returns a successful PlanResult', async () => {
+    const result = await runPhaseStepSession('Test prompt', PhaseStepType.Verify, makeConfig());
+
+    expect(result.success).toBe(true);
+    expect(result.sessionId).toBe('test-session');
+  });
+});
diff --git a/sdk/src/session-runner.ts b/sdk/src/session-runner.ts
index 716f0da81..693a5927b 100644
--- a/sdk/src/session-runner.ts
+++ b/sdk/src/session-runner.ts
@@ -277,7 +277,7 @@ export async function runPhaseStepSession(
   const cwd = options?.cwd ?? process.cwd();
 
   const queryStream = query({
-    prompt: prompt,
+    prompt: `Execute this phase step: ${phaseStep}`,
     options: {
       systemPrompt: {
         type: 'preset',