@boyingliu01/xp-gate 0.35.2 → 0.35.4

package/plugins/dsh/lib/dialect.d.ts

@@ -17,8 +17,28 @@ export declare function shq(value: string): string;
 /**
  * Pick the dialect for the current platform.
  *
- * `process.platform` is the reliable signal available inside a tool body; it
- * mirrors exactly where DSH enables each sandbox/executor pair.
+ * @deprecated Platform is NOT the right signal (#669). It encodes the assumption
+ * "win32 => PowerShell executor", which held when this was written but is now
+ * false: Windows can mount a Git Bash executor, and a PowerShell command string
+ * handed to bash fails to parse. Use {@link detectDialectForShell} wherever a
+ * shell is available; this remains only as the last-resort fallback.
  */
 export declare function detectDialect(platform?: NodeJS.Platform): ShellDialect;
+/** Structural view of a shell executor — only the fields that reveal its dialect. */
+export interface DialectBearingShell {
+    /** Explicit dialect/kind announcement, when the executor offers one. */
+    shellDialect?: unknown;
+    dialect?: unknown;
+    /** Absolute path to the resolved bash binary — the Git Bash executor's tell. */
+    bashPath?: unknown;
+}
+/**
+ * Pick the dialect from the executor that will actually run the command (#669).
+ *
+ * Detection is structural — an explicit announcement, or a resolved `bashPath` —
+ * never a class name, so this stays decoupled from any particular executor
+ * package. When nothing is recognisable it falls back to the platform default,
+ * which equals the old behaviour and is therefore never worse.
+ */
+export declare function detectDialectForShell(shell?: unknown, platform?: NodeJS.Platform): ShellDialect;
 //# sourceMappingURL=dialect.d.ts.map

package/plugins/dsh/lib/dialect.d.ts.map

@@ -1 +1 @@
-{"version":3,"file":"dialect.d.ts","sourceRoot":"","sources":["../src/dialect.ts"],"names":[],"mappings":"AAAA;;;;;;;GAOG;AAEH,MAAM,MAAM,YAAY,GAAG,OAAO,GAAG,YAAY,CAAA;AAEjD,iDAAiD;AACjD,wBAAgB,UAAU,CAAC,KAAK,EAAE,MAAM,EAAE,OAAO,EAAE,YAAY,GAAG,MAAM,CASvE;AAED;;;GAGG;AACH,wBAAgB,GAAG,CAAC,KAAK,EAAE,MAAM,GAAG,MAAM,CAEzC;AAED;;;;;GAKG;AACH,wBAAgB,aAAa,CAAC,QAAQ,GAAE,MAAM,CAAC,QAA2B,GAAG,YAAY,CAExF"}
+{"version":3,"file":"dialect.d.ts","sourceRoot":"","sources":["../src/dialect.ts"],"names":[],"mappings":"AAAA;;;;;;;GAOG;AAEH,MAAM,MAAM,YAAY,GAAG,OAAO,GAAG,YAAY,CAAA;AAEjD,iDAAiD;AACjD,wBAAgB,UAAU,CAAC,KAAK,EAAE,MAAM,EAAE,OAAO,EAAE,YAAY,GAAG,MAAM,CASvE;AAED;;;GAGG;AACH,wBAAgB,GAAG,CAAC,KAAK,EAAE,MAAM,GAAG,MAAM,CAEzC;AAED;;;;;;;;GAQG;AACH,wBAAgB,aAAa,CAAC,QAAQ,GAAE,MAAM,CAAC,QAA2B,GAAG,YAAY,CAExF;AAED,qFAAqF;AACrF,MAAM,WAAW,mBAAmB;IAClC,wEAAwE;IACxE,YAAY,CAAC,EAAE,OAAO,CAAA;IACtB,OAAO,CAAC,EAAE,OAAO,CAAA;IACjB,gFAAgF;IAChF,QAAQ,CAAC,EAAE,OAAO,CAAA;CACnB;AAYD;;;;;;;GAOG;AACH,wBAAgB,qBAAqB,CACnC,KAAK,CAAC,EAAE,OAAO,EACf,QAAQ,GAAE,MAAM,CAAC,QAA2B,GAC3C,YAAY,CAQd"}

package/plugins/dsh/lib/dialect.js

@@ -27,10 +27,45 @@ export function shq(value) {
 /**
  * Pick the dialect for the current platform.
  *
- * `process.platform` is the reliable signal available inside a tool body; it
- * mirrors exactly where DSH enables each sandbox/executor pair.
+ * @deprecated Platform is NOT the right signal (#669). It encodes the assumption
+ * "win32 => PowerShell executor", which held when this was written but is now
+ * false: Windows can mount a Git Bash executor, and a PowerShell command string
+ * handed to bash fails to parse. Use {@link detectDialectForShell} wherever a
+ * shell is available; this remains only as the last-resort fallback.
  */
 export function detectDialect(platform = process.platform) {
     return platform === "win32" ? "powershell" : "posix";
 }
+function announceDialect(shell) {
+    for (const raw of [shell.shellDialect, shell.dialect]) {
+        if (typeof raw !== "string")
+            continue;
+        const lower = raw.toLowerCase();
+        if (lower.includes("bash") || lower.includes("posix"))
+            return "posix";
+        if (lower.includes("pwsh") || lower.includes("powershell"))
+            return "powershell";
+    }
+    return null;
+}
+/**
+ * Pick the dialect from the executor that will actually run the command (#669).
+ *
+ * Detection is structural — an explicit announcement, or a resolved `bashPath` —
+ * never a class name, so this stays decoupled from any particular executor
+ * package. When nothing is recognisable it falls back to the platform default,
+ * which equals the old behaviour and is therefore never worse.
+ */
+export function detectDialectForShell(shell, platform = process.platform) {
+    const fallback = detectDialect(platform);
+    if (shell === null || typeof shell !== "object")
+        return fallback;
+    const view = shell;
+    const announced = announceDialect(view);
+    if (announced !== null)
+        return announced;
+    if (typeof view.bashPath === "string" && view.bashPath.length > 0)
+        return "posix";
+    return fallback;
+}
 //# sourceMappingURL=dialect.js.map

package/plugins/dsh/lib/gate-runner.d.ts.map

@@ -1 +1 @@
-{"version":3,"file":"gate-runner.d.ts","sourceRoot":"","sources":["../src/gate-runner.ts"],"names":[],"mappings":"AAEA,OAAO,KAAK,EAAE,gBAAgB,EAAE,aAAa,EAAE,cAAc,EAAE,MAAM,wBAAwB,CAAA;AAE7F;;;;;;;;GAQG;AACH,MAAM,WAAW,WAAW;IAC1B,OAAO,CAAC,OAAO,EAAE,gBAAgB,GAAG,aAAa,CAAA;IACjD,OAAO,CAAC,IAAI,EAAE,aAAa,GAAG,OAAO,CAAC;QAAE,MAAM,IAAI,OAAO,CAAC,cAAc,CAAC,CAAA;KAAE,CAAC,CAAA;CAC7E;AAED,kFAAkF;AAClF,MAAM,MAAM,WAAW,GAAG,MAAM,GAAG,aAAa,GAAG,eAAe,CAAA;AAElE;;;;;;;;;;;;;;GAcG;AACH,wBAAgB,eAAe,CAAC,MAAM,EAAE,cAAc,GAAG,WAAW,CAUnE;AA+BD;;;;;;GAMG;AACH,wBAAgB,gBAAgB,CAAC,MAAM,EAAE,cAAc,GAAG,MAAM,CAa/D;AAED;;;;GAIG;AACH,wBAAsB,SAAS,CAC7B,KAAK,EAAE,WAAW,EAClB,MAAM,EAAE,WAAW,EACnB,OAAO,EAAE,MAAM,EACf,OAAO,EAAE,MAAM,EACf,SAAS,EAAE,MAAM,GAChB,OAAO,CAAC,MAAM,CAAC,CAUjB"}
+{"version":3,"file":"gate-runner.d.ts","sourceRoot":"","sources":["../src/gate-runner.ts"],"names":[],"mappings":"AAEA,OAAO,KAAK,EAAE,gBAAgB,EAAE,aAAa,EAAE,cAAc,EAAE,MAAM,wBAAwB,CAAA;AAE7F;;;;;;;;GAQG;AACH,MAAM,WAAW,WAAW;IAC1B,OAAO,CAAC,OAAO,EAAE,gBAAgB,GAAG,aAAa,CAAA;IACjD,OAAO,CAAC,IAAI,EAAE,aAAa,GAAG,OAAO,CAAC;QAAE,MAAM,IAAI,OAAO,CAAC,cAAc,CAAC,CAAA;KAAE,CAAC,CAAA;CAC7E;AAED,kFAAkF;AAClF,MAAM,MAAM,WAAW,GAAG,MAAM,GAAG,aAAa,GAAG,eAAe,CAAA;AAElE;;;;;;;;;;;;;;GAcG;AACH,wBAAgB,eAAe,CAAC,MAAM,EAAE,cAAc,GAAG,WAAW,CAuBnE;AAyCD;;;;;;GAMG;AACH,wBAAgB,gBAAgB,CAAC,MAAM,EAAE,cAAc,GAAG,MAAM,CAa/D;AAED;;;;GAIG;AACH,wBAAsB,SAAS,CAC7B,KAAK,EAAE,WAAW,EAClB,MAAM,EAAE,WAAW,EACnB,OAAO,EAAE,MAAM,EACf,OAAO,EAAE,MAAM,EACf,SAAS,EAAE,MAAM,GAChB,OAAO,CAAC,MAAM,CAAC,CAUjB"}

package/plugins/dsh/lib/gate-runner.js

@@ -21,12 +21,34 @@ export function classifyGateRun(result) {
         return "INDETERMINATE";
     if (result.stdout.truncated)
         return "INDETERMINATE";
+    // Cross-check the reported exit code against the marker our own command
+    // suffix wrote into stdout (#669).
+    //
+    // The exit code travels a fragile chain: shell -> harness -> classify. The
+    // marker is the same fact recorded twice by independent paths, so a
+    // disagreement means one of them was mangled, and a run whose outcome cannot
+    // be established must not be reported as a pass. This is the backstop for the
+    // case where the exit code is lost outside our control -- the suffix fix makes
+    // the two agree, but only an explicit check proves they do.
+    const marker = parseExitMarker(stdout);
+    if (marker !== null && result.exitCode !== null && marker !== result.exitCode)
+        return "INDETERMINATE";
     if (result.exitCode === 0)
         return stdout.length > 0 ? "PASS" : "INDETERMINATE";
     // Nonzero: a verdict exists only when the run produced something to show.
     const stderr = result.stderr.text.trim();
     return stdout.length > 0 || stderr.length > 0 ? "GATE FAILED" : "INDETERMINATE";
 }
+/** Read the `[XP-Gate] exit:<n>` marker the command suffix appends, if present. */
+function parseExitMarker(stdout) {
+    const matches = [...stdout.matchAll(/\[XP-Gate\] exit:(-?\d+)/g)];
+    if (matches.length === 0)
+        return null;
+    // Take the LAST marker: a run whose output happens to contain an earlier one
+    // must be judged by the status our suffix wrote at the end.
+    const value = Number.parseInt(matches[matches.length - 1][1], 10);
+    return Number.isNaN(value) ? null : value;
+}
 /** Detailed diagnostics appended after the verdict line, when present. */
 function renderDetails(result) {
     const lines = [];

package/plugins/dsh/lib/index.d.ts.map

@@ -1 +1 @@
-{"version":3,"file":"index.d.ts","sourceRoot":"","sources":["../src/index.ts"],"names":[],"mappings":"AAOA,OAAO,CAAC,MAAM,0BAA0B,CAAA;AACxC,OAAO,KAAK,EAAE,OAAO,EAAE,MAAM,qBAAqB,CAAA;AA0BlD,eAAO,MAAM,IAAI,iBAAiB,CAAA;AAGlC,eAAO,MAAM,MAAM,UAA+B,CAAA;AAsZlD;;;;;;;;;;;GAWG;AACH,eAAO,MAAM,MAAM;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;aA4BjB,CAAA;AAEF,wBAAgB,KAAK,CAAC,GAAG,EAAE,OAAO,EAAE,MAAM,GAAE,OAAY,GAAG,IAAI,CA6B9D"}
+{"version":3,"file":"index.d.ts","sourceRoot":"","sources":["../src/index.ts"],"names":[],"mappings":"AAOA,OAAO,CAAC,MAAM,0BAA0B,CAAA;AACxC,OAAO,KAAK,EAAE,OAAO,EAAE,MAAM,qBAAqB,CAAA;AA0BlD,eAAO,MAAM,IAAI,iBAAiB,CAAA;AAGlC,eAAO,MAAM,MAAM,UAA+B,CAAA;AAsZlD;;;;;;;;;;;GAWG;AACH,eAAO,MAAM,MAAM;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;aA4BjB,CAAA;AAEF,wBAAgB,KAAK,CAAC,GAAG,EAAE,OAAO,EAAE,MAAM,GAAE,OAAY,GAAG,IAAI,CAgC9D"}

package/plugins/dsh/lib/index.js

@@ -3,7 +3,7 @@ import { join } from "node:path";
 import { fileURLToPath } from "node:url";
 import { defineTool } from "@deepseek-ai/dsh-tools";
 import z from "@deepseek-ai/schemastery";
-import { GATE_WHITELIST, buildCommand, detectDialect, resolveTarget } from "./command.js";
+import { GATE_WHITELIST, buildCommand, detectDialectForShell, resolveTarget } from "./command.js";
 import { runXpGate } from "./gate-runner.js";
 import { DelphiSetupError, expandEnvDetailed, listCredentialReferences, loadDelphiConfig, toDelphiOptions, } from "./delphi-config.js";
 import { DelphiConfigError } from "./delphi.js";
@@ -409,10 +409,13 @@ export const Config = z.object({
 });
 export function apply(ctx, config = {}) {
     const cfg = config;
-    // DSH runs a POSIX shell off-Windows and PowerShell on Windows; generating the
-    // wrong dialect makes every tool fail to parse rather than degrade gracefully.
-    const dialect = detectDialect();
     const shell = ctx.shell;
+    // The EXECUTOR decides the dialect, not the platform (#669). "win32 implies
+    // PowerShell" stopped being true once Windows could mount a Git Bash executor:
+    // a PowerShell command string handed to bash exits 127 on a line that does not
+    // parse, so every gate was reported as failed. Falls back to the platform
+    // default when the executor does not reveal itself.
+    const dialect = detectDialectForShell(shell);
     // ctx.credentials is optional: outside DSH there is no store, and the resolver
     // then reads the process environment directly. Inside DSH this is what makes
     // delphi-review work on a machine whose key lives in the credential store rather

package/plugins/dsh/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@boyingliu01/dsh-plugin-xp-gate",
-  "version": "0.35.2",
+  "version": "0.35.4",
   "description": "XP-Gate deterministic quality gates + AI workflow skills as a DeepSeek Harness (DSH) native plugin: gate-check / gate-principles / gate-arch tools + 12 SKILL.md skills",
   "type": "module",
   "main": "./lib/index.js",

package/plugins/dsh/skills/delphi-review/AGENTS.md

@@ -1,9 +1,9 @@
 # SKILLS/DELPHI-REVIEW KNOWLEDGE BASE
 
 **Generated:** 2026-10-30
-**Commit:** 0a400e4c
+**Commit:** aefe7f40
 **Branch:** main
-**Version:** 0.35.2.0
+**Version:** 0.35.4.0
 
 ## OVERVIEW
 Delphi Consensus Review — multi-round anonymous expert review (≥90% threshold, exactly 3 experts with distinct executable model IDs). Supports design, requirements, and code-walkthrough modes.

package/plugins/dsh/skills/sprint-flow/AGENTS.md

@@ -1,9 +1,9 @@
 # SKILLS/SPRINT-FLOW KNOWLEDGE BASE
 
 **Generated:** 2026-10-30
-**Commit:** 0a400e4c
+**Commit:** aefe7f40
 **Branch:** main
-**Version:** 0.35.2.0
+**Version:** 0.35.4.0
 
 ## OVERVIEW
 **6-phase** development pipeline (v2.0 compact redesign, Issue #290): PREP → DESIGN → BUILD → VERIFY → SHIP → CLOSE. Phase 3/6 BUILD default build mode is **ralph-loop** (REQ-level iteration, 40-67% token savings vs parallel). HARD-GATE between DESIGN (2/6) and BUILD (3/6): design must pass Delphi review (≥90% consensus) before any coding.

package/plugins/dsh/skills/test-specification-alignment/AGENTS.md

@@ -1,9 +1,9 @@
 # SKILLS/TEST-SPECIFICATION-ALIGNMENT KNOWLEDGE BASE
 
 **Generated:** 2026-10-30
-**Commit:** 0a400e4c
+**Commit:** aefe7f40
 **Branch:** main
-**Version:** 0.35.2.0
+**Version:** 0.35.4.0
 
 ## OVERVIEW
 Test-Specification Alignment Engine — two-stage validation ensuring tests accurately reflect requirements and design specs.

package/plugins/dsh/src/command.ts

@@ -3,7 +3,7 @@ import { type ShellDialect, shellQuote } from "./dialect.js"
 
 // Re-exported so existing importers of this module keep working; the dialect
 // primitives themselves live in ./dialect.ts.
-export { type ShellDialect, detectDialect, shellQuote, shq } from "./dialect.js"
+export { type ShellDialect, detectDialect, detectDialectForShell, shellQuote, shq } from "./dialect.js"
 
 /**
  * Allowlist of xp-gate gate aliases that are safe to invoke standalone
@@ -88,12 +88,18 @@ export function buildCommand(options: BuildCommandOptions): string {
     // Merge stderr into stdout so a failure that only writes to stderr still
     // reaches the renderer, then echo the status so the exit code survives the
     // pipeline.
-    return `${inner} 2>&1; printf '\\n[XP-Gate] exit:%s\\n' "$?"`
+    // Capture the status BEFORE printing it, then `exit` with it. Without the
+    // explicit `exit`, the compound command's status is printf's -- always 0 --
+    // so classifyGateRun() would read a failed gate as PASS (#669).
+    return `${inner} 2>&1; __rc=$?; printf '\\n[XP-Gate] exit:%s\\n' "$__rc"; exit $__rc`
   }
   // PowerShell: merge streams and report $LASTEXITCODE explicitly. The
   // .cmd/.ps1 shims npm writes under %APPDATA%\npm are visible here but not to a
   // Git-Bash PATH lookup, which is why Windows uses this dialect.
-  return `${inner} 2>&1; Write-Output "[XP-Gate] exit:$LASTEXITCODE"`
+  //
+  // Same defect as the POSIX branch (#669): Write-Output resets the status, so
+  // the failing gate's code must be captured and re-exited explicitly.
+  return `${inner} 2>&1; $__rc=$LASTEXITCODE; Write-Output "[XP-Gate] exit:$__rc"; exit $__rc`
 }
 
 /** Resolve a model-supplied path against the session cwd; absolutes pass through. */

package/plugins/dsh/src/dialect.ts

@@ -32,9 +32,52 @@ export function shq(value: string): string {
 /**
  * Pick the dialect for the current platform.
  *
- * `process.platform` is the reliable signal available inside a tool body; it
- * mirrors exactly where DSH enables each sandbox/executor pair.
+ * @deprecated Platform is NOT the right signal (#669). It encodes the assumption
+ * "win32 => PowerShell executor", which held when this was written but is now
+ * false: Windows can mount a Git Bash executor, and a PowerShell command string
+ * handed to bash fails to parse. Use {@link detectDialectForShell} wherever a
+ * shell is available; this remains only as the last-resort fallback.
  */
 export function detectDialect(platform: NodeJS.Platform = process.platform): ShellDialect {
   return platform === "win32" ? "powershell" : "posix"
 }
+
+/** Structural view of a shell executor — only the fields that reveal its dialect. */
+export interface DialectBearingShell {
+  /** Explicit dialect/kind announcement, when the executor offers one. */
+  shellDialect?: unknown
+  dialect?: unknown
+  /** Absolute path to the resolved bash binary — the Git Bash executor's tell. */
+  bashPath?: unknown
+}
+
+function announceDialect(shell: DialectBearingShell): ShellDialect | null {
+  for (const raw of [shell.shellDialect, shell.dialect]) {
+    if (typeof raw !== "string") continue
+    const lower = raw.toLowerCase()
+    if (lower.includes("bash") || lower.includes("posix")) return "posix"
+    if (lower.includes("pwsh") || lower.includes("powershell")) return "powershell"
+  }
+  return null
+}
+
+/**
+ * Pick the dialect from the executor that will actually run the command (#669).
+ *
+ * Detection is structural — an explicit announcement, or a resolved `bashPath` —
+ * never a class name, so this stays decoupled from any particular executor
+ * package. When nothing is recognisable it falls back to the platform default,
+ * which equals the old behaviour and is therefore never worse.
+ */
+export function detectDialectForShell(
+  shell?: unknown,
+  platform: NodeJS.Platform = process.platform,
+): ShellDialect {
+  const fallback = detectDialect(platform)
+  if (shell === null || typeof shell !== "object") return fallback
+  const view = shell as DialectBearingShell
+  const announced = announceDialect(view)
+  if (announced !== null) return announced
+  if (typeof view.bashPath === "string" && view.bashPath.length > 0) return "posix"
+  return fallback
+}

package/plugins/dsh/src/exit-code.test.ts

@@ -0,0 +1,83 @@
+/**
+ * @test REQ-669 Gate 判定不得因命令串自身缺陷而失真
+ * @intent 这两个缺陷都不是"慢"或"报错"，而是**给出错误的结论**——尤其是假绿：
+ *
+ *         1. 方言按 `process.platform` 猜。它假设 "win32 ⇒ PowerShell"，这个前提在
+ *            #442 时代成立；但 Windows 现在可以挂 Git Bash executor
+ *            （`@very12345/dsh-bash-windows`）。于是生成 PowerShell 命令串丢进 bash，
+ *            `Write-Output` 不存在 → 退出码 127 → 通过的 gate 报 GATE FAILED。
+ *
+ *         2. 退出码靠尾部文本标记传递，但**标记本身吃掉了退出码**。两个分支都用
+ *            `printf`/`Write-Output` 收尾且都不 `exit`，复合命令的退出码等于最后一条
+ *            命令的退出码 → 恒为 0。而 `classifyGateRun()` 只读 `result.exitCode`，
+ *            从不看 stdout 里的 `exit:N`。于是**失败的 gate 被判 PASS**。
+ *
+ *         缺陷二更危险：缺陷一至少会让人怀疑，缺陷二会让人相信。
+ *
+ *         所以这里不靠字符串匹配断言——那证明不了退出码真的传出去了。真正的证明是
+ *         起一个真实 shell 跑一遍，断言进程退出码等于内层命令的真实退出码。
+ * @covers AC-669-01, AC-669-02, AC-669-03, AC-669-04
+ */
+import { execFileSync } from "node:child_process"
+import { buildCommand, detectDialectForShell } from "./command.js"
+import { classifyGateRun } from "./gate-runner.js"
+
+/** 用一个必然退出的内层命令替换真实 xp-gate，只考察后缀对退出码的传递。 */
+function runInRealShell(command: string, shell: "bash" | "bash"): number {
+  try {
+    execFileSync(shell, ["-c", command], { stdio: "pipe", timeout: 20_000 })
+    return 0
+  } catch (error) {
+    const code = (error as { status?: number | null }).status
+    // `status` is null when the process was killed by a signal; surface that as
+    // a distinct value so a hang can never be mistaken for a clean exit.
+    return typeof code === "number" ? code : -1
+  }
+}
+
+describe("#669 命令串的退出码必须真实传递", () => {
+  it("AC-669-01: POSIX 后缀把内层退出码原样带出来（不是恒为 0）", () => {
+    // 用 `sh -c 'exit 7'` 冒充失败的 gate。修复前：文本标记说 7，但进程退出码是 0，
+    // 于是 classifyGateRun 看到 exitCode===0 + 有输出 → PASS。假绿。
+    const suffix = buildCommand({ subcommand: "principles", dialect: "posix" })
+      .replace(/^xp-gate principles/, "sh -c 'exit 7'")
+    expect(runInRealShell(suffix, "bash")).toBe(7)
+  })
+
+  it("AC-669-02: POSIX 后缀在成功时仍返回 0（不能为了传失败码而让成功变红）", () => {
+    const suffix = buildCommand({ subcommand: "principles", dialect: "posix" })
+      .replace(/^xp-gate principles/, "sh -c 'exit 0'")
+    expect(runInRealShell(suffix, "bash")).toBe(0)
+  })
+
+  it("AC-669-03: 方言由 executor 决定，而不是 process.platform", () => {
+    // Git Bash executor 在 win32 上暴露 bashPath；按平台猜会得到 powershell。
+    expect(detectDialectForShell({ bashPath: "C:\\Program Files\\Git\\bin\\bash.exe" }, "win32")).toBe("posix")
+    // 显式声明优先于一切推断。
+    expect(detectDialectForShell({ shellDialect: "bash" }, "win32")).toBe("posix")
+    expect(detectDialectForShell({ dialect: "pwsh" }, "linux")).toBe("powershell")
+    // 认不出来时回落到原平台默认 —— 等效旧行为，不会比现在更差。
+    expect(detectDialectForShell({}, "win32")).toBe("powershell")
+    expect(detectDialectForShell(null, "linux")).toBe("posix")
+    // PowerShell executor 带 bashPath 的假想组合：显式声明优先。
+    expect(detectDialectForShell({ shellDialect: "powershell", bashPath: "x" }, "win32")).toBe("powershell")
+  })
+
+  it("AC-669-04: classifyGateRun 对退出码与文本标记不一致的 run 报 INDETERMINATE", () => {
+    // 即使后缀修好，退出码仍要穿过 shell → harness → 分类器这条脆弱链路。
+    // 两者不一致时不能静默采信其中之一。
+    const base = {
+      timedOut: false,
+      timeoutMs: 1000,
+      stdout: { text: "some output\n[XP-Gate] exit:7\n", truncated: false },
+      stderr: { text: "", truncated: false },
+    }
+    // exitCode 说成功、标记说失败 → 冲突，不可判定。
+    expect(classifyGateRun({ ...base, exitCode: 0 })).toBe("INDETERMINATE")
+    // 两者都说失败 → 正常判 GATE FAILED。
+    expect(classifyGateRun({ ...base, exitCode: 7 })).toBe("GATE FAILED")
+    // 两者都说成功 → PASS。
+    const ok = { ...base, stdout: { text: "ok\n[XP-Gate] exit:0\n", truncated: false } }
+    expect(classifyGateRun({ ...ok, exitCode: 0 })).toBe("PASS")
+  })
+})

package/plugins/dsh/src/gate-runner.ts

@@ -39,6 +39,19 @@ export function classifyGateRun(result: ShellRunResult): GateVerdict {
 
   if (result.timedOut) return "INDETERMINATE"
   if (result.stdout.truncated) return "INDETERMINATE"
+
+  // Cross-check the reported exit code against the marker our own command
+  // suffix wrote into stdout (#669).
+  //
+  // The exit code travels a fragile chain: shell -> harness -> classify. The
+  // marker is the same fact recorded twice by independent paths, so a
+  // disagreement means one of them was mangled, and a run whose outcome cannot
+  // be established must not be reported as a pass. This is the backstop for the
+  // case where the exit code is lost outside our control -- the suffix fix makes
+  // the two agree, but only an explicit check proves they do.
+  const marker = parseExitMarker(stdout)
+  if (marker !== null && result.exitCode !== null && marker !== result.exitCode) return "INDETERMINATE"
+
   if (result.exitCode === 0) return stdout.length > 0 ? "PASS" : "INDETERMINATE"
 
   // Nonzero: a verdict exists only when the run produced something to show.
@@ -46,6 +59,16 @@ export function classifyGateRun(result: ShellRunResult): GateVerdict {
   return stdout.length > 0 || stderr.length > 0 ? "GATE FAILED" : "INDETERMINATE"
 }
 
+/** Read the `[XP-Gate] exit:<n>` marker the command suffix appends, if present. */
+function parseExitMarker(stdout: string): number | null {
+  const matches = [...stdout.matchAll(/\[XP-Gate\] exit:(-?\d+)/g)]
+  if (matches.length === 0) return null
+  // Take the LAST marker: a run whose output happens to contain an earlier one
+  // must be judged by the status our suffix wrote at the end.
+  const value = Number.parseInt(matches[matches.length - 1][1], 10)
+  return Number.isNaN(value) ? null : value
+}
+
 /** Detailed diagnostics appended after the verdict line, when present. */
 function renderDetails(result: ShellRunResult): string[] {
   const lines: string[] = []

package/plugins/dsh/src/index.ts

@@ -8,7 +8,7 @@ import type { PostToolDecision, ToolExecution, ToolExecutionResult } from "@deep
 import z from "@deepseek-ai/schemastery"
 import type { Context } from "@deepseek-ai/cordis"
 import type { ShellDialect } from "./command.js"
-import { GATE_WHITELIST, buildCommand, detectDialect, resolveTarget } from "./command.js"
+import { GATE_WHITELIST, buildCommand, detectDialectForShell, resolveTarget } from "./command.js"
 import { runXpGate } from "./gate-runner.js"
 import type { ResolvedDelphiConfig } from "./delphi-config.js"
 import {
@@ -485,10 +485,13 @@ export const Config = z.object({
 
 export function apply(ctx: Context, config: unknown = {}): void {
   const cfg = config as Partial<XpGatePluginConfig>
-  // DSH runs a POSIX shell off-Windows and PowerShell on Windows; generating the
-  // wrong dialect makes every tool fail to parse rather than degrade gracefully.
-  const dialect = detectDialect()
   const shell = ctx.shell
+  // The EXECUTOR decides the dialect, not the platform (#669). "win32 implies
+  // PowerShell" stopped being true once Windows could mount a Git Bash executor:
+  // a PowerShell command string handed to bash exits 127 on a line that does not
+  // parse, so every gate was reported as failed. Falls back to the platform
+  // default when the executor does not reveal itself.
+  const dialect = detectDialectForShell(shell)
   // ctx.credentials is optional: outside DSH there is no store, and the resolver
   // then reads the process environment directly. Inside DSH this is what makes
   // delphi-review work on a machine whose key lives in the credential store rather

package/plugins/opencode/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@boyingliu01/opencode-plugin",
-  "version": "0.35.2",
+  "version": "0.35.4",
   "type": "module",
   "main": "index.ts",
   "description": "XP-Gate quality gates + AI workflow skills + Sprint Flow TUI sidebar for OpenCode",

package/plugins/opencode/skills/delphi-review/AGENTS.md

@@ -1,9 +1,9 @@
 # SKILLS/DELPHI-REVIEW KNOWLEDGE BASE
 
 **Generated:** 2026-10-30
-**Commit:** 0a400e4c
+**Commit:** aefe7f40
 **Branch:** main
-**Version:** 0.35.2.0
+**Version:** 0.35.4.0
 
 ## OVERVIEW
 Delphi Consensus Review — multi-round anonymous expert review (≥90% threshold, exactly 3 experts with distinct executable model IDs). Supports design, requirements, and code-walkthrough modes.

package/plugins/opencode/skills/sprint-flow/AGENTS.md

@@ -1,9 +1,9 @@
 # SKILLS/SPRINT-FLOW KNOWLEDGE BASE
 
 **Generated:** 2026-10-30
-**Commit:** 0a400e4c
+**Commit:** aefe7f40
 **Branch:** main
-**Version:** 0.35.2.0
+**Version:** 0.35.4.0
 
 ## OVERVIEW
 **6-phase** development pipeline (v2.0 compact redesign, Issue #290): PREP → DESIGN → BUILD → VERIFY → SHIP → CLOSE. Phase 3/6 BUILD default build mode is **ralph-loop** (REQ-level iteration, 40-67% token savings vs parallel). HARD-GATE between DESIGN (2/6) and BUILD (3/6): design must pass Delphi review (≥90% consensus) before any coding.

package/plugins/opencode/skills/test-specification-alignment/AGENTS.md

@@ -1,9 +1,9 @@
 # SKILLS/TEST-SPECIFICATION-ALIGNMENT KNOWLEDGE BASE
 
 **Generated:** 2026-10-30
-**Commit:** 0a400e4c
+**Commit:** aefe7f40
 **Branch:** main
-**Version:** 0.35.2.0
+**Version:** 0.35.4.0
 
 ## OVERVIEW
 Test-Specification Alignment Engine — two-stage validation ensuring tests accurately reflect requirements and design specs.

package/plugins/qoder/plugin.json

@@ -1,6 +1,6 @@
 {
   "name": "xp-gate",
-  "version": "0.35.2",
+  "version": "0.35.4",
   "displayName": "XP-Gate",
   "description": "Extreme Programming quality gates + AI workflow skills for Qoder. Includes 12 pre-commit gates (Gate 0-11), 8 pre-push gates, Sprint Flow (6 phases), and Delphi multi-expert review (>=90% consensus).",
   "author": {

package/plugins/qoder/skills/delphi-review/AGENTS.md

@@ -1,9 +1,9 @@
 # SKILLS/DELPHI-REVIEW KNOWLEDGE BASE
 
 **Generated:** 2026-10-30
-**Commit:** 0a400e4c
+**Commit:** aefe7f40
 **Branch:** main
-**Version:** 0.35.2.0
+**Version:** 0.35.4.0
 
 ## OVERVIEW
 Delphi Consensus Review — multi-round anonymous expert review (≥90% threshold, exactly 3 experts with distinct executable model IDs). Supports design, requirements, and code-walkthrough modes.

package/plugins/qoder/skills/sprint-flow/AGENTS.md

@@ -1,9 +1,9 @@
 # SKILLS/SPRINT-FLOW KNOWLEDGE BASE
 
 **Generated:** 2026-10-30
-**Commit:** 0a400e4c
+**Commit:** aefe7f40
 **Branch:** main
-**Version:** 0.35.2.0
+**Version:** 0.35.4.0
 
 ## OVERVIEW
 **6-phase** development pipeline (v2.0 compact redesign, Issue #290): PREP → DESIGN → BUILD → VERIFY → SHIP → CLOSE. Phase 3/6 BUILD default build mode is **ralph-loop** (REQ-level iteration, 40-67% token savings vs parallel). HARD-GATE between DESIGN (2/6) and BUILD (3/6): design must pass Delphi review (≥90% consensus) before any coding.

package/plugins/qoder/skills/test-specification-alignment/AGENTS.md

@@ -1,9 +1,9 @@
 # SKILLS/TEST-SPECIFICATION-ALIGNMENT KNOWLEDGE BASE
 
 **Generated:** 2026-10-30
-**Commit:** 0a400e4c
+**Commit:** aefe7f40
 **Branch:** main
-**Version:** 0.35.2.0
+**Version:** 0.35.4.0
 
 ## OVERVIEW
 Test-Specification Alignment Engine — two-stage validation ensuring tests accurately reflect requirements and design specs.

package/principles/AGENTS.md

@@ -1,9 +1,9 @@
 # PRINCIPLES CHECKER MODULE
 
 **Generated:** 2026-10-30
-**Commit:** 0a400e4c
+**Commit:** aefe7f40
 **Branch:** main
-**Version:** 0.35.2.0
+**Version:** 0.35.4.0
 
 ## OVERVIEW
 Clean Code & SOLID principles checker — **Gate 4** of pre-commit. 14 rules × 9 language adapters, SARIF 2.1.0 output. Houses the **Boy Scout Rule** enforcement engine (Gate 6) and warning-baseline storage.

package/skills/delphi-review/AGENTS.md

@@ -1,9 +1,9 @@
 # SKILLS/DELPHI-REVIEW KNOWLEDGE BASE
 
 **Generated:** 2026-10-30
-**Commit:** 0a400e4c
+**Commit:** aefe7f40
 **Branch:** main
-**Version:** 0.35.2.0
+**Version:** 0.35.4.0
 
 ## OVERVIEW
 Delphi Consensus Review — multi-round anonymous expert review (≥90% threshold, exactly 3 experts with distinct executable model IDs). Supports design, requirements, and code-walkthrough modes.

package/skills/sprint-flow/AGENTS.md

@@ -1,9 +1,9 @@
 # SKILLS/SPRINT-FLOW KNOWLEDGE BASE
 
 **Generated:** 2026-10-30
-**Commit:** 0a400e4c
+**Commit:** aefe7f40
 **Branch:** main
-**Version:** 0.35.2.0
+**Version:** 0.35.4.0
 
 ## OVERVIEW
 **6-phase** development pipeline (v2.0 compact redesign, Issue #290): PREP → DESIGN → BUILD → VERIFY → SHIP → CLOSE. Phase 3/6 BUILD default build mode is **ralph-loop** (REQ-level iteration, 40-67% token savings vs parallel). HARD-GATE between DESIGN (2/6) and BUILD (3/6): design must pass Delphi review (≥90% consensus) before any coding.

package/skills/test-specification-alignment/AGENTS.md

@@ -1,9 +1,9 @@
 # SKILLS/TEST-SPECIFICATION-ALIGNMENT KNOWLEDGE BASE
 
 **Generated:** 2026-10-30
-**Commit:** 0a400e4c
+**Commit:** aefe7f40
 **Branch:** main
-**Version:** 0.35.2.0
+**Version:** 0.35.4.0
 
 ## OVERVIEW
 Test-Specification Alignment Engine — two-stage validation ensuring tests accurately reflect requirements and design specs.