fix: guide agents to quote the fixlab exec separator and avoid cmd /c

.github/agents/fixlab-autofix.agent.md

@@ -75,7 +75,10 @@ only for information that cannot be established safely from the repository.
    or create a pull request unless the developer explicitly requests it.
 
 Run verbose repository-owned tests, builds, linters, type checks, and package
-validation through `fixlab exec --stage <stage> -- <command> [args...]`. Keep
+validation through `fixlab exec --stage <stage> -- <command> [args...]`. In
+Windows PowerShell write the separator quoted (`'--'`) because an unquoted `--`
+is dropped, and never wrap `fixlab exec` in `cmd /c`; that wrapper is denied in
+non-interactive sessions. Keep
 the compact result in context and inspect its private evidence only when an
 exact diagnostic is needed. Use `--reuse` only for deterministic local
 validation with identical command and repository state. Never reuse startup,

agents/fixlab.md

@@ -51,7 +51,10 @@ instructions provide authoritative commands.
 10. Present evidence, skipped gates, and remaining risks.
 
 Run verbose repository-owned tests, builds, linters, type checks, and package
-validation through `fixlab exec --stage <stage> -- <command> [args...]`. Use
+validation through `fixlab exec --stage <stage> -- <command> [args...]`. In
+Windows PowerShell write the separator quoted (`'--'`) because an unquoted `--`
+is dropped, and never wrap `fixlab exec` in `cmd /c`; that wrapper is denied in
+non-interactive sessions. Use
 its compact result in agent context and refer to the private evidence path only
 when an exact diagnostic is needed. Add `--reuse` only for deterministic local
 validation when the command, arguments, Git state, working directory, profile,

bin/structured-command.js

@@ -284,7 +284,8 @@
   const separator = args.indexOf("--");
   if (separator < 0 || separator === args.length - 1) {
     throw new Error(
-      "usage: fixlab exec [repository] [--cwd <path>] [--stage <name>] [--reuse] [--timeout-seconds <seconds>] -- <command> [args...]"
+      "usage: fixlab exec [repository] [--cwd <path>] [--stage <name>] [--reuse] [--timeout-seconds <seconds>] -- <command> [args...]\n" +
+        "hint: in Windows PowerShell an unquoted -- is dropped; write '--' (quoted) and do not wrap the command in cmd /c."
     );
   }
 
@@ -758,14 +759,14 @@
         }
       }
     };
     const child = spawnProcess(executable, spawnedArguments, {
       cwd: options.cwd,
       env: process.env,
       shell: false,
       windowsVerbatimArguments: windowsCommandShell,
       windowsHide: true,
       stdio: ["ignore", "pipe", "pipe"]
     });
     child.stdout.on("data", stdoutCollector.write);
     child.stderr.on("data", stderrCollector.write);
     child.on("error", (error) => {

com.github.copilot/agents/fixlab.agent.md

@@ -51,7 +51,10 @@ instructions provide authoritative commands.
 10. Present evidence, skipped gates, and remaining risks.
 
 Run verbose repository-owned tests, builds, linters, type checks, and package
-validation through `fixlab exec --stage <stage> -- <command> [args...]`. Use
+validation through `fixlab exec --stage <stage> -- <command> [args...]`. In
+Windows PowerShell write the separator quoted (`'--'`) because an unquoted `--`
+is dropped, and never wrap `fixlab exec` in `cmd /c`; that wrapper is denied in
+non-interactive sessions. Use
 its compact result in agent context and refer to the private evidence path only
 when an exact diagnostic is needed. Add `--reuse` only for deterministic local
 validation when the command, arguments, Git state, working directory, profile,

dashboard/server.js

@@ -859,7 +859,7 @@ ${branchNamingGuidance}
 - Autonomously complete the lifecycle without asking the user to direct routine engineering steps.
 - ${playwrightOnly ? "Skip source diagnosis, separate reproduction, implementation, diff review, and non-browser validation. Mark diagnosis, reproduce, fix, and review skipped with the reason Playwright-only mode was selected." : "Inspect the affected surface, implement the smallest required code when edits are allowed, and self-review the effective diff for correctness, scope, and unrelated changes."}
 - ${playwrightOnly ? "Run only setup commands strictly required to launch the profile-defined applications and focused Playwright journey. Do not reinstall dependencies that are already available." : "Run the focused repository-owned tests, type-checks, and builds needed for the affected surface and risk. Preserve exact results."}
-- Run verbose repository-owned tests, builds, linters, type checks, and package validation through \`fixlab exec --stage <stage> -- <command> [args...]\`. Keep its compact result in context and inspect the private evidence path only when an exact diagnostic is needed.
+- Run verbose repository-owned tests, builds, linters, type checks, and package validation through \`fixlab exec --stage <stage> -- <command> [args...]\`. In Windows PowerShell quote the separator (\`'--'\`) because an unquoted \`--\` is dropped, and never wrap \`fixlab exec\` in \`cmd /c\`; that wrapper is denied in non-interactive sessions. Keep its compact result in context and inspect the private evidence path only when an exact diagnostic is needed.
 - Add \`--reuse\` only for deterministic local validation when command, arguments, Git state, working directory, profile, and stage are unchanged. Never reuse startup, health, authentication, Playwright or live-environment validation, external-data checks, dependency installation, deployments, or mutating commands.
 - Do not skip Playwright merely because an unrelated non-browser test, build, or backend startup is failed or blocked. If frontend startup, authentication, safe data, and the selected browser journey are independently ready, run the live-test gate and preserve the other blocker separately.
 - Use local-stack only for profile-defined application startup and health. Do not mark local-stack failed because a separate test, type-check, lint, or production build reports unrelated baseline diagnostics; preserve that exact validation limitation separately and continue browser execution when startup is healthy.

docs/troubleshooting.md

@@ -12,6 +12,15 @@ repository:
 fixlab exec --cwd .\frontend --stage test -- npm test
 ```
 
+In Windows PowerShell an unquoted `--` is dropped, so `fixlab exec` prints its
+usage text instead of running. Quote the separator (`'--'`). Do not wrap the
+command in `cmd /c`: the wrapper is denied when no interactive user can approve
+it, which blocks the job.
+
+```powershell
+fixlab exec --cwd .\frontend --stage test '--' npm test
+```
+
 The `Evidence:` path is Git-private and contains redacted `stdout.log`,
 `stderr.log`, and `result.json`. A missing stream file means that stream was
 empty. The terminal `Context:` line and `result.json.context` report raw and

templates/fixlab-autofix.agent.md

@@ -75,7 +75,10 @@ only for information that cannot be established safely from the repository.
    or create a pull request unless the developer explicitly requests it.
 
 Run verbose repository-owned tests, builds, linters, type checks, and package
-validation through `fixlab exec --stage <stage> -- <command> [args...]`. Keep
+validation through `fixlab exec --stage <stage> -- <command> [args...]`. In
+Windows PowerShell write the separator quoted (`'--'`) because an unquoted `--`
+is dropped, and never wrap `fixlab exec` in `cmd /c`; that wrapper is denied in
+non-interactive sessions. Keep
 the compact result in context and inspect its private evidence only when an
 exact diagnostic is needed. Use `--reuse` only for deterministic local
 validation with identical command and repository state. Never reuse startup,

test/dashboard-server.test.js

@@ -1171,6 +1171,7 @@ test("dashboard parses complete stage markers and passes a job", async () => {
     assert.match(receivedPrompt, /commit changes and repository-profile or instruction changes as invalidation boundaries/i);
     assert.match(receivedPrompt, /do not feed unbounded raw output back into prompts/i);
     assert.match(receivedPrompt, /fixlab exec --stage <stage>/i);
+    assert.match(receivedPrompt, /never wrap `fixlab exec` in `cmd \/c`/u);
     assert.match(receivedPrompt, /Never reuse startup, health, authentication/i);
     assert.match(receivedPrompt, /FIXLAB_STAGE\|stage\|status\|message/);
     assert.match(receivedPrompt, /FIXLAB_BUG\|id\|outcome\|owner\|summary/);

test/structured-command.test.js

@@ -318,7 +318,11 @@ test("structured command arguments enforce repository containment", () => {
     );
     assert.throws(
       () => parseStructuredCommandArguments([repository, process.execPath]),
-      /usage: fixlab exec/
+      /usage: fixlab exec[\s\S]*quoted/u
+    );
+    assert.throws(
+      () => parseStructuredCommandArguments([repository, process.execPath]),
+      /do not wrap the command in cmd \/c/u
     );
     assert.throws(
       () =>