refactor(agents): improve command argument handling and detection

Author: SDSLeon

src/supervisor/agents/claude/index.ts

@@ -108,14 +108,35 @@ export function createClaudeAdapter(): AgentAdapter {
     defaultOneShotModel: "haiku",
     buildOneShotCommand(model, effort, prompt) {
       if (!prompt) return undefined;
-      const args = ["-p", prompt, "--model", model];
+      // --no-session-persistence keeps title/commit/PR-summary calls out of
+      // the `/resume` picker. --fallback-model auto-degrades to Haiku if the
+      // primary is overloaded so async title generation does not silently
+      // fail when the API throttles.
+      const args = [
+        "-p",
+        prompt,
+        "--model",
+        model,
+        "--fallback-model",
+        "haiku",
+        "--no-session-persistence",
+      ];
       if (effort) {
         args.push("--effort", effort);
       }
       return { command: "claude", args, stdin: "" };
     },
     buildContextExtractionCommand(sessionRef, _location, model) {
-      const args = ["-p", "--resume", sessionRef.providerSessionId, "--model", model ?? "haiku"];
+      // The resumed session is read-only here; --no-session-persistence
+      // prevents the extraction turn from being written back to disk.
+      const args = [
+        "-p",
+        "--resume",
+        sessionRef.providerSessionId,
+        "--model",
+        model ?? "haiku",
+        "--no-session-persistence",
+      ];
       return { command: "claude", args };
     },
   };

src/supervisor/agents/commandBuilders.test.ts

@@ -273,7 +273,17 @@ describe("agent command builders", () => {
       createClaudeAdapter().buildOneShotCommand?.("haiku", "low", "Summarize this diff"),
     ).toEqual({
       command: "claude",
-      args: ["-p", "Summarize this diff", "--model", "haiku", "--effort", "low"],
+      args: [
+        "-p",
+        "Summarize this diff",
+        "--model",
+        "haiku",
+        "--fallback-model",
+        "haiku",
+        "--no-session-persistence",
+        "--effort",
+        "low",
+      ],
       stdin: "",
     });
 
@@ -362,7 +372,7 @@ describe("agent command builders", () => {
   );
 
   it.skipIf(process.platform !== "win32")(
-    "prefixes the initial Copilot interactive prompt with /plan in plan mode",
+    "passes --plan and an unmodified initial prompt when launching in plan mode",
     () => {
       const spec = launch(
         createCopilotAdapter(),
@@ -373,9 +383,10 @@ describe("agent command builders", () => {
       const { cmd, cmdArgs } = parseWindowsSpec(spec);
 
       expect(cmd).toBe("copilot");
+      expect(cmdArgs).toContain("--plan");
       expect(cmdArgs).toContain("-i");
-      expect(cmdArgs).toContain("/plan hi");
-      expect(cmdArgs).not.toContain("hi");
+      expect(cmdArgs).toContain("hi");
+      expect(cmdArgs).not.toContain("/plan hi");
     },
   );
 

src/supervisor/agents/copilot/argv.ts

@@ -1,27 +1,13 @@
 import type { ThreadConfig } from "@/shared/contracts";
 import { sanitizeModelName } from "./terminal";
 
-export function formatCopilotInteractivePrompt(prompt: string, config?: ThreadConfig): string {
-  if (config?.mode !== "plan") {
-    return prompt;
-  }
-
-  const trimmed = prompt.trimStart();
-  if (trimmed.startsWith("/")) {
-    return prompt;
-  }
-
-  return `/plan ${prompt}`;
-}
-
 export function buildCopilotArgs(
   config: ThreadConfig,
   prompt: string,
   sessionId: string,
   _launchOptions?: { suppressResumeConfigOverrides?: boolean },
 ): string[] {
   const args = [`--resume=${sessionId}`, "--allow-all-paths"];
-  const formattedPrompt = formatCopilotInteractivePrompt(prompt, config);
 
   // Copilot's TUI only reflects the selected model/effort when the resume
   // command also carries those flags, even if ACP already applied them.
@@ -35,11 +21,14 @@ export function buildCopilotArgs(
   if (config.effort) {
     args.push("--effort", config.effort);
   }
+  if (config.mode === "plan") {
+    args.push("--plan");
+  }
   if (config.approvalPolicy === "never") {
     args.push("--yolo");
   }
-  if (formattedPrompt.trim().length > 0) {
-    args.push("-i", formattedPrompt);
+  if (prompt.trim().length > 0) {
+    args.push("-i", prompt);
   }
 
   return args;

src/supervisor/agents/copilot/copilot.test.ts

@@ -398,18 +398,18 @@ describe("createCopilotAdapter", () => {
     },
   );
 
-  it("prefixes plan-mode interactive prompts with /plan", () => {
+  it("passes plan-mode prompts through unchanged (mode is set via --plan at launch)", () => {
     const adapter = createCopilotAdapter();
 
     expect(
       adapter.buildDirectInput?.("fix bug", undefined, {
         model: "gpt-5.4",
         mode: "plan",
       }),
-    ).toEqual(["/plan fix bug", "@wait:40", "\r"]);
+    ).toEqual(["fix bug", "@wait:40", "\r"]);
   });
 
-  it("does not double-prefix existing slash commands in plan mode", () => {
+  it("preserves user-typed slash commands verbatim", () => {
     const adapter = createCopilotAdapter();
 
     expect(

src/supervisor/agents/copilot/index.ts

@@ -13,7 +13,7 @@ import {
 } from "../base";
 import { resolveAgentBinaryPath } from "../binaryResolver";
 import { resolveInstallNodePath, warnIfPluginManifestMissing } from "../plugin/installerBase";
-import { buildCopilotArgs, formatCopilotInteractivePrompt } from "./argv";
+import { buildCopilotArgs } from "./argv";
 import { buildCopilotCommand, copilotDefaultCapabilities, copilotDetectionSpec } from "./detection";
 import {
   installCopilotPlugin,
@@ -118,8 +118,10 @@ export function createCopilotAdapter(): AgentAdapter {
     createInitialSessionRef() {
       return undefined;
     },
-    buildDirectInput(prompt, _segments, config) {
-      return [formatCopilotInteractivePrompt(prompt, config), "@wait:40", "\r"];
+    buildDirectInput(prompt) {
+      // Mode is set at launch via --plan / --mode. The TUI persists the
+      // selection across turns, so subsequent prompts pass through as-is.
+      return [prompt, "@wait:40", "\r"];
     },
     formatPromptSegments(segments: PromptSegment[]) {
       const attachments = segments.filter((segment) => segment.kind === "attachment");

src/supervisor/agents/gemini/argv.ts

@@ -8,6 +8,14 @@ export function buildGeminiArgs(
 ): string[] {
   const args: string[] = [];
 
+  // Gemini emits "Skipping project agents due to untrusted folder..." on
+  // stdout when the workspace is not pre-trusted. In --acp mode that string
+  // collides with JSON-RPC frames and breaks the stream parser. --skip-trust
+  // suppresses the prompt for this session (replaces the older
+  // GEMINI_CLI_TRUST_WORKSPACE=true env var, which only worked through env
+  // inheritance and was easy to lose across spawn wrappers).
+  args.push("--skip-trust");
+
   if (resumeSessionId) {
     args.push("--resume", resumeSessionId);
   } else if (assignedSessionId) {

src/supervisor/agents/gemini/detection.test.ts

@@ -51,6 +51,7 @@ describe("geminiDetectionSpec", () => {
 
     expect(buildAgentCommandMock).toHaveBeenCalledWith(location, "/Users/demo/.local/bin/gemini", [
       "--acp",
+      "--skip-trust",
     ]);
     expect(probeAcpCapabilitiesMock).toHaveBeenCalledWith(
       "/bin/zsh",

src/supervisor/agents/gemini/detection.ts

@@ -136,19 +136,18 @@ export const geminiDetectionSpec: DetectionSpec = {
     // Bypass Gemini's folder-trust check during the probe so the AgentRegistry
     // doesn't emit "Skipping project agents..." onto stdout, which can collide
     // with JSON-RPC frames and break the ACP parser.
-    const trustEnv = { GEMINI_CLI_TRUST_WORKSPACE: "true" };
+    const probeArgs = ["--acp", "--skip-trust"];
     const probeCmd =
       ctx.location.kind === "wsl"
-        ? buildAgentCommand(ctx.location, "gemini", ["--acp"], ctx.executablePath, trustEnv)
-        : buildAgentCommand(ctx.location, ctx.executablePath, ["--acp"]);
+        ? buildAgentCommand(ctx.location, "gemini", probeArgs, ctx.executablePath)
+        : buildAgentCommand(ctx.location, ctx.executablePath, probeArgs);
     const probeCwd = ctx.location.kind === "wsl" ? "/tmp" : homedir();
     const probeResult = await probeAcpCapabilities(probeCmd.command, probeCmd.args, probeCwd, {
       timeoutMs: 15_000,
       label:
         ctx.location.kind === "wsl"
           ? `gemini:wsl:${ctx.location.distro}`
           : `gemini:${ctx.location.kind}`,
-      ...(ctx.location.kind === "wsl" ? {} : { env: trustEnv }),
     });
     if (!probeResult) return undefined;
     const modelTokens = new Map<string, number>();

src/supervisor/agents/gemini/index.ts

@@ -44,13 +44,12 @@ export function createGeminiAdapter(): AgentAdapter {
     get capabilities() {
       return capabilities;
     },
-    // GEMINI_CLI_TRUST_WORKSPACE=true bypasses Gemini's folder-trust check.
-    // Without it, the AgentRegistry emits "Skipping project agents due to
-    // untrusted folder..." onto stdout, which can collide with JSON-RPC
-    // frames in --acp mode and break the ACP stream parser.
+    // Workspace trust is now suppressed via --skip-trust on every gemini
+    // invocation (see buildGeminiArgs and the --acp launch below). WSL still
+    // needs BROWSER=/bin/true so the OAuth flow does not try to xdg-open a
+    // browser inside the distro and hang the PTY.
     spawnEnv: {
-      native: { GEMINI_CLI_TRUST_WORKSPACE: "true" },
-      wsl: { BROWSER: "/bin/true", GEMINI_CLI_TRUST_WORKSPACE: "true" },
+      wsl: { BROWSER: "/bin/true" },
     },
     pluginId: "lightcode-status@gemini",
     pluginVersion: GEMINI_PLUGIN_VERSION,
@@ -108,9 +107,8 @@ export function createGeminiAdapter(): AgentAdapter {
       const command = buildAgentCommand(
         input.projectLocation,
         "gemini",
-        ["--acp"],
+        ["--acp", "--skip-trust"],
         resolveAgentBinaryPath(input.projectLocation, "gemini"),
-        { GEMINI_CLI_TRUST_WORKSPACE: "true" },
       );
       return createAcpStructuredSession(command, input);
     },