fix(exec): preserve single-token run summaries to keep package-script context

ly-wang19 · ly-wang19 · commit de8e9e3ea266 · 2026-06-28T23:47:00.000+08:00
Tighten stripLeadingExecDisplayVerb so it only strips `run` when the rest carries a binary plus arguments (the catch-all shape). Single-token `run <X>` summaries survive because package-script names (`npm run dev`, `npm run custom-script`) are indistinguishable from bare-binary invocations (`run python3`) at strip time. This restores arbitrary package-script run labels while still collapsing the reported `run python3 /path` case (issue #97319).
diff --git a/src/agents/tool-display-exec.ts b/src/agents/tool-display-exec.ts
@@ -290,32 +290,37 @@ function summarizeKnownExec(words: string[]): string {
 // `run` is the only framework-injected verb in `summarizeKnownExec` output; every
 // other leading word (`check git status`, `install dependencies`, `start app`,
 // `show last 20 lines`) carries diagnostic meaning and must survive stripping.
-// `summarizeKnownExec` also emits a small set of semantic `run <noun>` labels from
-// the npm/pnpm/yarn subcommand map (`run tests`, `run build`, `run lint`,
-// `run script`); those must also survive so the failure message keeps the
-// package-manager script context. Any other `run <binary> ...` shape is
-// framework-injected by the catch-all and is stripped so the backtick-wrapped
-// text is exactly what was run (issue #97319).
-const SEMANTIC_RUN_WORDS: ReadonlySet<string> = new Set(["tests", "build", "lint", "script"]);
+// `summarizeKnownExec` also emits `run <noun>` labels from the npm/pnpm/yarn
+// subcommand map — both fixed (`run tests`, `run build`, `run lint`, `run script`)
+// and arbitrary (`run dev`, `run ${scriptName}`) — which must survive so the
+// failure message keeps the package-manager script context. Only the catch-all
+// `run <binary> <args>` shape is stripped; that shape always carries at least
+// one whitespace-separated argument after the binary, so a `run <single-token>`
+// meta is never ambiguous and never stripped (issue #97319).
 
 /**
- * Strips a leading framework-injected `run` verb when it precedes a binary
- * (e.g. `run python3 /path` -> `python3 /path`) so the action label is not
- * backtick-wrapped together with the literal command (issue #97319).
+ * Strips a leading framework-injected `run` verb when the rest carries both a
+ * binary and its arguments (e.g. `run python3 /path` -> `python3 /path`) so the
+ * action label is not backtick-wrapped together with the literal command
+ * (issue #97319).
  *
  * Returns the original meta unchanged for any other shape — including action-
- * bearing summaries like `install dependencies`, `start app`, `run tests`,
- * `run build`, or `show last 20 lines` — so diagnostic verbs and package-manager
- * script labels stay visible in failure warnings.
+ * bearing summaries like `install dependencies`, `start app`, package-manager
+ * script labels like `run dev`, `run tests`, `run build`, or `show last 20
+ * lines` — so diagnostic verbs and arbitrary package-script names stay visible
+ * in failure warnings.
  */
 export function stripLeadingExecDisplayVerb(meta: string): string {
   const match = meta.match(/^run\s+(\S.*)$/);
   if (!match) {
     return meta;
   }
   const rest = match[1];
-  const firstWord = rest.match(/^([A-Za-z0-9._-]+)/)?.[1];
-  if (!firstWord || SEMANTIC_RUN_WORDS.has(firstWord)) {
+  // Only strip when the meta unambiguously came from the catch-all — i.e. the
+  // rest contains a binary AND its arguments. A single-token `run <X>` could be
+  // an arbitrary package-script name (`npm run dev`) or a fixed semantic label
+  // (`run tests`), so we preserve those to avoid dropping script context.
+  if (!/\s/.test(rest)) {
     return meta;
   }
   return rest;
diff --git a/src/agents/tool-display.test.ts b/src/agents/tool-display.test.ts
@@ -587,7 +587,7 @@ describe("compactRawCommand middle truncation", () => {
 });
 
 describe("stripLeadingExecDisplayVerb", () => {
-  it("strips `run` when followed by any binary so the framework verb stays outside the backticks", () => {
+  it("strips `run` when followed by a binary plus arguments so the framework verb stays outside the backticks", () => {
     expect(stripLeadingExecDisplayVerb("run python3 /path/to/script.py")).toBe(
       "python3 /path/to/script.py",
     );
@@ -601,19 +601,27 @@ describe("stripLeadingExecDisplayVerb", () => {
     expect(stripLeadingExecDisplayVerb("run my-custom-tool --flag value")).toBe(
       "my-custom-tool --flag value",
     );
-    expect(stripLeadingExecDisplayVerb("run ./bin/local-script")).toBe("./bin/local-script");
+  });
+
+  it("preserves single-token `run <X>` summaries because package-script names are indistinguishable from bare binaries", () => {
+    // `summarizeKnownExec` emits `run <scriptName>` for arbitrary npm/pnpm/yarn
+    // scripts (`npm run dev`, `npm run start`, `npm run custom-script`); those
+    // single-token metas cannot be told apart from a bare-binary `run python3`
+    // at strip time, so we keep `run` to preserve the package-script context.
+    expect(stripLeadingExecDisplayVerb("run tests")).toBe("run tests");
+    expect(stripLeadingExecDisplayVerb("run build")).toBe("run build");
+    expect(stripLeadingExecDisplayVerb("run lint")).toBe("run lint");
+    expect(stripLeadingExecDisplayVerb("run script")).toBe("run script");
+    expect(stripLeadingExecDisplayVerb("run dev")).toBe("run dev");
+    expect(stripLeadingExecDisplayVerb("run custom-script")).toBe("run custom-script");
   });
 
   it("preserves action-bearing exec summaries that `summarizeKnownExec` emits with semantic verbs", () => {
-    // `install dependencies`, `start app`, `run tests`, `show last 20 lines`,
+    // `install dependencies`, `start app`, `show last 20 lines`,
     // `check git status`, `find files named …` all carry diagnostic meaning
     // beyond the bare command — strip must not drop those verbs.
     expect(stripLeadingExecDisplayVerb("install dependencies")).toBe("install dependencies");
     expect(stripLeadingExecDisplayVerb("start app")).toBe("start app");
-    expect(stripLeadingExecDisplayVerb("run tests")).toBe("run tests");
-    expect(stripLeadingExecDisplayVerb("run build")).toBe("run build");
-    expect(stripLeadingExecDisplayVerb("run lint")).toBe("run lint");
-    expect(stripLeadingExecDisplayVerb("run script")).toBe("run script");
     expect(stripLeadingExecDisplayVerb("show some-file.txt (workspace)")).toBe(
       "show some-file.txt (workspace)",
     );
