fix: use separate prompts for agentic vs regular providers

Agentic providers get a file-read prompt with the test content file
copied into the working directory (gemini-cli sandboxes to cwd).
Regular providers keep the original shell/ls prompt that produces
tool-call log patterns.

Signed-off-by: Adrian Cole <adrian@tetrate.io>

Author: codefromthecrypt

scripts/test_providers.sh

@@ -262,11 +262,22 @@ run_test() {
 
   local testdir=$(mktemp -d)
 
+  # Agentic providers use a file-read prompt with known content marker;
+  # regular providers use the shell prompt that produces tool-call logs.
+  local prompt
+  if is_agentic_provider "$provider"; then
+    cp "$TEST_FILE" "$testdir/test-content.txt"
+    prompt="read ./test-content.txt and output its contents exactly"
+  else
+    echo "hello" > "$testdir/hello.txt"
+    prompt="Immediately use the shell tool to run 'ls'. Do not ask for confirmation."
+  fi
+
   # Run the test and capture output
   (
     export GOOSE_PROVIDER="$provider"
     export GOOSE_MODEL="$model"
-    cd "$testdir" && "$SCRIPT_DIR/target/debug/goose" run --text "read $TEST_FILE and output its contents exactly" --with-builtin "$BUILTINS" 2>&1
+    cd "$testdir" && "$SCRIPT_DIR/target/debug/goose" run --text "$prompt" --with-builtin "$BUILTINS" 2>&1
   ) > "$output_file" 2>&1
 
   # Check result: agentic providers return text containing the test content